Stowage/tutanota
2
0
Fork 0
mirror of https://github.com/tutao/tutanota.git synced 2025-12-07 13:49:47 +00:00
Code Issues Projects Releases Packages Wiki Activity
tutanota/libs/tensorflow.js
abp e196f8f8e4
add tensorflow cpu backend as fallback for webgl backend 
When the webgl backend is not available or unsupported,
we fall back to the tensorflow cpu backend.

Tensorflow cpu backend library review by abp and jhm.

Co-authored-by: jomapp <17314077+jomapp@users.noreply.github.com>
2025-11-20 16:06:18 +01:00

64377 lines
2.4 MiB
Vendored
Raw Blame History

function _mergeNamespaces(n, m) {
m.forEach(function (e) {
e && typeof e !== 'string' && !Array.isArray(e) && Object.keys(e).forEach(function (k) {
if (k !== 'default' && !(k in n)) {
var d = Object.getOwnPropertyDescriptor(e, k);
Object.defineProperty(n, k, d.get ? d : {
enumerable: true,
get: function () { return e[k]; }
});
}
});
});
return Object.freeze(n);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const EPSILON_FLOAT32$1 = 1e-7;
const EPSILON_FLOAT16$1 = 1e-4;
/** Convenient class for storing tensor-related data. */
class DataStorage {
constructor(backend, dataMover) {
this.backend = backend;
this.dataMover = dataMover;
this.data = new WeakMap();
this.dataIdsCount = 0;
}
get(dataId) {
if (!this.data.has(dataId)) {
this.dataMover.moveData(this.backend, dataId);
}
return this.data.get(dataId);
}
set(dataId, value) {
this.dataIdsCount++;
this.data.set(dataId, value);
}
has(dataId) {
return this.data.has(dataId);
}
delete(dataId) {
this.dataIdsCount--;
return this.data.delete(dataId);
}
numDataIds() {
return this.dataIdsCount;
}
}
/**
* The interface that defines the kernels that should be implemented when
* adding a new backend. New backends don't need to implement every one of the
* methods, this can be done gradually (throw an error for unimplemented
* methods).
*/
class KernelBackend {
refCount(dataId) {
return notYetImplemented('refCount');
}
incRef(dataId) {
return notYetImplemented('incRef');
}
timerAvailable() {
return true;
}
time(f) {
return notYetImplemented('time');
}
read(dataId) {
return notYetImplemented('read');
}
readSync(dataId) {
return notYetImplemented('readSync');
}
readToGPU(dataId, options) {
return notYetImplemented('readToGPU');
}
numDataIds() {
return notYetImplemented('numDataIds');
}
disposeData(dataId, force) {
return notYetImplemented('disposeData');
}
write(values, shape, dtype) {
return notYetImplemented('write');
}
move(dataId, values, shape, dtype, refCount) {
return notYetImplemented('move');
}
createTensorFromGPUData(values, shape, dtype) {
return notYetImplemented('createTensorFromGPUData');
}
memory() {
return notYetImplemented('memory');
}
/** Returns the highest precision for floats in bits (e.g. 16 or 32) */
floatPrecision() {
return notYetImplemented('floatPrecision');
}
/** Returns the smallest representable number. */
epsilon() {
return this.floatPrecision() === 32 ? EPSILON_FLOAT32$1 : EPSILON_FLOAT16$1;
}
dispose() {
return notYetImplemented('dispose');
}
}
function notYetImplemented(kernelName) {
throw new Error(`'${kernelName}' not yet implemented or not found in the registry. ` +
`This kernel may not be supported by the tfjs backend you have chosen`);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Shuffles the array in-place using Fisher-Yates algorithm.
*
* ```js
* const a = [1, 2, 3, 4, 5];
* tf.util.shuffle(a);
* console.log(a);
* ```
*
* @param array The array to shuffle in-place.
*
* @doc {heading: 'Util', namespace: 'util'}
*/
// tslint:disable-next-line:no-any
function shuffle(array) {
let counter = array.length;
let index = 0;
// While there are elements in the array
while (counter > 0) {
// Pick a random index
index = (Math.random() * counter) | 0;
// Decrease counter by 1
counter--;
// And swap the last element with it
swap(array, counter, index);
}
}
/** Clamps a value to a specified range. */
function clamp(min, x, max) {
return Math.max(min, Math.min(x, max));
}
function nearestLargerEven(val) {
return val % 2 === 0 ? val : val + 1;
}
function swap(object, left, right) {
const temp = object[left];
object[left] = object[right];
object[right] = temp;
}
function sum$3(arr) {
let sum = 0;
for (let i = 0; i < arr.length; i++) {
sum += arr[i];
}
return sum;
}
/**
* Asserts that the expression is true. Otherwise throws an error with the
* provided message.
*
* ```js
* const x = 2;
* tf.util.assert(x === 2, 'x is not 2');
* ```
*
* @param expr The expression to assert (as a boolean).
* @param msg A function that returns the message to report when throwing an
* error. We use a function for performance reasons.
*
* @doc {heading: 'Util', namespace: 'util'}
*/
function assert$1(expr, msg) {
if (!expr) {
throw new Error(typeof msg === 'string' ? msg : msg());
}
}
function assertShapesMatch(shapeA, shapeB, errorMessagePrefix = '') {
assert$1(arraysEqual(shapeA, shapeB), () => errorMessagePrefix + ` Shapes ${shapeA} and ${shapeB} must match`);
}
function assertNonNull(a) {
assert$1(a != null, () => `The input to the tensor constructor must be a non-null value.`);
}
/**
* Returns the size (number of elements) of the tensor given its shape.
*
* ```js
* const shape = [3, 4, 2];
* const size = tf.util.sizeFromShape(shape);
* console.log(size);
* ```
*
* @doc {heading: 'Util', namespace: 'util'}
*/
function sizeFromShape(shape) {
if (shape.length === 0) {
// Scalar.
return 1;
}
let size = shape[0];
for (let i = 1; i < shape.length; i++) {
size *= shape[i];
}
return size;
}
function arraysEqual(n1, n2) {
if (n1 === n2) {
return true;
}
if (n1 == null || n2 == null) {
return false;
}
if (n1.length !== n2.length) {
return false;
}
for (let i = 0; i < n1.length; i++) {
if (n1[i] !== n2[i]) {
return false;
}
}
return true;
}
function isInt(a) {
return a % 1 === 0;
}
function sizeToSquarishShape(size) {
const width = Math.ceil(Math.sqrt(size));
return [width, Math.ceil(size / width)];
}
function rightPad(a, size) {
if (size <= a.length) {
return a;
}
return a + ' '.repeat(size - a.length);
}
function repeatedTry(checkFn, delayFn = (counter) => 0, maxCounter, scheduleFn) {
return new Promise((resolve, reject) => {
let tryCount = 0;
const tryFn = () => {
if (checkFn()) {
resolve();
return;
}
tryCount++;
const nextBackoff = delayFn(tryCount);
if (maxCounter != null && tryCount >= maxCounter) {
reject();
return;
}
if (scheduleFn != null) {
scheduleFn(tryFn, nextBackoff);
}
else {
// google3 does not allow assigning another variable to setTimeout.
// Don't refactor this so scheduleFn has a default value of setTimeout.
setTimeout(tryFn, nextBackoff);
}
};
tryFn();
});
}
/**
* Given the full size of the array and a shape that may contain -1 as the
* implicit dimension, returns the inferred shape where -1 is replaced.
* E.g. For shape=[2, -1, 3] and size=24, it will return [2, 4, 3].
*
* @param shape The shape, which may contain -1 in some dimension.
* @param size The full size (number of elements) of the array.
* @return The inferred shape where -1 is replaced with the inferred size.
*/
function inferFromImplicitShape(shape, size) {
let shapeProd = 1;
let implicitIdx = -1;
for (let i = 0; i < shape.length; ++i) {
if (shape[i] >= 0) {
shapeProd *= shape[i];
}
else if (shape[i] === -1) {
if (implicitIdx !== -1) {
throw Error(`Shapes can only have 1 implicit size. ` +
`Found -1 at dim ${implicitIdx} and dim ${i}`);
}
implicitIdx = i;
}
else if (shape[i] < 0) {
throw Error(`Shapes can not be < 0. Found ${shape[i]} at dim ${i}`);
}
}
if (implicitIdx === -1) {
if (size > 0 && size !== shapeProd) {
throw Error(`Size(${size}) must match the product of shape ${shape}`);
}
return shape;
}
if (shapeProd === 0) {
throw Error(`Cannot infer the missing size in [${shape}] when ` +
`there are 0 elements`);
}
if (size % shapeProd !== 0) {
throw Error(`The implicit shape can't be a fractional number. ` +
`Got ${size} / ${shapeProd}`);
}
const newShape = shape.slice();
newShape[implicitIdx] = size / shapeProd;
return newShape;
}
function parseAxisParam(axis, shape) {
const rank = shape.length;
// Normalize input
axis = axis == null ? shape.map((s, i) => i) : [].concat(axis);
// Check for valid range
assert$1(axis.every(ax => ax >= -rank && ax < rank), () => `All values in axis param must be in range [-${rank}, ${rank}) but ` +
`got axis ${axis}`);
// Check for only integers
assert$1(axis.every(ax => isInt(ax)), () => `All values in axis param must be integers but ` +
`got axis ${axis}`);
// Handle negative axis.
return axis.map(a => a < 0 ? rank + a : a);
}
/** Reduces the shape by removing all dimensions of shape 1. */
function squeezeShape(shape, axis) {
const newShape = [];
const keptDims = [];
const isEmptyArray = axis != null && Array.isArray(axis) && axis.length === 0;
const axes = (axis == null || isEmptyArray) ?
null :
parseAxisParam(axis, shape).sort();
let j = 0;
for (let i = 0; i < shape.length; ++i) {
if (axes != null) {
if (axes[j] === i && shape[i] !== 1) {
throw new Error(`Can't squeeze axis ${i} since its dim '${shape[i]}' is not 1`);
}
if ((axes[j] == null || axes[j] > i) && shape[i] === 1) {
newShape.push(shape[i]);
keptDims.push(i);
}
if (axes[j] <= i) {
j++;
}
}
if (shape[i] !== 1) {
newShape.push(shape[i]);
keptDims.push(i);
}
}
return { newShape, keptDims };
}
function getTypedArrayFromDType(dtype, size) {
return getArrayFromDType(dtype, size);
}
function getArrayFromDType(dtype, size) {
let values = null;
if (dtype == null || dtype === 'float32') {
values = new Float32Array(size);
}
else if (dtype === 'int32') {
values = new Int32Array(size);
}
else if (dtype === 'bool') {
values = new Uint8Array(size);
}
else if (dtype === 'string') {
values = new Array(size);
}
else {
throw new Error(`Unknown data type ${dtype}`);
}
return values;
}
function checkConversionForErrors(vals, dtype) {
for (let i = 0; i < vals.length; i++) {
const num = vals[i];
if (isNaN(num) || !isFinite(num)) {
throw Error(`A tensor of type ${dtype} being uploaded contains ${num}.`);
}
}
}
/** Returns true if the dtype is valid. */
function isValidDtype(dtype) {
return dtype === 'bool' || dtype === 'complex64' || dtype === 'float32' ||
dtype === 'int32' || dtype === 'string';
}
/**
* Returns true if the new type can't encode the old type without loss of
* precision.
*/
function hasEncodingLoss(oldType, newType) {
if (newType === 'complex64') {
return false;
}
if (newType === 'float32' && oldType !== 'complex64') {
return false;
}
if (newType === 'int32' && oldType !== 'float32' && oldType !== 'complex64') {
return false;
}
if (newType === 'bool' && oldType === 'bool') {
return false;
}
return true;
}
function bytesPerElement(dtype) {
if (dtype === 'float32' || dtype === 'int32') {
return 4;
}
else if (dtype === 'complex64') {
return 8;
}
else if (dtype === 'bool') {
return 1;
}
else {
throw new Error(`Unknown dtype ${dtype}`);
}
}
/**
* Returns the approximate number of bytes allocated in the string array - 2
* bytes per character. Computing the exact bytes for a native string in JS
* is not possible since it depends on the encoding of the html page that
* serves the website.
*/
function bytesFromStringArray(arr) {
if (arr == null) {
return 0;
}
let bytes = 0;
arr.forEach(x => bytes += x.length);
return bytes;
}
/** Returns true if the value is a string. */
function isString(value) {
return typeof value === 'string' || value instanceof String;
}
function isBoolean(value) {
return typeof value === 'boolean';
}
function isNumber(value) {
return typeof value === 'number';
}
function inferDtype(values) {
if (Array.isArray(values)) {
return inferDtype(values[0]);
}
if (values instanceof Float32Array) {
return 'float32';
}
else if (values instanceof Int32Array || values instanceof Uint8Array ||
values instanceof Uint8ClampedArray) {
return 'int32';
}
else if (isNumber(values)) {
return 'float32';
}
else if (isString(values)) {
return 'string';
}
else if (isBoolean(values)) {
return 'bool';
}
return 'float32';
}
function isFunction(f) {
return !!(f && f.constructor && f.call && f.apply);
}
function nearestDivisor(size, start) {
for (let i = start; i < size; ++i) {
if (size % i === 0) {
return i;
}
}
return size;
}
function computeStrides(shape) {
const rank = shape.length;
if (rank < 2) {
return [];
}
// Last dimension has implicit stride of 1, thus having D-1 (instead of D)
// strides.
const strides = new Array(rank - 1);
strides[rank - 2] = shape[rank - 1];
for (let i = rank - 3; i >= 0; --i) {
strides[i] = strides[i + 1] * shape[i + 1];
}
return strides;
}
function createNestedArray(offset, shape, a, isComplex = false) {
const ret = new Array();
if (shape.length === 1) {
const d = shape[0] * (isComplex ? 2 : 1);
for (let i = 0; i < d; i++) {
ret[i] = a[offset + i];
}
}
else {
const d = shape[0];
const rest = shape.slice(1);
const len = rest.reduce((acc, c) => acc * c) * (isComplex ? 2 : 1);
for (let i = 0; i < d; i++) {
ret[i] = createNestedArray(offset + i * len, rest, a, isComplex);
}
}
return ret;
}
// Provide a nested array of TypedArray in given shape.
function toNestedArray(shape, a, isComplex = false) {
if (shape.length === 0) {
// Scalar type should return a single number.
return a[0];
}
const size = shape.reduce((acc, c) => acc * c) * (isComplex ? 2 : 1);
if (size === 0) {
// A tensor with shape zero should be turned into empty list.
return [];
}
if (size !== a.length) {
throw new Error(`[${shape}] does not match the input size ${a.length}${isComplex ? ' for a complex tensor' : ''}.`);
}
return createNestedArray(0, shape, a, isComplex);
}
function convertBackendValuesAndArrayBuffer(data, dtype) {
// If is type Uint8Array[], return it directly.
if (Array.isArray(data)) {
return data;
}
if (dtype === 'float32') {
return data instanceof Float32Array ? data : new Float32Array(data);
}
else if (dtype === 'int32') {
return data instanceof Int32Array ? data : new Int32Array(data);
}
else if (dtype === 'bool' || dtype === 'string') {
return Uint8Array.from(new Int32Array(data));
}
else {
throw new Error(`Unknown dtype ${dtype}`);
}
}
function makeOnesTypedArray(size, dtype) {
const array = makeZerosTypedArray(size, dtype);
for (let i = 0; i < array.length; i++) {
array[i] = 1;
}
return array;
}
function makeZerosTypedArray(size, dtype) {
if (dtype == null || dtype === 'float32' || dtype === 'complex64') {
return new Float32Array(size);
}
else if (dtype === 'int32') {
return new Int32Array(size);
}
else if (dtype === 'bool') {
return new Uint8Array(size);
}
else {
throw new Error(`Unknown data type ${dtype}`);
}
}
/**
* Make nested `TypedArray` filled with zeros.
* @param shape The shape information for the nested array.
* @param dtype dtype of the array element.
*/
function makeZerosNestedTypedArray(shape, dtype) {
const size = shape.reduce((prev, curr) => prev * curr, 1);
if (dtype == null || dtype === 'float32') {
return toNestedArray(shape, new Float32Array(size));
}
else if (dtype === 'int32') {
return toNestedArray(shape, new Int32Array(size));
}
else if (dtype === 'bool') {
return toNestedArray(shape, new Uint8Array(size));
}
else {
throw new Error(`Unknown data type ${dtype}`);
}
}
function assertNonNegativeIntegerDimensions(shape) {
shape.forEach(dimSize => {
assert$1(Number.isInteger(dimSize) && dimSize >= 0, () => `Tensor must have a shape comprised of positive integers but got ` +
`shape [${shape}].`);
});
}
/**
* Computes flat index for a given location (multidimentionsal index) in a
* Tensor/multidimensional array.
*
* @param locs Location in the tensor.
* @param rank Rank of the tensor.
* @param strides Tensor strides.
*/
function locToIndex(locs, rank, strides) {
if (rank === 0) {
return 0;
}
else if (rank === 1) {
return locs[0];
}
let index = locs[locs.length - 1];
for (let i = 0; i < locs.length - 1; ++i) {
index += strides[i] * locs[i];
}
return index;
}
/**
* Computes the location (multidimensional index) in a
* tensor/multidimentional array for a given flat index.
*
* @param index Index in flat array.
* @param rank Rank of tensor.
* @param strides Strides of tensor.
*/
function indexToLoc(index, rank, strides) {
if (rank === 0) {
return [];
}
else if (rank === 1) {
return [index];
}
const locs = new Array(rank);
for (let i = 0; i < locs.length - 1; ++i) {
locs[i] = Math.floor(index / strides[i]);
index -= locs[i] * strides[i];
}
locs[locs.length - 1] = index;
return locs;
}
/**
* This method asserts whether an object is a Promise instance.
* @param object
*/
// tslint:disable-next-line: no-any
function isPromise(object) {
// We chose to not use 'obj instanceOf Promise' for two reasons:
// 1. It only reliably works for es6 Promise, not other Promise
// implementations.
// 2. It doesn't work with framework that uses zone.js. zone.js monkey
// patch the async calls, so it is possible the obj (patched) is
// comparing to a pre-patched Promise.
return object && object.then && typeof object.then === 'function';
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Expects flags from URL in the format ?tfjsflags=FLAG1:1,FLAG2:true.
const TENSORFLOWJS_FLAGS_PREFIX = 'tfjsflags';
/**
* The environment contains evaluated flags as well as the registered platform.
* This is always used as a global singleton and can be retrieved with
* `tf.env()`.
*
* @doc {heading: 'Environment'}
*/
class Environment {
// tslint:disable-next-line: no-any
constructor(global) {
this.global = global;
this.flags = {};
this.flagRegistry = {};
this.urlFlags = {};
// Jasmine spies on this in 'environment_test.ts'
this.getQueryParams = getQueryParams;
this.populateURLFlags();
}
setPlatform(platformName, platform) {
if (this.platform != null) {
if (!(env().getBool('IS_TEST') || env().getBool('PROD'))) {
console.warn(`Platform ${this.platformName} has already been set. ` +
`Overwriting the platform with ${platformName}.`);
}
}
this.platformName = platformName;
this.platform = platform;
}
registerFlag(flagName, evaluationFn, setHook) {
this.flagRegistry[flagName] = { evaluationFn, setHook };
// Override the flag value from the URL. This has to happen here because
// the environment is initialized before flags get registered.
if (this.urlFlags[flagName] != null) {
const flagValue = this.urlFlags[flagName];
if (!(env().getBool('IS_TEST') || env().getBool('PROD'))) {
console.warn(`Setting feature override from URL ${flagName}: ${flagValue}.`);
}
this.set(flagName, flagValue);
}
}
async getAsync(flagName) {
if (flagName in this.flags) {
return this.flags[flagName];
}
this.flags[flagName] = await this.evaluateFlag(flagName);
return this.flags[flagName];
}
get(flagName) {
if (flagName in this.flags) {
return this.flags[flagName];
}
const flagValue = this.evaluateFlag(flagName);
if (isPromise(flagValue)) {
throw new Error(`Flag ${flagName} cannot be synchronously evaluated. ` +
`Please use getAsync() instead.`);
}
this.flags[flagName] = flagValue;
return this.flags[flagName];
}
getNumber(flagName) {
return this.get(flagName);
}
getBool(flagName) {
return this.get(flagName);
}
getString(flagName) {
return this.get(flagName);
}
getFlags() {
return this.flags;
}
// For backwards compatibility.
get features() {
return this.flags;
}
set(flagName, value) {
if (this.flagRegistry[flagName] == null) {
throw new Error(`Cannot set flag ${flagName} as it has not been registered.`);
}
this.flags[flagName] = value;
if (this.flagRegistry[flagName].setHook != null) {
this.flagRegistry[flagName].setHook(value);
}
}
evaluateFlag(flagName) {
if (this.flagRegistry[flagName] == null) {
throw new Error(`Cannot evaluate flag '${flagName}': no evaluation function found.`);
}
return this.flagRegistry[flagName].evaluationFn();
}
setFlags(flags) {
this.flags = Object.assign({}, flags);
}
reset() {
this.flags = {};
this.urlFlags = {};
this.populateURLFlags();
}
populateURLFlags() {
if (typeof this.global === 'undefined' ||
typeof this.global.location === 'undefined' ||
typeof this.global.location.search === 'undefined') {
return;
}
const urlParams = this.getQueryParams(this.global.location.search);
if (TENSORFLOWJS_FLAGS_PREFIX in urlParams) {
const keyValues = urlParams[TENSORFLOWJS_FLAGS_PREFIX].split(',');
keyValues.forEach(keyValue => {
const [key, value] = keyValue.split(':');
this.urlFlags[key] = parseValue(key, value);
});
}
}
}
function getQueryParams(queryString) {
const params = {};
queryString.replace(/[?&]([^=?&]+)(?:=([^&]*))?/g, (s, ...t) => {
decodeParam(params, t[0], t[1]);
return t.join('=');
});
return params;
}
function decodeParam(params, name, value) {
params[decodeURIComponent(name)] = decodeURIComponent(value || '');
}
function parseValue(flagName, value) {
const lowerCaseValue = value.toLowerCase();
if (lowerCaseValue === 'true' || lowerCaseValue === 'false') {
return lowerCaseValue === 'true';
}
else if (`${+lowerCaseValue}` === lowerCaseValue) {
return +lowerCaseValue;
}
else {
return value;
}
}
/**
* Returns the current environment (a global singleton).
*
* The environment object contains the evaluated feature values as well as the
* active platform.
*
* @doc {heading: 'Environment'}
*/
function env() {
return ENV$2;
}
let ENV$2 = null;
function setEnvironmentGlobal(environment) {
ENV$2 = environment;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Note that the identifier globalNameSpace is scoped to this module, but will
// always resolve to the same global object regardless of how the module is
// resolved.
// tslint:disable-next-line:no-any
let globalNameSpace;
// tslint:disable-next-line:no-any
function getGlobalNamespace() {
if (globalNameSpace == null) {
// tslint:disable-next-line:no-any
let ns;
if (typeof (window) !== 'undefined') {
ns = window;
}
else if (typeof (global) !== 'undefined') {
ns = global;
}
else if (typeof (process) !== 'undefined') {
ns = process;
}
else if (typeof (self) !== 'undefined') {
ns = self;
}
else {
throw new Error('Could not find a global object');
}
globalNameSpace = ns;
}
return globalNameSpace;
}
// tslint:disable-next-line:no-any
function getGlobalMap() {
const ns = getGlobalNamespace();
if (ns._tfGlobals == null) {
ns._tfGlobals = new Map();
}
return ns._tfGlobals;
}
/**
* Returns a globally accessible 'singleton' object.
*
* @param key the name of the object
* @param init a function to initialize to initialize this object
* the first time it is fetched.
*/
function getGlobal(key, init) {
const globalMap = getGlobalMap();
if (globalMap.has(key)) {
return globalMap.get(key);
}
else {
const singleton = init();
globalMap.set(key, singleton);
return globalMap.get(key);
}
}
const Abs = 'Abs';
const Acos = 'Acos';
const Acosh = 'Acosh';
const Add = 'Add';
const AddN = 'AddN';
const All = 'All';
const Any = 'Any';
const ArgMax = 'ArgMax';
const ArgMin = 'ArgMin';
const Asin = 'Asin';
const Asinh = 'Asinh';
const Atan = 'Atan';
const Atanh = 'Atanh';
const Atan2 = 'Atan2';
const AvgPool = 'AvgPool';
const AvgPoolGrad = 'AvgPoolGrad';
const AvgPool3D = 'AvgPool3D';
const AvgPool3DGrad = 'AvgPool3DGrad';
const BatchMatMul = 'BatchMatMul';
const BatchToSpaceND = 'BatchToSpaceND';
const Bincount = 'Bincount';
const BitwiseAnd = 'BitwiseAnd';
const BroadcastTo = 'BroadcastTo';
const BroadcastArgs = 'BroadcastArgs';
const Cast = 'Cast';
const Ceil = 'Ceil';
const ClipByValue = 'ClipByValue';
const Complex = 'Complex';
const ComplexAbs = 'ComplexAbs';
const Concat = 'Concat';
const Conv2D = 'Conv2D';
const Conv2DBackpropFilter = 'Conv2DBackpropFilter';
const Conv2DBackpropInput = 'Conv2DBackpropInput';
const Conv3D = 'Conv3D';
const Conv3DBackpropFilterV2 = 'Conv3DBackpropFilterV2';
const Conv3DBackpropInputV2 = 'Conv3DBackpropInputV2';
const Cos = 'Cos';
const Cosh = 'Cosh';
const Cumprod = 'Cumprod';
const Cumsum = 'Cumsum';
const CropAndResize = 'CropAndResize';
const DenseBincount = 'DenseBincount';
const DepthToSpace = 'DepthToSpace';
const DepthwiseConv2dNative = 'DepthwiseConv2dNative';
const DepthwiseConv2dNativeBackpropFilter = 'DepthwiseConv2dNativeBackpropFilter';
const DepthwiseConv2dNativeBackpropInput = 'DepthwiseConv2dNativeBackpropInput';
const Diag = 'Diag';
const Dilation2D = 'Dilation2D';
const Dilation2DBackpropInput = 'Dilation2DBackpropInput';
const Dilation2DBackpropFilter = 'Dilation2DBackpropFilter';
const Draw = 'Draw';
const RealDiv = 'RealDiv';
const Einsum = 'Einsum';
const Elu$1 = 'Elu';
const EluGrad = 'EluGrad';
const Erf = 'Erf';
const Equal = 'Equal';
const Exp = 'Exp';
const ExpandDims = 'ExpandDims';
const Expm1 = 'Expm1';
const FFT = 'FFT';
const Fill = 'Fill';
const FlipLeftRight = 'FlipLeftRight';
const Floor = 'Floor';
const FloorDiv = 'FloorDiv';
const FusedBatchNorm = 'FusedBatchNorm';
const GatherV2 = 'GatherV2';
const GatherNd = 'GatherNd';
const Greater = 'Greater';
const GreaterEqual = 'GreaterEqual';
const Identity$1 = 'Identity';
const IFFT = 'IFFT';
const Imag = 'Imag';
const IsFinite = 'IsFinite';
const IsInf = 'IsInf';
const IsNan = 'IsNan';
const LeakyRelu = 'LeakyRelu';
const Less = 'Less';
const LessEqual = 'LessEqual';
const LinSpace = 'LinSpace';
const Log = 'Log';
const Log1p = 'Log1p';
const LogicalAnd = 'LogicalAnd';
const LogicalNot = 'LogicalNot';
const LogicalOr = 'LogicalOr';
const LogSoftmax$1 = 'LogSoftmax';
const LRN = 'LRN';
const LRNGrad = 'LRNGrad';
const Max = 'Max';
const Maximum = 'Maximum';
const MaxPool = 'MaxPool';
const MaxPoolGrad = 'MaxPoolGrad';
const MaxPool3D = 'MaxPool3D';
const MaxPool3DGrad = 'MaxPool3DGrad';
const MaxPoolWithArgmax = 'MaxPoolWithArgmax';
const Mean = 'Mean';
const Min = 'Min';
const Minimum = 'Minimum';
const MirrorPad = 'MirrorPad';
const Mod = 'Mod';
const Multinomial = 'Multinomial';
const Multiply = 'Multiply';
const Neg = 'Neg';
const NotEqual = 'NotEqual';
const NonMaxSuppressionV3 = 'NonMaxSuppressionV3';
const NonMaxSuppressionV4 = 'NonMaxSuppressionV4';
const NonMaxSuppressionV5 = 'NonMaxSuppressionV5';
const OnesLike = 'OnesLike';
const OneHot = 'OneHot';
const Pack = 'Pack';
const PadV2 = 'PadV2';
const Pow = 'Pow';
const Prelu = 'Prelu';
const Prod = 'Prod';
const RaggedGather = 'RaggedGather';
const RaggedRange = 'RaggedRange';
const RaggedTensorToTensor = 'RaggedTensorToTensor';
const Range = 'Range';
const Real = 'Real';
const Reciprocal = 'Reciprocal';
const Relu$1 = 'Relu';
const Reshape$1 = 'Reshape';
const ResizeNearestNeighbor = 'ResizeNearestNeighbor';
const ResizeNearestNeighborGrad = 'ResizeNearestNeighborGrad';
const ResizeBilinear = 'ResizeBilinear';
const ResizeBilinearGrad = 'ResizeBilinearGrad';
const Relu6$1 = 'Relu6';
const Reverse = 'Reverse';
const Round = 'Round';
const Rsqrt = 'Rsqrt';
const ScatterNd = 'ScatterNd';
const TensorScatterUpdate = 'TensorScatterUpdate';
const SearchSorted = 'SearchSorted';
const Select = 'Select';
const Selu$1 = 'Selu';
const Slice = 'Slice';
const Sin = 'Sin';
const Sinh = 'Sinh';
const Sign = 'Sign';
const Sigmoid$1 = 'Sigmoid';
const Softplus$1 = 'Softplus';
const Sqrt = 'Sqrt';
const Sum = 'Sum';
const SpaceToBatchND = 'SpaceToBatchND';
const SplitV = 'SplitV';
const Softmax$1 = 'Softmax';
const SparseFillEmptyRows = 'SparseFillEmptyRows';
const SparseReshape = 'SparseReshape';
const SparseSegmentMean = 'SparseSegmentMean';
const SparseSegmentSum = 'SparseSegmentSum';
const SparseToDense = 'SparseToDense';
const SquaredDifference = 'SquaredDifference';
const Square = 'Square';
const StaticRegexReplace = 'StaticRegexReplace';
const StridedSlice = 'StridedSlice';
const StringNGrams = 'StringNGrams';
const StringSplit = 'StringSplit';
const StringToHashBucketFast = 'StringToHashBucketFast';
const Sub = 'Sub';
const Tan = 'Tan';
const Tanh$1 = 'Tanh';
const Tile = 'Tile';
const TopK = 'TopK';
const Transform = 'Transform';
const Transpose = 'Transpose';
const Unique = 'Unique';
const Unpack = 'Unpack';
const UnsortedSegmentSum = 'UnsortedSegmentSum';
const ZerosLike = 'ZerosLike';
/**
* TensorFlow.js-only kernels
*/
const Step = 'Step';
const FromPixels = 'FromPixels';
const RotateWithOffset = 'RotateWithOffset';
const _FusedMatMul = '_FusedMatMul';
const FusedConv2D = 'FusedConv2D';
const FusedDepthwiseConv2D = 'FusedDepthwiseConv2D';
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function warn(...msg) {
if (!(env().getBool('IS_TEST') || env().getBool('PROD'))) {
console.warn(...msg);
}
}
function log$3(...msg) {
if (!(env().getBool('IS_TEST') || env().getBool('PROD'))) {
console.log(...msg);
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const kernelRegistry = getGlobal('kernelRegistry', () => new Map());
const gradRegistry = getGlobal('gradRegistry', () => new Map());
/**
* Returns the kernel function (code) associated with the provided names.
*
* @param kernelName The official name of the kernel.
* @param backendName The official name of the backend.
*/
function getKernel(kernelName, backendName) {
const key = makeKey(kernelName, backendName);
return kernelRegistry.get(key);
}
/**
* Returns the registered gradient info associated with the provided kernel.
* @param kernelName The official TF kernel name.
*/
function getGradient(kernelName) {
return gradRegistry.get(kernelName);
}
function getKernelsForBackend(backendName) {
const it = kernelRegistry.entries();
const result = [];
while (true) {
const { done, value } = it.next();
if (done) {
break;
}
const [key, config] = value;
const [backend,] = key.split('_');
if (backend === backendName) {
result.push(config);
}
}
return result;
}
/**
* Registers the function (forward pass) for the kernel in a global registry.
*
* @param config A config object with the following properties:
* - `kernelName` The official name of the kernel.
* - `backendName` The official name of the backend.
* - `kernelFunc` The function to run during the forward pass of the kernel.
* - `setupFunc` Optional. Gets called once, after the backend initializes.
* - `disposeFunc` Optional. Gets called once, right before the backend is
* disposed.
*/
function registerKernel(config) {
const { kernelName, backendName } = config;
const key = makeKey(kernelName, backendName);
if (kernelRegistry.has(key)) {
warn(`The kernel '${kernelName}' for backend ` +
`'${backendName}' is already registered`);
}
kernelRegistry.set(key, config);
}
/**
* Registers a gradient function for a given kernel in the global registry,
* to be used during the back-propagation of that kernel.
*
* @param config An object with the following properties:
* - `kernelName` The name of the kernel that the gradient function is for.
* - `gradFunc` The function to run during back-propagation.
*/
function registerGradient(config) {
const { kernelName } = config;
if (gradRegistry.has(kernelName)) {
// TODO (yassogba) after 3.0 assess whether we need to keep this gated
// to debug mode.
if (env().getBool('DEBUG')) {
warn(`Overriding the gradient for '${kernelName}'`);
}
}
gradRegistry.set(kernelName, config);
}
function makeKey(kernelName, backendName) {
return `${backendName}_${kernelName}`;
}
// This env import is causing a circular dependency, which we are OK with
// We always set the platform to be PlatformBrowser
// We use a timeout to make sure the env is already initialized before setting the platform
setTimeout(() => env().setPlatform('browser', new PlatformStub()));
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function isTypedArrayBrowser(a) {
return a instanceof Float32Array || a instanceof Int32Array ||
a instanceof Uint8Array || a instanceof Uint8ClampedArray;
}
class PlatformStub {
constructor() {
}
fetch(path, init) {
throw new Error("fetch is not supported in this build.");
}
now() {
return performance.now();
}
encode(text, encoding) {
if (encoding !== 'utf-8' && encoding !== 'utf8') {
throw new Error(`Browser's encoder only supports utf-8, but got ${encoding}`);
}
if (this.textEncoder == null) {
this.textEncoder = new TextEncoder();
}
return this.textEncoder.encode(text);
}
decode(bytes, encoding) {
return new TextDecoder(encoding).decode(bytes);
}
setTimeoutCustom(functionRef, delay) {
if (typeof window === 'undefined' ||
!env().getBool('USE_SETTIMEOUTCUSTOM')) {
setTimeout(functionRef, delay);
return;
}
this.functionRefs.push(functionRef);
setTimeout(() => {
window.postMessage({name: this.messageName, index: this.functionRefs.length - 1}, location.origin);
}, delay);
if (!this.hasEventListener) {
this.hasEventListener = true;
window.addEventListener('message', (event) => {
if (event.source === window && event.data.name === this.messageName) {
event.stopPropagation();
const functionRef = this.functionRefs[event.data.index];
functionRef();
this.handledMessageCount++;
if (this.handledMessageCount === this.functionRefs.length) {
this.functionRefs = [];
this.handledMessageCount = 0;
}
}
}, true);
}
}
isTypedArray(a) {
return isTypedArrayBrowser(a)
}
}
var commonjsGlobal = typeof globalThis !== 'undefined' ? globalThis : typeof window !== 'undefined' ? window : typeof global !== 'undefined' ? global : typeof self !== 'undefined' ? self : {};
function getDefaultExportFromCjs (x) {
return x && x.__esModule && Object.prototype.hasOwnProperty.call(x, 'default') ? x['default'] : x;
}
var long = Long$1;
/**
* wasm optimizations, to do native i64 multiplication and divide
*/
var wasm = null;
try {
wasm = new WebAssembly.Instance(new WebAssembly.Module(new Uint8Array([
0, 97, 115, 109, 1, 0, 0, 0, 1, 13, 2, 96, 0, 1, 127, 96, 4, 127, 127, 127, 127, 1, 127, 3, 7, 6, 0, 1, 1, 1, 1, 1, 6, 6, 1, 127, 1, 65, 0, 11, 7, 50, 6, 3, 109, 117, 108, 0, 1, 5, 100, 105, 118, 95, 115, 0, 2, 5, 100, 105, 118, 95, 117, 0, 3, 5, 114, 101, 109, 95, 115, 0, 4, 5, 114, 101, 109, 95, 117, 0, 5, 8, 103, 101, 116, 95, 104, 105, 103, 104, 0, 0, 10, 191, 1, 6, 4, 0, 35, 0, 11, 36, 1, 1, 126, 32, 0, 173, 32, 1, 173, 66, 32, 134, 132, 32, 2, 173, 32, 3, 173, 66, 32, 134, 132, 126, 34, 4, 66, 32, 135, 167, 36, 0, 32, 4, 167, 11, 36, 1, 1, 126, 32, 0, 173, 32, 1, 173, 66, 32, 134, 132, 32, 2, 173, 32, 3, 173, 66, 32, 134, 132, 127, 34, 4, 66, 32, 135, 167, 36, 0, 32, 4, 167, 11, 36, 1, 1, 126, 32, 0, 173, 32, 1, 173, 66, 32, 134, 132, 32, 2, 173, 32, 3, 173, 66, 32, 134, 132, 128, 34, 4, 66, 32, 135, 167, 36, 0, 32, 4, 167, 11, 36, 1, 1, 126, 32, 0, 173, 32, 1, 173, 66, 32, 134, 132, 32, 2, 173, 32, 3, 173, 66, 32, 134, 132, 129, 34, 4, 66, 32, 135, 167, 36, 0, 32, 4, 167, 11, 36, 1, 1, 126, 32, 0, 173, 32, 1, 173, 66, 32, 134, 132, 32, 2, 173, 32, 3, 173, 66, 32, 134, 132, 130, 34, 4, 66, 32, 135, 167, 36, 0, 32, 4, 167, 11
])), {}).exports;
} catch (e) {
// no wasm support :(
}
/**
* Constructs a 64 bit two's-complement integer, given its low and high 32 bit values as *signed* integers.
* See the from* functions below for more convenient ways of constructing Longs.
* @exports Long
* @class A Long class for representing a 64 bit two's-complement integer value.
* @param {number} low The low (signed) 32 bits of the long
* @param {number} high The high (signed) 32 bits of the long
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @constructor
*/
function Long$1(low, high, unsigned) {
/**
* The low 32 bits as a signed value.
* @type {number}
*/
this.low = low | 0;
/**
* The high 32 bits as a signed value.
* @type {number}
*/
this.high = high | 0;
/**
* Whether unsigned or not.
* @type {boolean}
*/
this.unsigned = !!unsigned;
}
Object.defineProperty(Long$1.prototype, "__isLong__", { value: true });
/**
* @function
* @param {*} obj Object
* @returns {boolean}
* @inner
*/
function isLong(obj) {
return (obj && obj["__isLong__"]) === true;
}
/**
* Tests if the specified object is a Long.
* @function
* @param {*} obj Object
* @returns {boolean}
*/
Long$1.isLong = isLong;
/**
* A cache of the Long representations of small integer values.
* @type {!Object}
* @inner
*/
var INT_CACHE = {};
/**
* A cache of the Long representations of small unsigned integer values.
* @type {!Object}
* @inner
*/
var UINT_CACHE = {};
/**
* @param {number} value
* @param {boolean=} unsigned
* @returns {!Long}
* @inner
*/
function fromInt(value, unsigned) {
var obj, cachedObj, cache;
if (unsigned) {
value >>>= 0;
if (cache = (0 <= value && value < 256)) {
cachedObj = UINT_CACHE[value];
if (cachedObj)
return cachedObj;
}
obj = fromBits(value, (value | 0) < 0 ? -1 : 0, true);
if (cache)
UINT_CACHE[value] = obj;
return obj;
} else {
value |= 0;
if (cache = (-128 <= value && value < 128)) {
cachedObj = INT_CACHE[value];
if (cachedObj)
return cachedObj;
}
obj = fromBits(value, value < 0 ? -1 : 0, false);
if (cache)
INT_CACHE[value] = obj;
return obj;
}
}
/**
* Returns a Long representing the given 32 bit integer value.
* @function
* @param {number} value The 32 bit integer in question
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @returns {!Long} The corresponding Long value
*/
Long$1.fromInt = fromInt;
/**
* @param {number} value
* @param {boolean=} unsigned
* @returns {!Long}
* @inner
*/
function fromNumber(value, unsigned) {
if (isNaN(value))
return unsigned ? UZERO : ZERO;
if (unsigned) {
if (value < 0)
return UZERO;
if (value >= TWO_PWR_64_DBL)
return MAX_UNSIGNED_VALUE;
} else {
if (value <= -TWO_PWR_63_DBL)
return MIN_VALUE;
if (value + 1 >= TWO_PWR_63_DBL)
return MAX_VALUE;
}
if (value < 0)
return fromNumber(-value, unsigned).neg();
return fromBits((value % TWO_PWR_32_DBL) | 0, (value / TWO_PWR_32_DBL) | 0, unsigned);
}
/**
* Returns a Long representing the given value, provided that it is a finite number. Otherwise, zero is returned.
* @function
* @param {number} value The number in question
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @returns {!Long} The corresponding Long value
*/
Long$1.fromNumber = fromNumber;
/**
* @param {number} lowBits
* @param {number} highBits
* @param {boolean=} unsigned
* @returns {!Long}
* @inner
*/
function fromBits(lowBits, highBits, unsigned) {
return new Long$1(lowBits, highBits, unsigned);
}
/**
* Returns a Long representing the 64 bit integer that comes by concatenating the given low and high bits. Each is
* assumed to use 32 bits.
* @function
* @param {number} lowBits The low 32 bits
* @param {number} highBits The high 32 bits
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @returns {!Long} The corresponding Long value
*/
Long$1.fromBits = fromBits;
/**
* @function
* @param {number} base
* @param {number} exponent
* @returns {number}
* @inner
*/
var pow_dbl = Math.pow; // Used 4 times (4*8 to 15+4)
/**
* @param {string} str
* @param {(boolean|number)=} unsigned
* @param {number=} radix
* @returns {!Long}
* @inner
*/
function fromString(str, unsigned, radix) {
if (str.length === 0)
throw Error('empty string');
if (str === "NaN" || str === "Infinity" || str === "+Infinity" || str === "-Infinity")
return ZERO;
if (typeof unsigned === 'number') {
// For goog.math.long compatibility
radix = unsigned,
unsigned = false;
} else {
unsigned = !! unsigned;
}
radix = radix || 10;
if (radix < 2 || 36 < radix)
throw RangeError('radix');
var p;
if ((p = str.indexOf('-')) > 0)
throw Error('interior hyphen');
else if (p === 0) {
return fromString(str.substring(1), unsigned, radix).neg();
}
// Do several (8) digits each time through the loop, so as to
// minimize the calls to the very expensive emulated div.
var radixToPower = fromNumber(pow_dbl(radix, 8));
var result = ZERO;
for (var i = 0; i < str.length; i += 8) {
var size = Math.min(8, str.length - i),
value = parseInt(str.substring(i, i + size), radix);
if (size < 8) {
var power = fromNumber(pow_dbl(radix, size));
result = result.mul(power).add(fromNumber(value));
} else {
result = result.mul(radixToPower);
result = result.add(fromNumber(value));
}
}
result.unsigned = unsigned;
return result;
}
/**
* Returns a Long representation of the given string, written using the specified radix.
* @function
* @param {string} str The textual representation of the Long
* @param {(boolean|number)=} unsigned Whether unsigned or not, defaults to signed
* @param {number=} radix The radix in which the text is written (2-36), defaults to 10
* @returns {!Long} The corresponding Long value
*/
Long$1.fromString = fromString;
/**
* @function
* @param {!Long|number|string|!{low: number, high: number, unsigned: boolean}} val
* @param {boolean=} unsigned
* @returns {!Long}
* @inner
*/
function fromValue(val, unsigned) {
if (typeof val === 'number')
return fromNumber(val, unsigned);
if (typeof val === 'string')
return fromString(val, unsigned);
// Throws for non-objects, converts non-instanceof Long:
return fromBits(val.low, val.high, typeof unsigned === 'boolean' ? unsigned : val.unsigned);
}
/**
* Converts the specified value to a Long using the appropriate from* function for its type.
* @function
* @param {!Long|number|string|!{low: number, high: number, unsigned: boolean}} val Value
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @returns {!Long}
*/
Long$1.fromValue = fromValue;
// NOTE: the compiler should inline these constant values below and then remove these variables, so there should be
// no runtime penalty for these.
/**
* @type {number}
* @const
* @inner
*/
var TWO_PWR_16_DBL = 1 << 16;
/**
* @type {number}
* @const
* @inner
*/
var TWO_PWR_24_DBL = 1 << 24;
/**
* @type {number}
* @const
* @inner
*/
var TWO_PWR_32_DBL = TWO_PWR_16_DBL * TWO_PWR_16_DBL;
/**
* @type {number}
* @const
* @inner
*/
var TWO_PWR_64_DBL = TWO_PWR_32_DBL * TWO_PWR_32_DBL;
/**
* @type {number}
* @const
* @inner
*/
var TWO_PWR_63_DBL = TWO_PWR_64_DBL / 2;
/**
* @type {!Long}
* @const
* @inner
*/
var TWO_PWR_24 = fromInt(TWO_PWR_24_DBL);
/**
* @type {!Long}
* @inner
*/
var ZERO = fromInt(0);
/**
* Signed zero.
* @type {!Long}
*/
Long$1.ZERO = ZERO;
/**
* @type {!Long}
* @inner
*/
var UZERO = fromInt(0, true);
/**
* Unsigned zero.
* @type {!Long}
*/
Long$1.UZERO = UZERO;
/**
* @type {!Long}
* @inner
*/
var ONE = fromInt(1);
/**
* Signed one.
* @type {!Long}
*/
Long$1.ONE = ONE;
/**
* @type {!Long}
* @inner
*/
var UONE = fromInt(1, true);
/**
* Unsigned one.
* @type {!Long}
*/
Long$1.UONE = UONE;
/**
* @type {!Long}
* @inner
*/
var NEG_ONE = fromInt(-1);
/**
* Signed negative one.
* @type {!Long}
*/
Long$1.NEG_ONE = NEG_ONE;
/**
* @type {!Long}
* @inner
*/
var MAX_VALUE = fromBits(0xFFFFFFFF|0, 0x7FFFFFFF|0, false);
/**
* Maximum signed value.
* @type {!Long}
*/
Long$1.MAX_VALUE = MAX_VALUE;
/**
* @type {!Long}
* @inner
*/
var MAX_UNSIGNED_VALUE = fromBits(0xFFFFFFFF|0, 0xFFFFFFFF|0, true);
/**
* Maximum unsigned value.
* @type {!Long}
*/
Long$1.MAX_UNSIGNED_VALUE = MAX_UNSIGNED_VALUE;
/**
* @type {!Long}
* @inner
*/
var MIN_VALUE = fromBits(0, 0x80000000|0, false);
/**
* Minimum signed value.
* @type {!Long}
*/
Long$1.MIN_VALUE = MIN_VALUE;
/**
* @alias Long.prototype
* @inner
*/
var LongPrototype = Long$1.prototype;
/**
* Converts the Long to a 32 bit integer, assuming it is a 32 bit integer.
* @returns {number}
*/
LongPrototype.toInt = function toInt() {
return this.unsigned ? this.low >>> 0 : this.low;
};
/**
* Converts the Long to a the nearest floating-point representation of this value (double, 53 bit mantissa).
* @returns {number}
*/
LongPrototype.toNumber = function toNumber() {
if (this.unsigned)
return ((this.high >>> 0) * TWO_PWR_32_DBL) + (this.low >>> 0);
return this.high * TWO_PWR_32_DBL + (this.low >>> 0);
};
/**
* Converts the Long to a string written in the specified radix.
* @param {number=} radix Radix (2-36), defaults to 10
* @returns {string}
* @override
* @throws {RangeError} If `radix` is out of range
*/
LongPrototype.toString = function toString(radix) {
radix = radix || 10;
if (radix < 2 || 36 < radix)
throw RangeError('radix');
if (this.isZero())
return '0';
if (this.isNegative()) { // Unsigned Longs are never negative
if (this.eq(MIN_VALUE)) {
// We need to change the Long value before it can be negated, so we remove
// the bottom-most digit in this base and then recurse to do the rest.
var radixLong = fromNumber(radix),
div = this.div(radixLong),
rem1 = div.mul(radixLong).sub(this);
return div.toString(radix) + rem1.toInt().toString(radix);
} else
return '-' + this.neg().toString(radix);
}
// Do several (6) digits each time through the loop, so as to
// minimize the calls to the very expensive emulated div.
var radixToPower = fromNumber(pow_dbl(radix, 6), this.unsigned),
rem = this;
var result = '';
while (true) {
var remDiv = rem.div(radixToPower),
intval = rem.sub(remDiv.mul(radixToPower)).toInt() >>> 0,
digits = intval.toString(radix);
rem = remDiv;
if (rem.isZero())
return digits + result;
else {
while (digits.length < 6)
digits = '0' + digits;
result = '' + digits + result;
}
}
};
/**
* Gets the high 32 bits as a signed integer.
* @returns {number} Signed high bits
*/
LongPrototype.getHighBits = function getHighBits() {
return this.high;
};
/**
* Gets the high 32 bits as an unsigned integer.
* @returns {number} Unsigned high bits
*/
LongPrototype.getHighBitsUnsigned = function getHighBitsUnsigned() {
return this.high >>> 0;
};
/**
* Gets the low 32 bits as a signed integer.
* @returns {number} Signed low bits
*/
LongPrototype.getLowBits = function getLowBits() {
return this.low;
};
/**
* Gets the low 32 bits as an unsigned integer.
* @returns {number} Unsigned low bits
*/
LongPrototype.getLowBitsUnsigned = function getLowBitsUnsigned() {
return this.low >>> 0;
};
/**
* Gets the number of bits needed to represent the absolute value of this Long.
* @returns {number}
*/
LongPrototype.getNumBitsAbs = function getNumBitsAbs() {
if (this.isNegative()) // Unsigned Longs are never negative
return this.eq(MIN_VALUE) ? 64 : this.neg().getNumBitsAbs();
var val = this.high != 0 ? this.high : this.low;
for (var bit = 31; bit > 0; bit--)
if ((val & (1 << bit)) != 0)
break;
return this.high != 0 ? bit + 33 : bit + 1;
};
/**
* Tests if this Long's value equals zero.
* @returns {boolean}
*/
LongPrototype.isZero = function isZero() {
return this.high === 0 && this.low === 0;
};
/**
* Tests if this Long's value equals zero. This is an alias of {@link Long#isZero}.
* @returns {boolean}
*/
LongPrototype.eqz = LongPrototype.isZero;
/**
* Tests if this Long's value is negative.
* @returns {boolean}
*/
LongPrototype.isNegative = function isNegative() {
return !this.unsigned && this.high < 0;
};
/**
* Tests if this Long's value is positive.
* @returns {boolean}
*/
LongPrototype.isPositive = function isPositive() {
return this.unsigned || this.high >= 0;
};
/**
* Tests if this Long's value is odd.
* @returns {boolean}
*/
LongPrototype.isOdd = function isOdd() {
return (this.low & 1) === 1;
};
/**
* Tests if this Long's value is even.
* @returns {boolean}
*/
LongPrototype.isEven = function isEven() {
return (this.low & 1) === 0;
};
/**
* Tests if this Long's value equals the specified's.
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.equals = function equals(other) {
if (!isLong(other))
other = fromValue(other);
if (this.unsigned !== other.unsigned && (this.high >>> 31) === 1 && (other.high >>> 31) === 1)
return false;
return this.high === other.high && this.low === other.low;
};
/**
* Tests if this Long's value equals the specified's. This is an alias of {@link Long#equals}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.eq = LongPrototype.equals;
/**
* Tests if this Long's value differs from the specified's.
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.notEquals = function notEquals(other) {
return !this.eq(/* validates */ other);
};
/**
* Tests if this Long's value differs from the specified's. This is an alias of {@link Long#notEquals}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.neq = LongPrototype.notEquals;
/**
* Tests if this Long's value differs from the specified's. This is an alias of {@link Long#notEquals}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.ne = LongPrototype.notEquals;
/**
* Tests if this Long's value is less than the specified's.
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.lessThan = function lessThan(other) {
return this.comp(/* validates */ other) < 0;
};
/**
* Tests if this Long's value is less than the specified's. This is an alias of {@link Long#lessThan}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.lt = LongPrototype.lessThan;
/**
* Tests if this Long's value is less than or equal the specified's.
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.lessThanOrEqual = function lessThanOrEqual(other) {
return this.comp(/* validates */ other) <= 0;
};
/**
* Tests if this Long's value is less than or equal the specified's. This is an alias of {@link Long#lessThanOrEqual}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.lte = LongPrototype.lessThanOrEqual;
/**
* Tests if this Long's value is less than or equal the specified's. This is an alias of {@link Long#lessThanOrEqual}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.le = LongPrototype.lessThanOrEqual;
/**
* Tests if this Long's value is greater than the specified's.
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.greaterThan = function greaterThan(other) {
return this.comp(/* validates */ other) > 0;
};
/**
* Tests if this Long's value is greater than the specified's. This is an alias of {@link Long#greaterThan}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.gt = LongPrototype.greaterThan;
/**
* Tests if this Long's value is greater than or equal the specified's.
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.greaterThanOrEqual = function greaterThanOrEqual(other) {
return this.comp(/* validates */ other) >= 0;
};
/**
* Tests if this Long's value is greater than or equal the specified's. This is an alias of {@link Long#greaterThanOrEqual}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.gte = LongPrototype.greaterThanOrEqual;
/**
* Tests if this Long's value is greater than or equal the specified's. This is an alias of {@link Long#greaterThanOrEqual}.
* @function
* @param {!Long|number|string} other Other value
* @returns {boolean}
*/
LongPrototype.ge = LongPrototype.greaterThanOrEqual;
/**
* Compares this Long's value with the specified's.
* @param {!Long|number|string} other Other value
* @returns {number} 0 if they are the same, 1 if the this is greater and -1
* if the given one is greater
*/
LongPrototype.compare = function compare(other) {
if (!isLong(other))
other = fromValue(other);
if (this.eq(other))
return 0;
var thisNeg = this.isNegative(),
otherNeg = other.isNegative();
if (thisNeg && !otherNeg)
return -1;
if (!thisNeg && otherNeg)
return 1;
// At this point the sign bits are the same
if (!this.unsigned)
return this.sub(other).isNegative() ? -1 : 1;
// Both are positive if at least one is unsigned
return (other.high >>> 0) > (this.high >>> 0) || (other.high === this.high && (other.low >>> 0) > (this.low >>> 0)) ? -1 : 1;
};
/**
* Compares this Long's value with the specified's. This is an alias of {@link Long#compare}.
* @function
* @param {!Long|number|string} other Other value
* @returns {number} 0 if they are the same, 1 if the this is greater and -1
* if the given one is greater
*/
LongPrototype.comp = LongPrototype.compare;
/**
* Negates this Long's value.
* @returns {!Long} Negated Long
*/
LongPrototype.negate = function negate() {
if (!this.unsigned && this.eq(MIN_VALUE))
return MIN_VALUE;
return this.not().add(ONE);
};
/**
* Negates this Long's value. This is an alias of {@link Long#negate}.
* @function
* @returns {!Long} Negated Long
*/
LongPrototype.neg = LongPrototype.negate;
/**
* Returns the sum of this and the specified Long.
* @param {!Long|number|string} addend Addend
* @returns {!Long} Sum
*/
LongPrototype.add = function add(addend) {
if (!isLong(addend))
addend = fromValue(addend);
// Divide each number into 4 chunks of 16 bits, and then sum the chunks.
var a48 = this.high >>> 16;
var a32 = this.high & 0xFFFF;
var a16 = this.low >>> 16;
var a00 = this.low & 0xFFFF;
var b48 = addend.high >>> 16;
var b32 = addend.high & 0xFFFF;
var b16 = addend.low >>> 16;
var b00 = addend.low & 0xFFFF;
var c48 = 0, c32 = 0, c16 = 0, c00 = 0;
c00 += a00 + b00;
c16 += c00 >>> 16;
c00 &= 0xFFFF;
c16 += a16 + b16;
c32 += c16 >>> 16;
c16 &= 0xFFFF;
c32 += a32 + b32;
c48 += c32 >>> 16;
c32 &= 0xFFFF;
c48 += a48 + b48;
c48 &= 0xFFFF;
return fromBits((c16 << 16) | c00, (c48 << 16) | c32, this.unsigned);
};
/**
* Returns the difference of this and the specified Long.
* @param {!Long|number|string} subtrahend Subtrahend
* @returns {!Long} Difference
*/
LongPrototype.subtract = function subtract(subtrahend) {
if (!isLong(subtrahend))
subtrahend = fromValue(subtrahend);
return this.add(subtrahend.neg());
};
/**
* Returns the difference of this and the specified Long. This is an alias of {@link Long#subtract}.
* @function
* @param {!Long|number|string} subtrahend Subtrahend
* @returns {!Long} Difference
*/
LongPrototype.sub = LongPrototype.subtract;
/**
* Returns the product of this and the specified Long.
* @param {!Long|number|string} multiplier Multiplier
* @returns {!Long} Product
*/
LongPrototype.multiply = function multiply(multiplier) {
if (this.isZero())
return ZERO;
if (!isLong(multiplier))
multiplier = fromValue(multiplier);
// use wasm support if present
if (wasm) {
var low = wasm.mul(this.low,
this.high,
multiplier.low,
multiplier.high);
return fromBits(low, wasm.get_high(), this.unsigned);
}
if (multiplier.isZero())
return ZERO;
if (this.eq(MIN_VALUE))
return multiplier.isOdd() ? MIN_VALUE : ZERO;
if (multiplier.eq(MIN_VALUE))
return this.isOdd() ? MIN_VALUE : ZERO;
if (this.isNegative()) {
if (multiplier.isNegative())
return this.neg().mul(multiplier.neg());
else
return this.neg().mul(multiplier).neg();
} else if (multiplier.isNegative())
return this.mul(multiplier.neg()).neg();
// If both longs are small, use float multiplication
if (this.lt(TWO_PWR_24) && multiplier.lt(TWO_PWR_24))
return fromNumber(this.toNumber() * multiplier.toNumber(), this.unsigned);
// Divide each long into 4 chunks of 16 bits, and then add up 4x4 products.
// We can skip products that would overflow.
var a48 = this.high >>> 16;
var a32 = this.high & 0xFFFF;
var a16 = this.low >>> 16;
var a00 = this.low & 0xFFFF;
var b48 = multiplier.high >>> 16;
var b32 = multiplier.high & 0xFFFF;
var b16 = multiplier.low >>> 16;
var b00 = multiplier.low & 0xFFFF;
var c48 = 0, c32 = 0, c16 = 0, c00 = 0;
c00 += a00 * b00;
c16 += c00 >>> 16;
c00 &= 0xFFFF;
c16 += a16 * b00;
c32 += c16 >>> 16;
c16 &= 0xFFFF;
c16 += a00 * b16;
c32 += c16 >>> 16;
c16 &= 0xFFFF;
c32 += a32 * b00;
c48 += c32 >>> 16;
c32 &= 0xFFFF;
c32 += a16 * b16;
c48 += c32 >>> 16;
c32 &= 0xFFFF;
c32 += a00 * b32;
c48 += c32 >>> 16;
c32 &= 0xFFFF;
c48 += a48 * b00 + a32 * b16 + a16 * b32 + a00 * b48;
c48 &= 0xFFFF;
return fromBits((c16 << 16) | c00, (c48 << 16) | c32, this.unsigned);
};
/**
* Returns the product of this and the specified Long. This is an alias of {@link Long#multiply}.
* @function
* @param {!Long|number|string} multiplier Multiplier
* @returns {!Long} Product
*/
LongPrototype.mul = LongPrototype.multiply;
/**
* Returns this Long divided by the specified. The result is signed if this Long is signed or
* unsigned if this Long is unsigned.
* @param {!Long|number|string} divisor Divisor
* @returns {!Long} Quotient
*/
LongPrototype.divide = function divide(divisor) {
if (!isLong(divisor))
divisor = fromValue(divisor);
if (divisor.isZero())
throw Error('division by zero');
// use wasm support if present
if (wasm) {
// guard against signed division overflow: the largest
// negative number / -1 would be 1 larger than the largest
// positive number, due to two's complement.
if (!this.unsigned &&
this.high === -2147483648 &&
divisor.low === -1 && divisor.high === -1) {
// be consistent with non-wasm code path
return this;
}
var low = (this.unsigned ? wasm.div_u : wasm.div_s)(
this.low,
this.high,
divisor.low,
divisor.high
);
return fromBits(low, wasm.get_high(), this.unsigned);
}
if (this.isZero())
return this.unsigned ? UZERO : ZERO;
var approx, rem, res;
if (!this.unsigned) {
// This section is only relevant for signed longs and is derived from the
// closure library as a whole.
if (this.eq(MIN_VALUE)) {
if (divisor.eq(ONE) || divisor.eq(NEG_ONE))
return MIN_VALUE; // recall that -MIN_VALUE == MIN_VALUE
else if (divisor.eq(MIN_VALUE))
return ONE;
else {
// At this point, we have |other| >= 2, so |this/other| < |MIN_VALUE|.
var halfThis = this.shr(1);
approx = halfThis.div(divisor).shl(1);
if (approx.eq(ZERO)) {
return divisor.isNegative() ? ONE : NEG_ONE;
} else {
rem = this.sub(divisor.mul(approx));
res = approx.add(rem.div(divisor));
return res;
}
}
} else if (divisor.eq(MIN_VALUE))
return this.unsigned ? UZERO : ZERO;
if (this.isNegative()) {
if (divisor.isNegative())
return this.neg().div(divisor.neg());
return this.neg().div(divisor).neg();
} else if (divisor.isNegative())
return this.div(divisor.neg()).neg();
res = ZERO;
} else {
// The algorithm below has not been made for unsigned longs. It's therefore
// required to take special care of the MSB prior to running it.
if (!divisor.unsigned)
divisor = divisor.toUnsigned();
if (divisor.gt(this))
return UZERO;
if (divisor.gt(this.shru(1))) // 15 >>> 1 = 7 ; with divisor = 8 ; true
return UONE;
res = UZERO;
}
// Repeat the following until the remainder is less than other: find a
// floating-point that approximates remainder / other *from below*, add this
// into the result, and subtract it from the remainder. It is critical that
// the approximate value is less than or equal to the real value so that the
// remainder never becomes negative.
rem = this;
while (rem.gte(divisor)) {
// Approximate the result of division. This may be a little greater or
// smaller than the actual value.
approx = Math.max(1, Math.floor(rem.toNumber() / divisor.toNumber()));
// We will tweak the approximate result by changing it in the 48-th digit or
// the smallest non-fractional digit, whichever is larger.
var log2 = Math.ceil(Math.log(approx) / Math.LN2),
delta = (log2 <= 48) ? 1 : pow_dbl(2, log2 - 48),
// Decrease the approximation until it is smaller than the remainder. Note
// that if it is too large, the product overflows and is negative.
approxRes = fromNumber(approx),
approxRem = approxRes.mul(divisor);
while (approxRem.isNegative() || approxRem.gt(rem)) {
approx -= delta;
approxRes = fromNumber(approx, this.unsigned);
approxRem = approxRes.mul(divisor);
}
// We know the answer can't be zero... and actually, zero would cause
// infinite recursion since we would make no progress.
if (approxRes.isZero())
approxRes = ONE;
res = res.add(approxRes);
rem = rem.sub(approxRem);
}
return res;
};
/**
* Returns this Long divided by the specified. This is an alias of {@link Long#divide}.
* @function
* @param {!Long|number|string} divisor Divisor
* @returns {!Long} Quotient
*/
LongPrototype.div = LongPrototype.divide;
/**
* Returns this Long modulo the specified.
* @param {!Long|number|string} divisor Divisor
* @returns {!Long} Remainder
*/
LongPrototype.modulo = function modulo(divisor) {
if (!isLong(divisor))
divisor = fromValue(divisor);
// use wasm support if present
if (wasm) {
var low = (this.unsigned ? wasm.rem_u : wasm.rem_s)(
this.low,
this.high,
divisor.low,
divisor.high
);
return fromBits(low, wasm.get_high(), this.unsigned);
}
return this.sub(this.div(divisor).mul(divisor));
};
/**
* Returns this Long modulo the specified. This is an alias of {@link Long#modulo}.
* @function
* @param {!Long|number|string} divisor Divisor
* @returns {!Long} Remainder
*/
LongPrototype.mod = LongPrototype.modulo;
/**
* Returns this Long modulo the specified. This is an alias of {@link Long#modulo}.
* @function
* @param {!Long|number|string} divisor Divisor
* @returns {!Long} Remainder
*/
LongPrototype.rem = LongPrototype.modulo;
/**
* Returns the bitwise NOT of this Long.
* @returns {!Long}
*/
LongPrototype.not = function not() {
return fromBits(~this.low, ~this.high, this.unsigned);
};
/**
* Returns the bitwise AND of this Long and the specified.
* @param {!Long|number|string} other Other Long
* @returns {!Long}
*/
LongPrototype.and = function and(other) {
if (!isLong(other))
other = fromValue(other);
return fromBits(this.low & other.low, this.high & other.high, this.unsigned);
};
/**
* Returns the bitwise OR of this Long and the specified.
* @param {!Long|number|string} other Other Long
* @returns {!Long}
*/
LongPrototype.or = function or(other) {
if (!isLong(other))
other = fromValue(other);
return fromBits(this.low | other.low, this.high | other.high, this.unsigned);
};
/**
* Returns the bitwise XOR of this Long and the given one.
* @param {!Long|number|string} other Other Long
* @returns {!Long}
*/
LongPrototype.xor = function xor(other) {
if (!isLong(other))
other = fromValue(other);
return fromBits(this.low ^ other.low, this.high ^ other.high, this.unsigned);
};
/**
* Returns this Long with bits shifted to the left by the given amount.
* @param {number|!Long} numBits Number of bits
* @returns {!Long} Shifted Long
*/
LongPrototype.shiftLeft = function shiftLeft(numBits) {
if (isLong(numBits))
numBits = numBits.toInt();
if ((numBits &= 63) === 0)
return this;
else if (numBits < 32)
return fromBits(this.low << numBits, (this.high << numBits) | (this.low >>> (32 - numBits)), this.unsigned);
else
return fromBits(0, this.low << (numBits - 32), this.unsigned);
};
/**
* Returns this Long with bits shifted to the left by the given amount. This is an alias of {@link Long#shiftLeft}.
* @function
* @param {number|!Long} numBits Number of bits
* @returns {!Long} Shifted Long
*/
LongPrototype.shl = LongPrototype.shiftLeft;
/**
* Returns this Long with bits arithmetically shifted to the right by the given amount.
* @param {number|!Long} numBits Number of bits
* @returns {!Long} Shifted Long
*/
LongPrototype.shiftRight = function shiftRight(numBits) {
if (isLong(numBits))
numBits = numBits.toInt();
if ((numBits &= 63) === 0)
return this;
else if (numBits < 32)
return fromBits((this.low >>> numBits) | (this.high << (32 - numBits)), this.high >> numBits, this.unsigned);
else
return fromBits(this.high >> (numBits - 32), this.high >= 0 ? 0 : -1, this.unsigned);
};
/**
* Returns this Long with bits arithmetically shifted to the right by the given amount. This is an alias of {@link Long#shiftRight}.
* @function
* @param {number|!Long} numBits Number of bits
* @returns {!Long} Shifted Long
*/
LongPrototype.shr = LongPrototype.shiftRight;
/**
* Returns this Long with bits logically shifted to the right by the given amount.
* @param {number|!Long} numBits Number of bits
* @returns {!Long} Shifted Long
*/
LongPrototype.shiftRightUnsigned = function shiftRightUnsigned(numBits) {
if (isLong(numBits))
numBits = numBits.toInt();
numBits &= 63;
if (numBits === 0)
return this;
else {
var high = this.high;
if (numBits < 32) {
var low = this.low;
return fromBits((low >>> numBits) | (high << (32 - numBits)), high >>> numBits, this.unsigned);
} else if (numBits === 32)
return fromBits(high, 0, this.unsigned);
else
return fromBits(high >>> (numBits - 32), 0, this.unsigned);
}
};
/**
* Returns this Long with bits logically shifted to the right by the given amount. This is an alias of {@link Long#shiftRightUnsigned}.
* @function
* @param {number|!Long} numBits Number of bits
* @returns {!Long} Shifted Long
*/
LongPrototype.shru = LongPrototype.shiftRightUnsigned;
/**
* Returns this Long with bits logically shifted to the right by the given amount. This is an alias of {@link Long#shiftRightUnsigned}.
* @function
* @param {number|!Long} numBits Number of bits
* @returns {!Long} Shifted Long
*/
LongPrototype.shr_u = LongPrototype.shiftRightUnsigned;
/**
* Converts this Long to signed.
* @returns {!Long} Signed long
*/
LongPrototype.toSigned = function toSigned() {
if (!this.unsigned)
return this;
return fromBits(this.low, this.high, false);
};
/**
* Converts this Long to unsigned.
* @returns {!Long} Unsigned long
*/
LongPrototype.toUnsigned = function toUnsigned() {
if (this.unsigned)
return this;
return fromBits(this.low, this.high, true);
};
/**
* Converts this Long to its byte representation.
* @param {boolean=} le Whether little or big endian, defaults to big endian
* @returns {!Array.<number>} Byte representation
*/
LongPrototype.toBytes = function toBytes(le) {
return le ? this.toBytesLE() : this.toBytesBE();
};
/**
* Converts this Long to its little endian byte representation.
* @returns {!Array.<number>} Little endian byte representation
*/
LongPrototype.toBytesLE = function toBytesLE() {
var hi = this.high,
lo = this.low;
return [
lo & 0xff,
lo >>> 8 & 0xff,
lo >>> 16 & 0xff,
lo >>> 24 ,
hi & 0xff,
hi >>> 8 & 0xff,
hi >>> 16 & 0xff,
hi >>> 24
];
};
/**
* Converts this Long to its big endian byte representation.
* @returns {!Array.<number>} Big endian byte representation
*/
LongPrototype.toBytesBE = function toBytesBE() {
var hi = this.high,
lo = this.low;
return [
hi >>> 24 ,
hi >>> 16 & 0xff,
hi >>> 8 & 0xff,
hi & 0xff,
lo >>> 24 ,
lo >>> 16 & 0xff,
lo >>> 8 & 0xff,
lo & 0xff
];
};
/**
* Creates a Long from its byte representation.
* @param {!Array.<number>} bytes Byte representation
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @param {boolean=} le Whether little or big endian, defaults to big endian
* @returns {Long} The corresponding Long value
*/
Long$1.fromBytes = function fromBytes(bytes, unsigned, le) {
return le ? Long$1.fromBytesLE(bytes, unsigned) : Long$1.fromBytesBE(bytes, unsigned);
};
/**
* Creates a Long from its little endian byte representation.
* @param {!Array.<number>} bytes Little endian byte representation
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @returns {Long} The corresponding Long value
*/
Long$1.fromBytesLE = function fromBytesLE(bytes, unsigned) {
return new Long$1(
bytes[0] |
bytes[1] << 8 |
bytes[2] << 16 |
bytes[3] << 24,
bytes[4] |
bytes[5] << 8 |
bytes[6] << 16 |
bytes[7] << 24,
unsigned
);
};
/**
* Creates a Long from its big endian byte representation.
* @param {!Array.<number>} bytes Big endian byte representation
* @param {boolean=} unsigned Whether unsigned or not, defaults to signed
* @returns {Long} The corresponding Long value
*/
Long$1.fromBytesBE = function fromBytesBE(bytes, unsigned) {
return new Long$1(
bytes[4] << 24 |
bytes[5] << 16 |
bytes[6] << 8 |
bytes[7],
bytes[0] << 24 |
bytes[1] << 16 |
bytes[2] << 8 |
bytes[3],
unsigned
);
};
var long$1 = /*@__PURE__*/getDefaultExportFromCjs(long);
var LongExports = /*#__PURE__*/_mergeNamespaces({
__proto__: null,
default: long$1
}, [long]);
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Workaround for allowing cjs module to be included in bundle created by
// rollup.
// tslint:disable-next-line
const Long =
// tslint:disable-next-line
long$1 || LongExports;
function hexToLong(hex) {
return Long.fromString(hex, true, 16);
}
// Some primes between 2^63 and 2^64 for various uses.
// Hex 0xc3a5c85c97cb3127
const k0 = hexToLong('c3a5c85c97cb3127');
// Hex 0xb492b66fbe98f273
const k1 = hexToLong('b492b66fbe98f273');
// Hex 0x9ae16a3b2f90404f
const k2 = hexToLong('9ae16a3b2f90404f');
function shiftMix(val) {
return val.xor(val.shru(47));
}
function fetch(s, offset, numBytes) {
const bytes = s.slice(offset, offset + numBytes);
return Long.fromBytes(Array.from(bytes), true, true);
}
function fetch64(s, offset) {
return fetch(s, offset, 8);
}
function fetch32(s, offset) {
return fetch(s, offset, 4);
}
function rotate64(val, shift) {
// Avoid shifting by 64: doing so yields an undefined result.
return shift === 0 ? val : val.shru(shift).or(val.shl(64 - shift));
}
function hashLen16(u, v, mul = hexToLong('9ddfea08eb382d69')) {
// Murmur-inspired hashing.
let a = u.xor(v).mul(mul);
a = a.xor(a.shru(47));
let b = v.xor(a).mul(mul);
b = b.xor(b.shru(47));
b = b.mul(mul);
return b;
}
// Return a 16-byte hash for 48 bytes. Quick and dirty.
// Callers do best to use "random-looking" values for a and b.
function weakHashLen32WithSeeds(w, x, y, z, a, b) {
a = a.add(w);
b = rotate64(b.add(a).add(z), 21);
const c = a;
a = a.add(x);
a = a.add(y);
b = b.add(rotate64(a, 44));
return [a.add(z), b.add(c)];
}
function weakHashLen32WithSeedsStr(s, offset, a, b) {
return weakHashLen32WithSeeds(fetch64(s, offset), fetch64(s, offset + 8), fetch64(s, offset + 16), fetch64(s, offset + 24), a, b);
}
function hashLen0to16(s, len = s.length) {
if (len >= 8) {
const mul = k2.add(len * 2);
const a = fetch64(s, 0).add(k2);
const b = fetch64(s, len - 8);
const c = rotate64(b, 37).mul(mul).add(a);
const d = rotate64(a, 25).add(b).mul(mul);
return hashLen16(c, d, mul);
}
if (len >= 4) {
const mul = k2.add(len * 2);
const a = fetch32(s, 0);
return hashLen16(a.shl(3).add(len), fetch32(s, len - 4), mul);
}
if (len > 0) {
const a = s[0];
const b = s[len >> 1];
const c = s[len - 1];
const y = a + (b << 8);
const z = len + (c << 2);
return shiftMix(k2.mul(y).xor(k0.mul(z))).mul(k2);
}
return k2;
}
function hashLen17to32(s, len = s.length) {
const mul = k2.add(len * 2);
const a = fetch64(s, 0).mul(k1);
const b = fetch64(s, 8);
const c = fetch64(s, len - 8).mul(mul);
const d = fetch64(s, len - 16).mul(k2);
return hashLen16(rotate64(a.add(b), 43).add(rotate64(c, 30)).add(d), a.add(rotate64(b.add(k2), 18)).add(c), mul);
}
function hashLen33to64(s, len = s.length) {
const mul = k2.add(len * 2);
const a = fetch64(s, 0).mul(k2);
const b = fetch64(s, 8);
const c = fetch64(s, len - 8).mul(mul);
const d = fetch64(s, len - 16).mul(k2);
const y = rotate64(a.add(b), 43).add(rotate64(c, 30)).add(d);
const z = hashLen16(y, a.add(rotate64(b.add(k2), 18)).add(c), mul);
const e = fetch64(s, 16).mul(mul);
const f = fetch64(s, 24);
const g = y.add(fetch64(s, len - 32)).mul(mul);
const h = z.add(fetch64(s, len - 24)).mul(mul);
return hashLen16(rotate64(e.add(f), 43).add(rotate64(g, 30)).add(h), e.add(rotate64(f.add(a), 18)).add(g), mul);
}
function fingerPrint64(s, len = s.length) {
const seed = Long.fromNumber(81, true);
if (len <= 32) {
if (len <= 16) {
return hashLen0to16(s, len);
}
else {
return hashLen17to32(s, len);
}
}
else if (len <= 64) {
return hashLen33to64(s, len);
}
// For strings over 64 bytes we loop. Internal state consists of
// 56 bytes: v, w, x, y, and z.
let x = seed;
let y = seed.mul(k1).add(113);
let z = shiftMix(y.mul(k2).add(113)).mul(k2);
let v = [Long.UZERO, Long.UZERO];
let w = [Long.UZERO, Long.UZERO];
x = x.mul(k2).add(fetch64(s, 0));
let offset = 0;
// Set end so that after the loop we have 1 to 64 bytes left to process.
const end = ((len - 1) >> 6) * 64;
const last64 = end + ((len - 1) & 63) - 63;
do {
x = rotate64(x.add(y).add(v[0]).add(fetch64(s, offset + 8)), 37).mul(k1);
y = rotate64(y.add(v[1]).add(fetch64(s, offset + 48)), 42).mul(k1);
x = x.xor(w[1]);
y = y.add(v[0]).add(fetch64(s, offset + 40));
z = rotate64(z.add(w[0]), 33).mul(k1);
v = weakHashLen32WithSeedsStr(s, offset, v[1].mul(k1), x.add(w[0]));
w = weakHashLen32WithSeedsStr(s, offset + 32, z.add(w[1]), y.add(fetch64(s, offset + 16)));
[z, x] = [x, z];
offset += 64;
} while (offset !== end);
const mul = k1.add(z.and(0xff).shl(1));
// Point to the last 64 bytes of input.
offset = last64;
w[0] = w[0].add((len - 1) & 63);
v[0] = v[0].add(w[0]);
w[0] = w[0].add(v[0]);
x = rotate64(x.add(y).add(v[0]).add(fetch64(s, offset + 8)), 37).mul(mul);
y = rotate64(y.add(v[1]).add(fetch64(s, offset + 48)), 42).mul(mul);
x = x.xor(w[1].mul(9));
y = y.add(v[0].mul(9).add(fetch64(s, offset + 40)));
z = rotate64(z.add(w[0]), 33).mul(mul);
v = weakHashLen32WithSeedsStr(s, offset, v[1].mul(mul), x.add(w[0]));
w = weakHashLen32WithSeedsStr(s, offset + 32, z.add(w[1]), y.add(fetch64(s, offset + 16)));
[z, x] = [x, z];
return hashLen16(hashLen16(v[0], w[0], mul).add(shiftMix(y).mul(k0)).add(z), hashLen16(v[1], w[1], mul).add(x), mul);
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Create typed array for scalar value. Used for storing in `DataStorage`.
*/
function createScalarValue(value, dtype) {
if (dtype === 'string') {
return encodeString(value);
}
return toTypedArray([value], dtype);
}
function noConversionNeeded(a, dtype) {
return (a instanceof Float32Array && dtype === 'float32') ||
(a instanceof Int32Array && dtype === 'int32') ||
(a instanceof Uint8Array && dtype === 'bool');
}
function toTypedArray(a, dtype) {
if (dtype === 'string') {
throw new Error('Cannot convert a string[] to a TypedArray');
}
if (Array.isArray(a)) {
a = flatten$1(a);
}
if (env().getBool('DEBUG')) {
checkConversionForErrors(a, dtype);
}
if (noConversionNeeded(a, dtype)) {
return a;
}
if (dtype == null || dtype === 'float32' || dtype === 'complex64') {
return new Float32Array(a);
}
else if (dtype === 'int32') {
return new Int32Array(a);
}
else if (dtype === 'bool') {
const bool = new Uint8Array(a.length);
for (let i = 0; i < bool.length; ++i) {
if (Math.round(a[i]) !== 0) {
bool[i] = 1;
}
}
return bool;
}
else {
throw new Error(`Unknown data type ${dtype}`);
}
}
/**
* Returns the current high-resolution time in milliseconds relative to an
* arbitrary time in the past. It works across different platforms (node.js,
* browsers).
*
* ```js
* console.log(tf.util.now());
* ```
*
* @doc {heading: 'Util', namespace: 'util'}
*/
function now() {
return env().platform.now();
}
/**
* Encodes the provided string into bytes using the provided encoding scheme.
*
* @param s The string to encode.
* @param encoding The encoding scheme. Defaults to utf-8.
*
* @doc {heading: 'Util'}
*/
function encodeString(s, encoding = 'utf-8') {
encoding = encoding || 'utf-8';
return env().platform.encode(s, encoding);
}
/**
* Decodes the provided bytes into a string using the provided encoding scheme.
* @param bytes The bytes to decode.
*
* @param encoding The encoding scheme. Defaults to utf-8.
*
* @doc {heading: 'Util'}
*/
function decodeString(bytes, encoding = 'utf-8') {
encoding = encoding || 'utf-8';
return env().platform.decode(bytes, encoding);
}
function isTypedArray(a) {
// TODO(mattsoulanille): Remove this fallback in 5.0.0
if (env().platform.isTypedArray != null) {
return env().platform.isTypedArray(a);
}
else {
return isTypedArrayBrowser(a);
}
}
// NOTE: We explicitly type out what T extends instead of any so that
// util.flatten on a nested array of number doesn't try to infer T as a
// number[][], causing us to explicitly type util.flatten<number>().
/**
* Flattens an arbitrarily nested array.
*
* ```js
* const a = [[1, 2], [3, 4], [5, [6, [7]]]];
* const flat = tf.util.flatten(a);
* console.log(flat);
* ```
*
* @param arr The nested array to flatten.
* @param result The destination array which holds the elements.
* @param skipTypedArray If true, avoids flattening the typed arrays. Defaults
* to false.
*
* @doc {heading: 'Util', namespace: 'util'}
*/
function flatten$1(arr, result = [], skipTypedArray = false) {
if (result == null) {
result = [];
}
if (typeof arr === 'boolean' || typeof arr === 'number' ||
typeof arr === 'string' || isPromise(arr) || arr == null ||
isTypedArray(arr) && skipTypedArray) {
result.push(arr);
}
else if (Array.isArray(arr) || isTypedArray(arr)) {
for (let i = 0; i < arr.length; ++i) {
flatten$1(arr[i], result, skipTypedArray);
}
}
else {
let maxIndex = -1;
for (const key of Object.keys(arr)) {
// 0 or positive integer.
if (/^([1-9]+[0-9]*|0)$/.test(key)) {
maxIndex = Math.max(maxIndex, Number(key));
}
}
for (let i = 0; i <= maxIndex; i++) {
// tslint:disable-next-line: no-unnecessary-type-assertion
flatten$1(arr[i], result, skipTypedArray);
}
}
return result;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Profiler {
constructor(backendTimer, logger) {
this.backendTimer = backendTimer;
this.logger = logger;
if (logger == null) {
this.logger = new Logger();
}
}
profileKernel(kernelName, inputs, f) {
let outputs;
const holdResultWrapperFn = () => {
outputs = f();
};
let timer;
const start = now();
if (this.backendTimer.timerAvailable()) {
timer = this.backendTimer.time(holdResultWrapperFn);
}
else {
holdResultWrapperFn();
for (const output of outputs) {
output.dataSync();
}
timer = Promise.resolve({ kernelMs: now() - start });
}
if (env().getBool('CHECK_COMPUTATION_FOR_ERRORS')) {
for (let i = 0; i < outputs.length; i++) {
const output = outputs[i];
// Dangling promise here because we don't want to propagate up
// asynchronicity.
output.data().then(tensorVals => {
checkComputationForErrors(tensorVals, output.dtype, kernelName);
});
}
}
const kernelProfile = {
kernelName,
outputs,
inputs,
timeMs: timer.then(timing => timing.kernelMs),
extraInfo: timer.then(timing => timing.getExtraProfileInfo != null ?
timing.getExtraProfileInfo() :
'')
};
return kernelProfile;
}
logKernelProfile(kernelProfile) {
const { kernelName, outputs, timeMs, inputs, extraInfo } = kernelProfile;
outputs.forEach(result => {
Promise.all([result.data(), timeMs, extraInfo]).then(valueContainer => {
this.logger.logKernelProfile(kernelName, result, valueContainer[0], valueContainer[1], inputs, valueContainer[2]);
});
});
}
}
function checkComputationForErrors(vals, dtype, kernelName) {
if (dtype !== 'float32') {
// Only floating point computations will generate NaN values
return false;
}
for (let i = 0; i < vals.length; i++) {
const num = vals[i];
if (isNaN(num) || !isFinite(num)) {
// Throwing custom exception so behavior is testable.
console.warn(`Found ${num} in the result of '${kernelName}'`);
return true;
}
}
return false;
}
class Logger {
logKernelProfile(name, result, vals, timeMs, inputs, extraInfo) {
const time = typeof timeMs === 'number' ? rightPad(`${timeMs}ms`, 9) :
timeMs['error'];
const paddedName = rightPad(name, 25);
const rank = result.rank;
const size = result.size;
const shape = rightPad(result.shape.toString(), 14);
let inputShapesDescription = '';
for (const name in inputs) {
const input = inputs[name];
if (input != null) {
// The input might be a non-tensor (e.g HTMLImageElement), in which case
// we claim the output shape as input shape.
const inputShape = input.shape || result.shape;
const inputRank = inputShape.length;
inputShapesDescription +=
`${name}: ${inputRank}D ${inputRank > 0 ? inputShape : ''} `;
}
}
console.log(`%c${paddedName}\t%c${time}\t%c${rank}D ${shape}\t%c${size}\t%c${inputShapesDescription}\t%c${extraInfo}`, 'font-weight:bold', 'color:red', 'color:blue', 'color: orange', 'color: green', 'color: steelblue');
}
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes a list of TapeNodes that connect x to y, filtering everything else
* out and preserving the order of the original tape elements.
*
* @param tape The tape elements to filter.
* @param xs The input Tensors.
* @param y The output Tensor.
*/
function getFilteredNodesXToY(tape, xs, y) {
// Forward pass to compute all the nodes and Tensors that are transitively a
// function of x.
const tensorsFromX = {};
const nodesFromX = {};
for (let i = 0; i < xs.length; i++) {
tensorsFromX[xs[i].id] = true;
}
for (let i = 0; i < tape.length; i++) {
const node = tape[i];
const nodeInputs = node.inputs;
for (const inputName in nodeInputs) {
const input = nodeInputs[inputName];
let anyInputFromX = false;
for (let j = 0; j < xs.length; j++) {
if (tensorsFromX[input.id]) {
node.outputs.forEach(output => tensorsFromX[output.id] = true);
anyInputFromX = true;
nodesFromX[node.id] = true;
break;
}
}
if (anyInputFromX) {
break;
}
}
}
// Backward pass to find all of the nodes and Tensors that lead to y.
const tensorsLeadToY = {};
tensorsLeadToY[y.id] = true;
const nodesToY = {};
for (let i = tape.length - 1; i >= 0; i--) {
const node = tape[i];
const nodeInputs = node.inputs;
// If any of the outputs lead to y, mark all of the inputs as leading to y.
for (let j = 0; j < node.outputs.length; j++) {
if (tensorsLeadToY[node.outputs[j].id]) {
for (const inputName in nodeInputs) {
tensorsLeadToY[nodeInputs[inputName].id] = true;
nodesToY[node.id] = true;
}
break;
}
}
}
// Return the paths that come from x and lead to y.
const filteredTape = [];
for (let i = 0; i < tape.length; i++) {
const node = tape[i];
if (nodesFromX[node.id] && nodesToY[node.id]) {
// Prune the inputs from the node that aren't a function of x.
const prunedInputs = {};
for (const inputName in node.inputs) {
const nodeInput = node.inputs[inputName];
if (tensorsFromX[nodeInput.id]) {
prunedInputs[inputName] = nodeInput;
}
}
// Copy the node and overwrite inputsAndArgs to the pruned version.
const prunedNode = Object.assign({}, node);
prunedNode.inputs = prunedInputs;
prunedNode.outputs = node.outputs;
filteredTape.push(prunedNode);
}
}
return filteredTape;
}
/**
* Backpropagate gradients through the filtered TapeNodes.
*
* @param tensorAccumulatedGradientMap A map of Tensor to its gradient. This map
* is mutated by this method.
* @param filteredTape The filtered TapeNodes to backprop through.
*/
function backpropagateGradients(tensorAccumulatedGradientMap, filteredTape, tidy, add) {
// Walk the tape backward and keep a map of Tensor to its gradient.
for (let i = filteredTape.length - 1; i >= 0; i--) {
const node = filteredTape[i];
const dys = [];
node.outputs.forEach(o => {
const gradTensor = tensorAccumulatedGradientMap[o.id];
if (gradTensor != null) {
dys.push(gradTensor);
}
else {
// This particular output is not in the back-propagation subgraph, so it
// does not affect the final output, thus we put null for its dy.
dys.push(null);
}
});
if (node.gradient == null) {
throw new Error(`Cannot compute gradient: gradient function not found ` +
`for ${node.kernelName}.`);
}
// Backprop dy through this node and accumulate gradients over the inputs.
const inputGradients = node.gradient(dys);
for (const inputName in node.inputs) {
if (!(inputName in inputGradients)) {
throw new Error(`Cannot backprop through input ${inputName}. ` +
`Available gradients found: ${Object.keys(inputGradients)}.`);
}
// Call the gradient function.
const dx = tidy(() => inputGradients[inputName]());
if (dx.dtype !== 'float32') {
throw new Error(`Error in gradient for op ${node.kernelName}. The gradient of input ` +
`${inputName} must have 'float32' dtype, but has '${dx.dtype}'`);
}
const x = node.inputs[inputName];
if (!arraysEqual(dx.shape, x.shape)) {
throw new Error(`Error in gradient for op ${node.kernelName}. The gradient of input ` +
`'${inputName}' has shape '${dx.shape}', which does not match ` +
`the shape of the input '${x.shape}'`);
}
if (tensorAccumulatedGradientMap[x.id] == null) {
tensorAccumulatedGradientMap[x.id] = dx;
}
else {
const curGradient = tensorAccumulatedGradientMap[x.id];
tensorAccumulatedGradientMap[x.id] = add(curGradient, dx);
curGradient.dispose();
}
}
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Maximum number of values before we decide to show ellipsis.
const FORMAT_LIMIT_NUM_VALS = 20;
// Number of first and last values to show when displaying a, b,...,y, z.
const FORMAT_NUM_FIRST_LAST_VALS = 3;
// Number of significant digits to show.
const FORMAT_NUM_SIG_DIGITS = 7;
function tensorToString(vals, shape, dtype, verbose) {
const strides = computeStrides(shape);
const padPerCol = computeMaxSizePerColumn(vals, shape, dtype, strides);
const rank = shape.length;
const valsLines = subTensorToString(vals, shape, dtype, strides, padPerCol);
const lines = ['Tensor'];
if (verbose) {
lines.push(` dtype: ${dtype}`);
lines.push(` rank: ${rank}`);
lines.push(` shape: [${shape}]`);
lines.push(` values:`);
}
lines.push(valsLines.map(l => ' ' + l).join('\n'));
return lines.join('\n');
}
function computeMaxSizePerColumn(vals, shape, dtype, strides) {
const n = sizeFromShape(shape);
const numCols = strides[strides.length - 1];
const padPerCol = new Array(numCols).fill(0);
const rank = shape.length;
const valuesOrTuples = dtype === 'complex64' ? createComplexTuples(vals) : vals;
if (rank > 1) {
for (let row = 0; row < n / numCols; row++) {
const offset = row * numCols;
for (let j = 0; j < numCols; j++) {
padPerCol[j] = Math.max(padPerCol[j], valToString(valuesOrTuples[offset + j], 0, dtype).length);
}
}
}
return padPerCol;
}
function valToString(val, pad, dtype) {
let valStr;
if (Array.isArray(val)) {
valStr = `${parseFloat(val[0].toFixed(FORMAT_NUM_SIG_DIGITS))} + ` +
`${parseFloat(val[1].toFixed(FORMAT_NUM_SIG_DIGITS))}j`;
}
else if (isString(val)) {
valStr = `'${val}'`;
}
else if (dtype === 'bool') {
valStr = boolNumToString(val);
}
else {
valStr = parseFloat(val.toFixed(FORMAT_NUM_SIG_DIGITS)).toString();
}
return rightPad(valStr, pad);
}
function boolNumToString(v) {
return v === 0 ? 'false' : 'true';
}
function subTensorToString(vals, shape, dtype, strides, padPerCol, isLast = true) {
const storagePerElement = dtype === 'complex64' ? 2 : 1;
const size = shape[0];
const rank = shape.length;
if (rank === 0) {
if (dtype === 'complex64') {
const complexTuple = createComplexTuples(vals);
return [valToString(complexTuple[0], 0, dtype)];
}
if (dtype === 'bool') {
return [boolNumToString(vals[0])];
}
return [vals[0].toString()];
}
if (rank === 1) {
if (size > FORMAT_LIMIT_NUM_VALS) {
const firstValsSize = FORMAT_NUM_FIRST_LAST_VALS * storagePerElement;
let firstVals = Array.from(vals.slice(0, firstValsSize));
let lastVals = Array.from(vals.slice((size - FORMAT_NUM_FIRST_LAST_VALS) * storagePerElement, size * storagePerElement));
if (dtype === 'complex64') {
firstVals = createComplexTuples(firstVals);
lastVals = createComplexTuples(lastVals);
}
return [
'[' +
firstVals.map((x, i) => valToString(x, padPerCol[i], dtype))
.join(', ') +
', ..., ' +
lastVals
.map((x, i) => valToString(x, padPerCol[size - FORMAT_NUM_FIRST_LAST_VALS + i], dtype))
.join(', ') +
']'
];
}
const displayVals = dtype === 'complex64' ? createComplexTuples(vals) :
Array.from(vals);
return [
'[' +
displayVals.map((x, i) => valToString(x, padPerCol[i], dtype))
.join(', ') +
']'
];
}
// The array is rank 2 or more.
const subshape = shape.slice(1);
const substrides = strides.slice(1);
const stride = strides[0] * storagePerElement;
const lines = [];
if (size > FORMAT_LIMIT_NUM_VALS) {
for (let i = 0; i < FORMAT_NUM_FIRST_LAST_VALS; i++) {
const start = i * stride;
const end = start + stride;
lines.push(...subTensorToString(vals.slice(start, end), subshape, dtype, substrides, padPerCol, false /* isLast */));
}
lines.push('...');
for (let i = size - FORMAT_NUM_FIRST_LAST_VALS; i < size; i++) {
const start = i * stride;
const end = start + stride;
lines.push(...subTensorToString(vals.slice(start, end), subshape, dtype, substrides, padPerCol, i === size - 1 /* isLast */));
}
}
else {
for (let i = 0; i < size; i++) {
const start = i * stride;
const end = start + stride;
lines.push(...subTensorToString(vals.slice(start, end), subshape, dtype, substrides, padPerCol, i === size - 1 /* isLast */));
}
}
const sep = rank === 2 ? ',' : '';
lines[0] = '[' + (size > 0 ? lines[0] + sep : '');
for (let i = 1; i < lines.length - 1; i++) {
lines[i] = ' ' + lines[i] + sep;
}
let newLineSep = ',\n';
for (let i = 2; i < rank; i++) {
newLineSep += '\n';
}
lines[lines.length - 1] =
' ' + lines[lines.length - 1] + ']' + (isLast ? '' : newLineSep);
return lines;
}
function createComplexTuples(vals) {
const complexTuples = [];
for (let i = 0; i < vals.length; i += 2) {
complexTuples.push([vals[i], vals[i + 1]]);
}
return complexTuples;
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Workaround for: https://github.com/bazelbuild/rules_nodejs/issues/1265
/// <reference types="@webgpu/types/dist" />
/**
* A mutable object, similar to `tf.Tensor`, that allows users to set values
* at locations before converting to an immutable `tf.Tensor`.
*
* See `tf.buffer` for creating a tensor buffer.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
class TensorBuffer {
constructor(shape, dtype, values) {
this.dtype = dtype;
this.shape = shape.slice();
this.size = sizeFromShape(shape);
if (values != null) {
const n = values.length;
assert$1(n === this.size, () => `Length of values '${n}' does not match the size ` +
`inferred by the shape '${this.size}'.`);
}
if (dtype === 'complex64') {
throw new Error(`complex64 dtype TensorBuffers are not supported. Please create ` +
`a TensorBuffer for the real and imaginary parts separately and ` +
`call tf.complex(real, imag).`);
}
this.values = values || getArrayFromDType(dtype, this.size);
this.strides = computeStrides(shape);
}
/**
* Sets a value in the buffer at a given location.
*
* @param value The value to set.
* @param locs The location indices.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
set(value, ...locs) {
if (locs.length === 0) {
locs = [0];
}
assert$1(locs.length === this.rank, () => `The number of provided coordinates (${locs.length}) must ` +
`match the rank (${this.rank})`);
const index = this.locToIndex(locs);
this.values[index] = value;
}
/**
* Returns the value in the buffer at the provided location.
*
* @param locs The location indices.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
get(...locs) {
if (locs.length === 0) {
locs = [0];
}
let i = 0;
for (const loc of locs) {
if (loc < 0 || loc >= this.shape[i]) {
const msg = `Requested out of range element at ${locs}. ` +
` Buffer shape=${this.shape}`;
throw new Error(msg);
}
i++;
}
let index = locs[locs.length - 1];
for (let i = 0; i < locs.length - 1; ++i) {
index += this.strides[i] * locs[i];
}
return this.values[index];
}
locToIndex(locs) {
if (this.rank === 0) {
return 0;
}
else if (this.rank === 1) {
return locs[0];
}
let index = locs[locs.length - 1];
for (let i = 0; i < locs.length - 1; ++i) {
index += this.strides[i] * locs[i];
}
return index;
}
indexToLoc(index) {
if (this.rank === 0) {
return [];
}
else if (this.rank === 1) {
return [index];
}
const locs = new Array(this.shape.length);
for (let i = 0; i < locs.length - 1; ++i) {
locs[i] = Math.floor(index / this.strides[i]);
index -= locs[i] * this.strides[i];
}
locs[locs.length - 1] = index;
return locs;
}
get rank() {
return this.shape.length;
}
/**
* Creates an immutable `tf.Tensor` object from the buffer.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
toTensor() {
return trackerFn().makeTensor(this.values, this.shape, this.dtype);
}
}
// For tracking tensor creation and disposal.
let trackerFn = null;
// Used by chaining methods to call into ops.
let opHandler$1 = null;
/**
* An external consumer can register itself as the tensor tracker. This way
* the Tensor class can notify the tracker for every tensor created and
* disposed.
*/
function setTensorTracker(fn) {
trackerFn = fn;
}
/**
* An external consumer can register itself as the op handler. This way the
* Tensor class can have chaining methods that call into ops via the op
* handler.
*/
function setOpHandler(handler) {
opHandler$1 = handler;
}
/**
* A `tf.Tensor` object represents an immutable, multidimensional array of
* numbers that has a shape and a data type.
*
* For performance reasons, functions that create tensors do not necessarily
* perform a copy of the data passed to them (e.g. if the data is passed as a
* `Float32Array`), and changes to the data will change the tensor. This is not
* a feature and is not supported. To avoid this behavior, use the tensor before
* changing the input data or create a copy with `copy = tf.add(yourTensor, 0)`.
*
* See `tf.tensor` for details on how to create a `tf.Tensor`.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
class Tensor {
constructor(shape, dtype, dataId, id) {
/** Whether this tensor has been globally kept. */
this.kept = false;
this.isDisposedInternal = false;
this.shape = shape.slice();
this.dtype = dtype || 'float32';
this.size = sizeFromShape(shape);
this.strides = computeStrides(shape);
this.dataId = dataId;
this.id = id;
this.rankType = (this.rank < 5 ? this.rank.toString() : 'higher');
}
get rank() {
return this.shape.length;
}
/**
* Returns a promise of `tf.TensorBuffer` that holds the underlying data.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
async buffer() {
const vals = await this.data();
return opHandler$1.buffer(this.shape, this.dtype, vals);
}
/**
* Returns a `tf.TensorBuffer` that holds the underlying data.
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
bufferSync() {
return opHandler$1.buffer(this.shape, this.dtype, this.dataSync());
}
/**
* Returns the tensor data as a nested array. The transfer of data is done
* asynchronously.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
async array() {
const vals = await this.data();
return toNestedArray(this.shape, vals, this.dtype === 'complex64');
}
/**
* Returns the tensor data as a nested array. The transfer of data is done
* synchronously.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
arraySync() {
return toNestedArray(this.shape, this.dataSync(), this.dtype === 'complex64');
}
/**
* Asynchronously downloads the values from the `tf.Tensor`. Returns a
* promise of `TypedArray` that resolves when the computation has finished.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
async data() {
this.throwIfDisposed();
const data = trackerFn().read(this.dataId);
if (this.dtype === 'string') {
const bytes = await data;
try {
return bytes.map(b => decodeString(b));
}
catch (_a) {
throw new Error('Failed to decode the string bytes into utf-8. ' +
'To get the original bytes, call tensor.bytes().');
}
}
return data;
}
/**
* Copy the tensor's data to a new GPU resource. Comparing to the `dataSync()`
* and `data()`, this method prevents data from being downloaded to CPU.
*
* For WebGL backend, the data will be stored on a densely packed texture.
* This means that the texture will use the RGBA channels to store value.
*
* For WebGPU backend, the data will be stored on a buffer. There is no
* parameter, so can not use a user-defined size to create the buffer.
*
* @param options:
* For WebGL,
* - customTexShape: Optional. If set, will use the user defined
* texture shape to create the texture.
*
* @returns For WebGL backend, a GPUData contains the new texture and
* its information.
* {
* tensorRef: The tensor that is associated with this texture,
* texture: WebGLTexture,
* texShape: [number, number] // [height, width]
* }
*
* For WebGPU backend, a GPUData contains the new buffer.
* {
* tensorRef: The tensor that is associated with this buffer,
* buffer: GPUBuffer,
* }
*
* Remember to dispose the GPUData after it is used by
* `res.tensorRef.dispose()`.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
dataToGPU(options) {
this.throwIfDisposed();
return trackerFn().readToGPU(this.dataId, options);
}
/**
* Synchronously downloads the values from the `tf.Tensor`. This blocks the
* UI thread until the values are ready, which can cause performance issues.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
dataSync() {
this.throwIfDisposed();
const data = trackerFn().readSync(this.dataId);
if (this.dtype === 'string') {
try {
return data.map(b => decodeString(b));
}
catch (_a) {
throw new Error('Failed to decode the string bytes into utf-8. ' +
'To get the original bytes, call tensor.bytes().');
}
}
return data;
}
/** Returns the underlying bytes of the tensor's data. */
async bytes() {
this.throwIfDisposed();
const data = await trackerFn().read(this.dataId);
if (this.dtype === 'string') {
return data;
}
else {
return new Uint8Array(data.buffer);
}
}
/**
* Disposes `tf.Tensor` from memory.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
dispose() {
if (this.isDisposed) {
return;
}
if (this.kerasMask) {
this.kerasMask.dispose();
}
trackerFn().disposeTensor(this);
this.isDisposedInternal = true;
}
get isDisposed() {
return this.isDisposedInternal;
}
throwIfDisposed() {
if (this.isDisposed) {
throw new Error(`Tensor is disposed.`);
}
}
/**
* Prints the `tf.Tensor`. See `tf.print` for details.
*
* @param verbose Whether to print verbose information about the tensor,
* including dtype and size.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
print(verbose = false) {
return opHandler$1.print(this, verbose);
}
/**
* Returns a copy of the tensor. See `tf.clone` for details.
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
clone() {
this.throwIfDisposed();
return opHandler$1.clone(this);
}
/**
* Returns a human-readable description of the tensor. Useful for logging.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
toString(verbose = false) {
const vals = this.dataSync();
return tensorToString(vals, this.shape, this.dtype, verbose);
}
cast(dtype) {
this.throwIfDisposed();
return opHandler$1.cast(this, dtype);
}
variable(trainable = true, name, dtype) {
this.throwIfDisposed();
return trackerFn().makeVariable(this, trainable, name, dtype);
}
}
Object.defineProperty(Tensor, Symbol.hasInstance, {
value: (instance) => {
// Implementation note: we should use properties of the object that will be
// defined before the constructor body has finished executing (methods).
// This is because when this code is transpiled by babel, babel will call
// classCallCheck before the constructor body is run.
// See https://github.com/tensorflow/tfjs/issues/3384 for backstory.
return !!instance && instance.data != null && instance.dataSync != null &&
instance.throwIfDisposed != null;
}
});
function getGlobalTensorClass() {
// Use getGlobal so that we can augment the Tensor class across package
// boundaries because the node resolution alg may result in different modules
// being returned for this file depending on the path they are loaded from.
return getGlobal('Tensor', () => {
return Tensor;
});
}
// Global side effect. Cache global reference to Tensor class
getGlobalTensorClass();
/**
* A mutable `tf.Tensor`, useful for persisting state, e.g. for training.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
class Variable extends Tensor {
constructor(initialValue, trainable, name, tensorId) {
super(initialValue.shape, initialValue.dtype, initialValue.dataId, tensorId);
this.trainable = trainable;
this.name = name;
}
/**
* Assign a new `tf.Tensor` to this variable. The new `tf.Tensor` must have
* the same shape and dtype as the old `tf.Tensor`.
*
* @param newValue New tensor to be assigned to this variable.
*
* @doc {heading: 'Tensors', subheading: 'Classes'}
*/
assign(newValue) {
if (newValue.dtype !== this.dtype) {
throw new Error(`dtype of the new value (${newValue.dtype}) and ` +
`previous value (${this.dtype}) must match`);
}
if (!arraysEqual(newValue.shape, this.shape)) {
throw new Error(`shape of the new value (${newValue.shape}) and ` +
`previous value (${this.shape}) must match`);
}
trackerFn().disposeTensor(this);
this.dataId = newValue.dataId;
trackerFn().incRef(this, null /* backend */);
}
dispose() {
trackerFn().disposeVariable(this);
this.isDisposedInternal = true;
}
}
Object.defineProperty(Variable, Symbol.hasInstance, {
value: (instance) => {
return instance instanceof Tensor && instance.assign != null &&
instance.assign instanceof Function;
}
});
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
var Rank;
(function (Rank) {
Rank["R0"] = "R0";
Rank["R1"] = "R1";
Rank["R2"] = "R2";
Rank["R3"] = "R3";
Rank["R4"] = "R4";
Rank["R5"] = "R5";
Rank["R6"] = "R6";
})(Rank || (Rank = {}));
// Looks for upcasting types. Used, for example, in operations with mixed dtype
// inputs.
var UpcastInt32AndMap;
(function (UpcastInt32AndMap) {
UpcastInt32AndMap["float32"] = "float32";
UpcastInt32AndMap["int32"] = "int32";
UpcastInt32AndMap["bool"] = "int32";
UpcastInt32AndMap["complex64"] = "complex64";
})(UpcastInt32AndMap || (UpcastInt32AndMap = {}));
var UpcastBoolAndMap;
(function (UpcastBoolAndMap) {
UpcastBoolAndMap["float32"] = "float32";
UpcastBoolAndMap["int32"] = "int32";
UpcastBoolAndMap["bool"] = "bool";
UpcastBoolAndMap["complex64"] = "complex64";
})(UpcastBoolAndMap || (UpcastBoolAndMap = {}));
var UpcastFloat32AndMap;
(function (UpcastFloat32AndMap) {
UpcastFloat32AndMap["float32"] = "float32";
UpcastFloat32AndMap["int32"] = "float32";
UpcastFloat32AndMap["bool"] = "float32";
UpcastFloat32AndMap["complex64"] = "complex64";
})(UpcastFloat32AndMap || (UpcastFloat32AndMap = {}));
var UpcastComplex64AndMap;
(function (UpcastComplex64AndMap) {
UpcastComplex64AndMap["float32"] = "complex64";
UpcastComplex64AndMap["int32"] = "complex64";
UpcastComplex64AndMap["bool"] = "complex64";
UpcastComplex64AndMap["complex64"] = "complex64";
})(UpcastComplex64AndMap || (UpcastComplex64AndMap = {}));
const upcastTypeMap = {
'float32': UpcastFloat32AndMap,
'int32': UpcastInt32AndMap,
'bool': UpcastBoolAndMap,
'complex64': UpcastComplex64AndMap
};
function upcastType(typeA, typeB) {
if (typeA === 'string' || typeB === 'string') {
if (typeA === 'string' && typeB === 'string') {
return 'string';
}
throw new Error(`Can not upcast ${typeA} with ${typeB}`);
}
return upcastTypeMap[typeA][typeB];
}
/** Returns the output type after summation. */
function sumOutType(type) {
return upcastType(type, 'int32');
}
function isWebGLData(values) {
return values != null && typeof values === 'object' && 'texture' in values &&
values.texture instanceof WebGLTexture;
}
function isWebGPUData(values) {
return typeof GPUBuffer !== 'undefined' && values != null &&
typeof values === 'object' && 'buffer' in values &&
values.buffer instanceof GPUBuffer;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function makeTypesMatch(a, b) {
if (a.dtype === b.dtype) {
return [a, b];
}
const dtype = upcastType(a.dtype, b.dtype);
return [a.cast(dtype), b.cast(dtype)];
}
/**
* Extracts any `Tensor`s found within the provided object.
*
* @param container an object that may be a `Tensor` or may directly contain
* `Tensor`s, such as a `Tensor[]` or `{key: Tensor, ...}`. In general it
* is safe to pass any object here, except that `Promise`s are not
* supported.
* @returns An array of `Tensors` found within the passed object. If the
* argument is simply a `Tensor', a list containing that `Tensor` is
* returned. If the object is not a `Tensor` or does not
* contain `Tensors`, an empty list is returned.
*/
function getTensorsInContainer(result) {
const list = [];
const seen = new Set();
walkTensorContainer(result, list, seen);
return list;
}
function walkTensorContainer(container, list, seen) {
if (container == null) {
return;
}
if (container instanceof Tensor) {
list.push(container);
return;
}
if (!isIterable(container)) {
return;
}
// Iteration over keys works also for arrays.
const iterable = container;
for (const k in iterable) {
const val = iterable[k];
if (!seen.has(val)) {
seen.add(val);
walkTensorContainer(val, list, seen);
}
}
}
// tslint:disable-next-line:no-any
function isIterable(obj) {
return Array.isArray(obj) || typeof obj === 'object';
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function isRegisteredKernelInvocation(kernelInvocation) {
return kernelInvocation.kernelName != null;
}
class EngineState {
constructor() {
// Public since optimizers will use it.
this.registeredVariables = {};
this.nextTapeNodeId = 0;
this.numBytes = 0;
this.numTensors = 0;
this.numStringTensors = 0;
this.numDataBuffers = 0;
// Number of nested tf.grad() statements when computing higher-order
// gradients. E.g. `1` for first-order gradients and `2` for second-order
// gradients. Used to track if the tape should be removed after a backprop.
this.gradientDepth = 0;
// Number of nested kernel calls. When kernel depth is greater than 1, we turn
// off the tape.
this.kernelDepth = 0;
this.scopeStack = [];
/**
* Keeps track of the number of data moves during a kernel execution. We
* maintain a stack since kernels can call other kernels, recursively.
*/
this.numDataMovesStack = [];
this.nextScopeId = 0;
this.tensorInfo = new WeakMap();
this.profiling = false;
this.activeProfile = {
newBytes: 0,
newTensors: 0,
peakBytes: 0,
kernels: [],
result: null,
get kernelNames() {
return Array.from(new Set(this.kernels.map(k => k.name)));
}
};
}
dispose() {
for (const variableName in this.registeredVariables) {
this.registeredVariables[variableName].dispose();
}
}
}
class Engine {
constructor(ENV) {
this.ENV = ENV;
this.registry = {};
this.registryFactory = {};
this.pendingBackendInitId = 0;
this.state = new EngineState();
}
async ready() {
if (this.pendingBackendInit != null) {
return this.pendingBackendInit.then(() => { });
}
if (this.backendInstance != null) {
return;
}
const sortedBackends = this.getSortedBackends();
for (let i = 0; i < sortedBackends.length; i++) {
const backendName = sortedBackends[i];
const success = await this.initializeBackend(backendName).success;
if (success) {
await this.setBackend(backendName);
return;
}
}
throw new Error(`Could not initialize any backends, all backend initializations ` +
`failed.`);
}
get backend() {
if (this.pendingBackendInit != null) {
throw new Error(`Backend '${this.backendName}' has not yet been initialized. Make ` +
`sure to await tf.ready() or await tf.setBackend() before calling ` +
`other methods`);
}
if (this.backendInstance == null) {
const { name, asyncInit } = this.initializeBackendsAndReturnBest();
if (asyncInit) {
throw new Error(`The highest priority backend '${name}' has not yet been ` +
`initialized. Make sure to await tf.ready() or ` +
`await tf.setBackend() before calling other methods`);
}
this.setBackend(name);
}
return this.backendInstance;
}
backendNames() {
return Object.keys(this.registryFactory);
}
findBackend(backendName) {
if (!(backendName in this.registry)) {
// If the backend hasn't been initialized but we have a registry entry for
// it, initialize it and return it.
if (backendName in this.registryFactory) {
const { asyncInit } = this.initializeBackend(backendName);
if (asyncInit) {
// Backend is not ready yet.
return null;
}
}
else {
return null;
}
}
return this.registry[backendName];
}
findBackendFactory(backendName) {
if (!(backendName in this.registryFactory)) {
return null;
}
return this.registryFactory[backendName].factory;
}
registerBackend(backendName, factory, priority = 1) {
if (backendName in this.registryFactory) {
warn(`${backendName} backend was already registered. ` +
`Reusing existing backend factory.`);
return false;
}
this.registryFactory[backendName] = { factory, priority };
return true;
}
async setBackend(backendName) {
if (this.registryFactory[backendName] == null) {
throw new Error(`Backend name '${backendName}' not found in registry`);
}
this.backendName = backendName;
if (this.registry[backendName] == null) {
this.backendInstance = null;
const { success, asyncInit } = this.initializeBackend(backendName);
const result = asyncInit ? await success : success;
if (!result) {
return false;
}
}
this.backendInstance = this.registry[backendName];
this.setupRegisteredKernels();
// Reset the profiler.
this.profiler = new Profiler(this.backendInstance);
return true;
}
setupRegisteredKernels() {
const kernels = getKernelsForBackend(this.backendName);
kernels.forEach(kernel => {
if (kernel.setupFunc != null) {
kernel.setupFunc(this.backendInstance);
}
});
}
disposeRegisteredKernels(backendName) {
const kernels = getKernelsForBackend(backendName);
kernels.forEach(kernel => {
if (kernel.disposeFunc != null) {
kernel.disposeFunc(this.registry[backendName]);
}
});
}
/**
* Initializes a backend by looking up the backend name in the factory
* registry and calling the factory method. Returns a boolean representing
* whether the initialization of the backend succeeded. Throws an error if
* there is no backend in the factory registry.
*/
initializeBackend(backendName) {
const registryFactoryEntry = this.registryFactory[backendName];
if (registryFactoryEntry == null) {
throw new Error(`Cannot initialize backend ${backendName}, no registration found.`);
}
try {
const backend = registryFactoryEntry.factory();
/* Test if the factory returns a promise.
Done in a more liberal way than
previous 'Promise.resolve(backend)===backend'
as we needed to account for custom Promise
implementations (e.g. Angular) */
if (backend && !(backend instanceof KernelBackend) &&
typeof backend.then === 'function') {
const promiseId = ++this.pendingBackendInitId;
const success = backend
.then(backendInstance => {
// Outdated promise. Another backend was set in the meantime.
if (promiseId < this.pendingBackendInitId) {
return false;
}
this.registry[backendName] = backendInstance;
this.pendingBackendInit = null;
return true;
})
.catch(err => {
// Outdated promise. Another backend was set in the meantime.
if (promiseId < this.pendingBackendInitId) {
return false;
}
this.pendingBackendInit = null;
warn(`Initialization of backend ${backendName} failed`);
warn(err.stack || err.message);
return false;
});
this.pendingBackendInit = success;
return { success, asyncInit: true };
}
else {
this.registry[backendName] = backend;
return { success: true, asyncInit: false };
}
}
catch (err) {
warn(`Initialization of backend ${backendName} failed`);
warn(err.stack || err.message);
return { success: false, asyncInit: false };
}
}
removeBackend(backendName) {
if (!(backendName in this.registryFactory)) {
throw new Error(`${backendName} backend not found in registry`);
}
if (this.backendName === backendName && this.pendingBackendInit != null) {
// There is a pending promise of the backend we want to remove. Make it
// obsolete.
this.pendingBackendInitId++;
}
if (backendName in this.registry) {
this.disposeRegisteredKernels(backendName);
this.registry[backendName].dispose();
delete this.registry[backendName];
}
delete this.registryFactory[backendName];
// Unset the backend if it is active.
if (this.backendName === backendName) {
this.pendingBackendInit = null;
this.backendName = null;
this.backendInstance = null;
}
}
getSortedBackends() {
if (Object.keys(this.registryFactory).length === 0) {
throw new Error('No backend found in registry.');
}
return Object.keys(this.registryFactory).sort((a, b) => {
// Highest priority comes first.
return this.registryFactory[b].priority -
this.registryFactory[a].priority;
});
}
initializeBackendsAndReturnBest() {
const sortedBackends = this.getSortedBackends();
for (let i = 0; i < sortedBackends.length; i++) {
const backendName = sortedBackends[i];
const { success, asyncInit } = this.initializeBackend(backendName);
if (asyncInit || success) {
return { name: backendName, asyncInit };
}
}
throw new Error(`Could not initialize any backends, all backend initializations ` +
`failed.`);
}
moveData(backend, dataId) {
const info = this.state.tensorInfo.get(dataId);
const srcBackend = info.backend;
const values = this.readSync(dataId);
const refCount = srcBackend.refCount(dataId);
// Delete the tensor from the old backend and move it to the new
// backend.
srcBackend.disposeData(dataId, true);
info.backend = backend;
backend.move(dataId, values, info.shape, info.dtype, refCount);
if (this.shouldCheckForMemLeaks()) {
// Track the number of moves during a kernel execution to correctly
// detect memory leaks.
this.state.numDataMovesStack[this.state.numDataMovesStack.length - 1]++;
}
}
tidy(nameOrFn, fn) {
let name = null;
if (fn == null) {
// Called with only 1 argument.
if (typeof nameOrFn !== 'function') {
throw new Error('Please provide a function to tidy()');
}
fn = nameOrFn;
}
else {
// Called with 2 arguments.
if (typeof nameOrFn !== 'string' && !(nameOrFn instanceof String)) {
throw new Error('When calling with two arguments, the first argument ' +
'to tidy() must be a string');
}
if (typeof fn !== 'function') {
throw new Error('When calling with two arguments, the 2nd argument ' +
'to tidy() must be a function');
}
name = nameOrFn;
// TODO(nsthorat,smilkov): Do operation logging and performance
// profiling.
}
let result;
return this.scopedRun(() => this.startScope(name), () => this.endScope(result), () => {
result = fn();
if (result instanceof Promise) {
console.error('Cannot return a Promise inside of tidy.');
}
return result;
});
}
scopedRun(start, end, f) {
start();
try {
const res = f();
end();
return res;
}
catch (ex) {
end();
throw ex;
}
}
nextTensorId() {
return Engine.nextTensorId++;
}
nextVariableId() {
return Engine.nextVariableId++;
}
/**
* This method is called instead of the public-facing tensor.clone() when
* saving a tensor for backwards pass. It makes sure to add the clone
* operation to the tape regardless of being called inside a kernel
* execution.
*/
clone(x) {
const y = ENGINE.runKernel(Identity$1, { x });
const inputs = { x };
const grad = (dy) => ({
x: () => {
const dtype = 'float32';
const gradInputs = { x: dy };
const attrs = { dtype };
return ENGINE.runKernel(Cast, gradInputs,
// tslint:disable-next-line: no-unnecessary-type-assertion
attrs);
}
});
const saved = [];
this.addTapeNode(this.state.activeScope.name, inputs, [y], grad, saved, {});
return y;
}
/**
* Execute a kernel with the given name and return the output tensor.
*
* @param kernelName The name of the kernel to execute.
* @param inputs A map of input names to tensors.
* @param attrs A map of attribute names to their values. An attribute is a
* primitive (non-tensor) input to the kernel.
* @param inputsToSave A list of tensors, inputs to save for the backprop
* computation.
* @param outputsToSave A list of booleans, specifying which output to save
* for the backprop computation. These are booleans since the output
* tensors are not visible to the user.
*/
runKernel(kernelName, inputs, attrs) {
const hasKernel = getKernel(kernelName, this.backendName) != null;
if (!hasKernel) {
throw new Error(`Kernel '${kernelName}' not registered for backend '${this.backendName}'`);
}
return this.runKernelFunc({ kernelName, inputs, attrs });
}
shouldCheckForMemLeaks() {
return this.ENV.getBool('IS_TEST');
}
checkKernelForMemLeak(kernelName, numDataIdsBefore, outInfos) {
const numDataIdsAfter = this.backend.numDataIds();
// Count the number of data ids associated with the result of the kernel.
let numOutputDataIds = 0;
outInfos.forEach(info => {
// Complex numbers allocate 3 data ids, one for 'real', one for
// 'imaginary', and one for the container that holds the former two.
numOutputDataIds += (info.dtype === 'complex64' ? 3 : 1);
});
// Account for the number of moves during kernel execution. A "data move"
// can happen in the middle of a kernel execution, placing a new (key,value)
// pair in the data storage. Since data moves have net zero effect (we
// always remove the data from the old backend), we have to cancel them out
// when detecting memory leaks.
const numMoves = this.state.numDataMovesStack[this.state.numDataMovesStack.length - 1];
const dataIdsLeaked = numDataIdsAfter - numDataIdsBefore - numOutputDataIds - numMoves;
if (dataIdsLeaked > 0) {
throw new Error(`Backend '${this.backendName}' has an internal memory leak ` +
`(${dataIdsLeaked} data ids) after running '${kernelName}'`);
}
}
/**
* Internal helper method to execute a kernel Func
*
* Use `runKernel` to execute kernels from outside of engine.
*/
runKernelFunc(kernelParams) {
let outputs;
let saved = [];
const isTapeOn = this.isTapeOn();
const startingBytecount = this.state.numBytes;
const startingNumTensors = this.state.numTensors;
if (this.shouldCheckForMemLeaks()) {
this.state.numDataMovesStack.push(0);
}
let kernelFunc;
let out;
const kernelOrScopeName = isRegisteredKernelInvocation(kernelParams) ?
kernelParams.kernelName :
this.state.activeScope != null ? this.state.activeScope.name : '';
// Create the kernelFunc from either a registered kernel OR passed in
// forward/backward functions (used by custom grad). In this context a
// kernelFunc wraps a kernel implementation with some bookkeeping.
if (isRegisteredKernelInvocation(kernelParams)) {
const { kernelName, inputs, attrs } = kernelParams;
const kernel = getKernel(kernelName, this.backendName);
assert$1(kernel != null, () => `Cannot find registered kernel '${kernelName}' for backend '${this.backendName}'`);
kernelFunc = () => {
const numDataIdsBefore = this.backend.numDataIds();
out = kernel.kernelFunc({ inputs, attrs, backend: this.backend });
const outInfos = Array.isArray(out) ? out : [out];
if (this.shouldCheckForMemLeaks()) {
this.checkKernelForMemLeak(kernelName, numDataIdsBefore, outInfos);
}
const outTensors = outInfos.map((outInfo) => {
// todo (yassogba) remove this option (Tensor) when node backend
// methods have been modularized and they all return tensorInfo.
// TensorInfos do not have a rank attribute.
if (outInfo.rank != null) {
return outInfo;
}
return this.makeTensorFromTensorInfo(outInfo);
});
// Save any required inputs and outputs.
// Do not save unless we are recording to the tape. Otherwise it would
// cause a mem leak since there would be no backprop for these tensors
// (which would otherwise dispose them).
if (isTapeOn) {
const tensorsToSave = this.getTensorsForGradient(kernelName, inputs, outTensors);
saved = this.saveTensorsForBackwardMode(tensorsToSave);
}
return outTensors;
};
}
else {
const { forwardFunc } = kernelParams;
// Running a customGrad op.
const saveFunc = (tensors) => {
// Do not save unless we are recording to the tape. Otherwise it would
// cause a mem leak since we would never run backprop, which disposes
// the kept tensors.
if (!isTapeOn) {
return;
}
saved = tensors.map(tensor => this.keep(this.clone(tensor)));
};
kernelFunc = () => {
const numDataIdsBefore = this.backend.numDataIds();
out = this.tidy(() => forwardFunc(this.backend, saveFunc));
const outs = (Array.isArray(out) ? out : [out]);
if (this.shouldCheckForMemLeaks()) {
// Scope name is used to print a more helpful error message if needed.
this.checkKernelForMemLeak(kernelOrScopeName, numDataIdsBefore, outs);
}
return outs;
};
}
//
// Run the kernelFunc. Optionally profiling it.
//
const { inputs, attrs } = kernelParams;
const backwardsFunc = isRegisteredKernelInvocation(kernelParams) ?
null :
kernelParams.backwardsFunc;
let kernelProfile;
this.scopedRun(
// Stop recording to a tape when running a kernel.
() => this.state.kernelDepth++, () => this.state.kernelDepth--, () => {
if (!this.ENV.getBool('DEBUG') && !this.state.profiling) {
outputs = kernelFunc();
}
else {
kernelProfile = this.profiler.profileKernel(kernelOrScopeName, inputs, () => kernelFunc());
if (this.ENV.getBool('DEBUG')) {
this.profiler.logKernelProfile(kernelProfile);
}
outputs = kernelProfile.outputs;
}
});
if (isTapeOn) {
this.addTapeNode(kernelOrScopeName, inputs, outputs, backwardsFunc, saved, attrs);
}
if (this.state.profiling) {
this.state.activeProfile.kernels.push({
name: kernelOrScopeName,
bytesAdded: this.state.numBytes - startingBytecount,
totalBytesSnapshot: this.state.numBytes,
tensorsAdded: this.state.numTensors - startingNumTensors,
totalTensorsSnapshot: this.state.numTensors,
inputShapes: Object.keys(inputs).map(key => inputs[key] != null ? inputs[key].shape : null),
outputShapes: outputs.map(item => item.shape),
kernelTimeMs: kernelProfile.timeMs,
extraInfo: kernelProfile.extraInfo
});
}
return (Array.isArray(out) ? outputs : outputs[0]);
}
/**
* Saves tensors used in forward mode for use in backward mode.
*
* @param tensors the list of tensors to save.
*/
saveTensorsForBackwardMode(tensors) {
const saved = tensors.map(tensor => this.keep(this.clone(tensor)));
return saved;
}
/**
* Returns a list of tensors to save for a given gradient calculation.
*
* @param kernelName name of kernel to look up gradient for.
* @param inputs a map of input tensors.
* @param outputs an array of output tensors from forward mode of kernel.
*/
getTensorsForGradient(kernelName, inputs, outputs) {
const gradConfig = getGradient(kernelName);
if (gradConfig != null) {
const inputsToSave = gradConfig.inputsToSave || [];
const outputsToSave = gradConfig.outputsToSave || [];
// If saveAllInputs is true, all inputs will be saved. Otherwise, inputs
// specified in inputsToSave will be saved.
let inputTensorsToSave;
if (gradConfig.saveAllInputs) {
assert$1(Array.isArray(inputs), () => 'saveAllInputs is true, expected inputs to be an array.');
inputTensorsToSave = Object.keys(inputs).map((key) => inputs[key]);
}
else {
inputTensorsToSave = inputsToSave.map((inputName) => inputs[inputName]);
}
const outputTensorsToSave = outputs.filter((_, i) => outputsToSave[i]);
return inputTensorsToSave.concat(outputTensorsToSave);
}
// We return an empty list rather than throw an error because the kernel we
// are looking up may not actually be relevant to backproping through the
// overall function
//
// See 'does not error if irrelevant (pruned) ops are missing grads' test
// in gradients_test.ts for an example.
return [];
}
/**
* Internal method used by public APIs for tensor creation. Makes a new
* tensor with the provided shape, dtype and values. It always
* creates a new data id and writes the values to the underlying backend.
*/
makeTensor(values, shape, dtype, backend) {
if (values == null) {
throw new Error('Values passed to engine.makeTensor() are null');
}
dtype = dtype || 'float32';
backend = backend || this.backend;
let backendVals = values;
if (dtype === 'string' && isString(values[0])) {
backendVals = values.map(d => encodeString(d));
}
const dataId = backend.write(backendVals, shape, dtype);
const t = new Tensor(shape, dtype, dataId, this.nextTensorId());
this.trackTensor(t, backend);
// Count bytes for string tensors.
if (dtype === 'string') {
const info = this.state.tensorInfo.get(dataId);
const newBytes = bytesFromStringArray(backendVals);
this.state.numBytes += newBytes - info.bytes;
info.bytes = newBytes;
}
return t;
}
/**
* Internal method used by backends. Makes a new tensor
* that is a wrapper around an existing data id. It doesn't create
* a new data id, only increments the ref count used in memory tracking.
* @deprecated
*/
makeTensorFromDataId(dataId, shape, dtype, backend) {
dtype = dtype || 'float32';
const tensorInfo = { dataId, shape, dtype };
return this.makeTensorFromTensorInfo(tensorInfo, backend);
}
/**
* Internal method used by backends. Makes a new tensor that is a wrapper
* around an existing data id in TensorInfo. It doesn't create a new data id,
* only increments the ref count used in memory tracking.
*/
makeTensorFromTensorInfo(tensorInfo, backend) {
const { dataId, shape, dtype } = tensorInfo;
const t = new Tensor(shape, dtype, dataId, this.nextTensorId());
this.trackTensor(t, backend);
return t;
}
makeVariable(initialValue, trainable = true, name, dtype) {
name = name || this.nextVariableId().toString();
if (dtype != null && dtype !== initialValue.dtype) {
initialValue = initialValue.cast(dtype);
}
const v = new Variable(initialValue, trainable, name, this.nextTensorId());
if (this.state.registeredVariables[v.name] != null) {
throw new Error(`Variable with name ${v.name} was already registered`);
}
this.state.registeredVariables[v.name] = v;
this.incRef(v, this.backend);
return v;
}
trackTensor(a, backend) {
this.state.numTensors++;
if (a.dtype === 'string') {
this.state.numStringTensors++;
}
// Bytes for complex numbers are counted by their components. Bytes for
// string tensors are counted when writing values.
let bytes = 0;
if (a.dtype !== 'complex64' && a.dtype !== 'string') {
bytes = a.size * bytesPerElement(a.dtype);
}
this.state.numBytes += bytes;
if (!this.state.tensorInfo.has(a.dataId)) {
this.state.numDataBuffers++;
this.state.tensorInfo.set(a.dataId, {
backend: backend || this.backend,
dtype: a.dtype,
shape: a.shape,
bytes
});
}
if (!(a instanceof Variable)) {
this.track(a);
}
}
// Track the tensor by dataId and increase the refCount for the dataId in the
// backend.
// TODO(pyu10055): This is currently used by makeVariable method, to increase
// refCount on the backend for the dataId. It can potentially be replaced with
// Identity op indead of calling backend directly.
incRef(a, backend) {
this.trackTensor(a, backend);
this.backend.incRef(a.dataId);
}
removeDataId(dataId, backend) {
if (this.state.tensorInfo.has(dataId) &&
this.state.tensorInfo.get(dataId).backend === backend) {
this.state.tensorInfo.delete(dataId);
this.state.numDataBuffers--;
}
}
disposeTensor(a) {
if (!this.state.tensorInfo.has(a.dataId)) {
return;
}
const info = this.state.tensorInfo.get(a.dataId);
this.state.numTensors--;
if (a.dtype === 'string') {
this.state.numStringTensors--;
this.state.numBytes -= info.bytes;
}
// Don't count bytes for complex numbers as they are counted by their
// components.
if (a.dtype !== 'complex64' && a.dtype !== 'string') {
const bytes = a.size * bytesPerElement(a.dtype);
this.state.numBytes -= bytes;
}
// Remove the reference to dataId if backend dispose the data successfully
if (info.backend.disposeData(a.dataId)) {
this.removeDataId(a.dataId, info.backend);
}
// TODO(nsthorat): Construct an error and save the stack trace for
// debugging when in debug mode. Creating a stack trace is too expensive
// to do unconditionally.
}
disposeVariables() {
for (const varName in this.state.registeredVariables) {
const v = this.state.registeredVariables[varName];
this.disposeVariable(v);
}
}
disposeVariable(v) {
this.disposeTensor(v);
if (this.state.registeredVariables[v.name] != null) {
delete this.state.registeredVariables[v.name];
}
}
memory() {
const info = this.backend.memory();
info.numTensors = this.state.numTensors;
info.numDataBuffers = this.state.numDataBuffers;
info.numBytes = this.state.numBytes;
if (this.state.numStringTensors > 0) {
info.unreliable = true;
if (info.reasons == null) {
info.reasons = [];
}
info.reasons.push('Memory usage by string tensors is approximate ' +
'(2 bytes per character)');
}
return info;
}
async profile(query) {
this.state.profiling = true;
const startBytes = this.state.numBytes;
const startNumTensors = this.state.numTensors;
this.state.activeProfile.kernels = [];
this.state.activeProfile.result = await query();
this.state.profiling = false;
this.state.activeProfile.peakBytes = Math.max(...this.state.activeProfile.kernels.map(d => d.totalBytesSnapshot));
this.state.activeProfile.newBytes = this.state.numBytes - startBytes;
this.state.activeProfile.newTensors =
this.state.numTensors - startNumTensors;
for (const kernel of this.state.activeProfile.kernels) {
kernel.kernelTimeMs = await kernel.kernelTimeMs;
kernel.extraInfo = await kernel.extraInfo;
}
return this.state.activeProfile;
}
isTapeOn() {
return this.state.gradientDepth > 0 && this.state.kernelDepth === 0;
}
addTapeNode(kernelName, inputs, outputs, gradientsFunc, saved, attrs) {
const tapeNode = { id: this.state.nextTapeNodeId++, kernelName, inputs, outputs, saved };
const gradConfig = getGradient(kernelName);
if (gradConfig != null) {
gradientsFunc = gradConfig.gradFunc;
}
if (gradientsFunc != null) {
tapeNode.gradient = (dys) => {
// TODO(smilkov): To optimize back-prop, pass dys that are not used in
// the backprop graph to the user as null instead of zeros
dys = dys.map((dy, i) => {
if (dy == null) {
const output = outputs[i];
const vals = makeZerosTypedArray(output.size, output.dtype);
return this.makeTensor(vals, output.shape, output.dtype);
}
return dy;
});
// Grad functions of ops with single outputs expect a dy, while ops
// with multiple outputs expect dys (array of dy).
return gradientsFunc(dys.length > 1 ? dys : dys[0], saved, attrs);
};
}
this.state.activeTape.push(tapeNode);
}
keep(result) {
result.kept = true;
return result;
}
startTape() {
if (this.state.gradientDepth === 0) {
this.state.activeTape = [];
}
this.state.gradientDepth++;
}
endTape() {
this.state.gradientDepth--;
}
/**
* Start a scope. Use this with endScope() to achieve the same functionality
* as scope() without the need for a function closure.
*/
startScope(name) {
const scopeInfo = {
track: [],
name: 'unnamed scope',
id: this.state.nextScopeId++
};
if (name) {
scopeInfo.name = name;
}
this.state.scopeStack.push(scopeInfo);
this.state.activeScope = scopeInfo;
}
/**
* End a scope. Use this with startScope() to achieve the same functionality
* as scope() without the need for a function closure.
*/
endScope(result) {
const tensorsToTrackInParent = getTensorsInContainer(result);
const tensorsToTrackInParentSet = new Set(tensorsToTrackInParent.map(t => t.id));
// Dispose the arrays tracked in this scope.
for (let i = 0; i < this.state.activeScope.track.length; i++) {
const tensor = this.state.activeScope.track[i];
if (!tensor.kept && !tensorsToTrackInParentSet.has(tensor.id)) {
tensor.dispose();
}
}
const oldScope = this.state.scopeStack.pop();
this.state.activeScope = this.state.scopeStack.length === 0 ?
null :
this.state.scopeStack[this.state.scopeStack.length - 1];
// Track the current result in the parent scope.
tensorsToTrackInParent.forEach(tensor => {
// Only track the tensor if was allocated in the inner scope and is not
// globally kept.
if (!tensor.kept && tensor.scopeId === oldScope.id) {
this.track(tensor);
}
});
}
/**
* Returns gradients of `f` with respect to each of the `xs`. The gradients
* returned are of the same length as `xs`, but some might be null if `f`
* was not a function of that `x`. It also takes optional dy to multiply the
* gradient, which defaults to `1`.
*/
gradients(f, xs, dy, allowNoGradients = false) {
assert$1(xs.length > 0, () => 'gradients() received an empty list of xs.');
if (dy != null && dy.dtype !== 'float32') {
throw new Error(`dy must have 'float32' dtype, but has '${dy.dtype}'`);
}
const y = this.scopedRun(() => this.startTape(), () => this.endTape(), () => this.tidy('forward', f));
assert$1(y instanceof Tensor, () => 'The result y returned by f() must be a tensor.');
// Filter out the nodes that don't connect x => y.
const filteredTape = getFilteredNodesXToY(this.state.activeTape, xs, y);
if (!allowNoGradients && filteredTape.length === 0 && xs.length > 0) {
throw new Error('Cannot compute gradient of y=f(x) with respect to x. Make sure ' +
'that the f you passed encloses all operations that lead from x ' +
'to y.');
}
return this.tidy('backward', () => {
const accumulatedGradientMap = {};
accumulatedGradientMap[y.id] = (dy == null) ? ones$1(y.shape) : dy;
// Backprop gradients through the filtered nodes.
backpropagateGradients(accumulatedGradientMap, filteredTape,
// Pass the tidy function to avoid circular dep with `tape.ts`.
f => this.tidy(f),
// Pass an add function to avoide a circular dep with `tape.ts`.
add$2);
const grads = xs.map(x => accumulatedGradientMap[x.id]);
if (this.state.gradientDepth === 0) {
// This means that we are not computing higher-order gradients
// and can clean up the tape.
this.state.activeTape.forEach(node => {
for (const tensor of node.saved) {
tensor.dispose();
}
});
this.state.activeTape = null;
}
return { value: y, grads };
});
}
customGrad(f) {
assert$1(isFunction(f), () => 'The f passed in customGrad(f) must be a function.');
return (...inputs) => {
assert$1(inputs.every(t => t instanceof Tensor), () => 'The args passed in customGrad(f)(x1, x2,...) must all be ' +
'tensors');
let res;
const inputMap = {};
inputs.forEach((input, i) => {
inputMap[i] = input;
});
const forwardFunc = (_, save) => {
res = f(...[...inputs, save]);
assert$1(res.value instanceof Tensor, () => 'The function f passed in customGrad(f) must return an ' +
'object where `obj.value` is a tensor');
assert$1(isFunction(res.gradFunc), () => 'The function f passed in customGrad(f) must return an ' +
'object where `obj.gradFunc` is a function.');
return res.value;
};
const backwardsFunc = (dy, saved) => {
const gradRes = res.gradFunc(dy, saved);
const grads = Array.isArray(gradRes) ? gradRes : [gradRes];
assert$1(grads.length === inputs.length, () => 'The function f passed in customGrad(f) must return an ' +
'object where `obj.gradFunc` is a function that returns ' +
'the same number of tensors as inputs passed to f(...).');
assert$1(grads.every(t => t instanceof Tensor), () => 'The function f passed in customGrad(f) must return an ' +
'object where `obj.gradFunc` is a function that returns ' +
'a list of only tensors.');
const gradMap = {};
grads.forEach((grad, i) => {
gradMap[i] = () => grad;
});
return gradMap;
};
return this.runKernelFunc({
forwardFunc,
backwardsFunc,
inputs: inputMap,
});
};
}
readSync(dataId) {
// Route the read to the correct backend.
const info = this.state.tensorInfo.get(dataId);
return info.backend.readSync(dataId);
}
read(dataId) {
// Route the read to the correct backend.
const info = this.state.tensorInfo.get(dataId);
return info.backend.read(dataId);
}
readToGPU(dataId, options) {
// Route the read to the correct backend.
const info = this.state.tensorInfo.get(dataId);
return info.backend.readToGPU(dataId, options);
}
async time(query) {
const start = now();
const timingInfo = await this.backend.time(query);
timingInfo.wallMs = now() - start;
return timingInfo;
}
/**
* Tracks a Tensor in the current scope to be automatically cleaned up
* when the current scope ends, and returns the value.
*
* @param result The Tensor to track in the current scope.
*/
track(result) {
if (this.state.activeScope != null) {
result.scopeId = this.state.activeScope.id;
this.state.activeScope.track.push(result);
}
return result;
}
get registeredVariables() {
return this.state.registeredVariables;
}
/**
* Resets the engine state. Removes all backends but does not remove
* registered backend factories.
*/
reset() {
// Make any pending promise obsolete.
this.pendingBackendInitId++;
this.state.dispose();
this.ENV.reset();
this.state = new EngineState();
for (const backendName in this.registry) {
this.disposeRegisteredKernels(backendName);
this.registry[backendName].dispose();
delete this.registry[backendName];
}
this.backendName = null;
this.backendInstance = null;
this.pendingBackendInit = null;
}
}
Engine.nextTensorId = 0;
Engine.nextVariableId = 0;
function ones$1(shape) {
const values = makeOnesTypedArray(sizeFromShape(shape), 'float32');
return ENGINE.makeTensor(values, shape, 'float32');
}
function getOrMakeEngine() {
const ns = getGlobalNamespace();
if (ns._tfengine == null) {
const environment = new Environment(ns);
ns._tfengine = new Engine(environment);
}
setEnvironmentGlobal(ns._tfengine.ENV);
// Tell the current tensor interface that the global engine is responsible
// for tracking.
setTensorTracker(() => ns._tfengine);
return ns._tfengine;
}
const ENGINE = getOrMakeEngine();
/**
* A implementation of the add op for use within engine and tape.
*
* This allows us to avoid a circular dependency between add.ts and engine.
* It is exported to be available in tape tests.
*/
function add$2(a, b) {
// We duplicate Add here to avoid a circular dependency with add.ts.
const inputs = { a, b };
return ENGINE.runKernel(Add, inputs);
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// tslint:disable-next-line:no-any
function _isNavigatorDefined() {
return typeof navigator !== 'undefined' && navigator != null;
}
function isMobile(nav) {
if (nav || _isNavigatorDefined()) {
if (!nav) {
nav = navigator;
}
if (nav.product === 'ReactNative') {
return true;
}
const a = nav.userAgent || nav.vendor ||
// tslint:disable-next-line:no-any
(typeof window !== 'undefined' ? window.opera : '');
// Use `navigator.userAgentData.mobile` as fallback.
if (!a) {
// tslint:disable-next-line:no-any
const navAny = nav;
return navAny.userAgentData && navAny.userAgentData.mobile;
}
// tslint:disable-next-line:max-line-length
return /(android|bb\d+|meego).+mobile|avantgo|bada\/|blackberry|blazer|compal|elaine|fennec|hiptop|iemobile|ip(hone|od)|iris|kindle|lge |maemo|midp|mmp|mobile.+firefox|netfront|opera m(ob|in)i|palm( os)?|phone|p(ixi|re)\/|plucker|pocket|psp|series(4|6)0|symbian|treo|up\.(browser|link)|vodafone|wap|windows ce|xda|xiino/i
.test(a) ||
// tslint:disable-next-line:max-line-length
/1207|6310|6590|3gso|4thp|50[1-6]i|770s|802s|a wa|abac|ac(er|oo|s\-)|ai(ko|rn)|al(av|ca|co)|amoi|an(ex|ny|yw)|aptu|ar(ch|go)|as(te|us)|attw|au(di|\-m|r |s )|avan|be(ck|ll|nq)|bi(lb|rd)|bl(ac|az)|br(e|v)w|bumb|bw\-(n|u)|c55\/|capi|ccwa|cdm\-|cell|chtm|cldc|cmd\-|co(mp|nd)|craw|da(it|ll|ng)|dbte|dc\-s|devi|dica|dmob|do(c|p)o|ds(12|\-d)|el(49|ai)|em(l2|ul)|er(ic|k0)|esl8|ez([4-7]0|os|wa|ze)|fetc|fly(\-|_)|g1 u|g560|gene|gf\-5|g\-mo|go(\.w|od)|gr(ad|un)|haie|hcit|hd\-(m|p|t)|hei\-|hi(pt|ta)|hp( i|ip)|hs\-c|ht(c(\-| |_|a|g|p|s|t)|tp)|hu(aw|tc)|i\-(20|go|ma)|i230|iac( |\-|\/)|ibro|idea|ig01|ikom|im1k|inno|ipaq|iris|ja(t|v)a|jbro|jemu|jigs|kddi|keji|kgt( |\/)|klon|kpt |kwc\-|kyo(c|k)|le(no|xi)|lg( g|\/(k|l|u)|50|54|\-[a-w])|libw|lynx|m1\-w|m3ga|m50\/|ma(te|ui|xo)|mc(01|21|ca)|m\-cr|me(rc|ri)|mi(o8|oa|ts)|mmef|mo(01|02|bi|de|do|t(\-| |o|v)|zz)|mt(50|p1|v )|mwbp|mywa|n10[0-2]|n20[2-3]|n30(0|2)|n50(0|2|5)|n7(0(0|1)|10)|ne((c|m)\-|on|tf|wf|wg|wt)|nok(6|i)|nzph|o2im|op(ti|wv)|oran|owg1|p800|pan(a|d|t)|pdxg|pg(13|\-([1-8]|c))|phil|pire|pl(ay|uc)|pn\-2|po(ck|rt|se)|prox|psio|pt\-g|qa\-a|qc(07|12|21|32|60|\-[2-7]|i\-)|qtek|r380|r600|raks|rim9|ro(ve|zo)|s55\/|sa(ge|ma|mm|ms|ny|va)|sc(01|h\-|oo|p\-)|sdk\/|se(c(\-|0|1)|47|mc|nd|ri)|sgh\-|shar|sie(\-|m)|sk\-0|sl(45|id)|sm(al|ar|b3|it|t5)|so(ft|ny)|sp(01|h\-|v\-|v )|sy(01|mb)|t2(18|50)|t6(00|10|18)|ta(gt|lk)|tcl\-|tdg\-|tel(i|m)|tim\-|t\-mo|to(pl|sh)|ts(70|m\-|m3|m5)|tx\-9|up(\.b|g1|si)|utst|v400|v750|veri|vi(rg|te)|vk(40|5[0-3]|\-v)|vm40|voda|vulc|vx(52|53|60|61|70|80|81|83|85|98)|w3c(\-| )|webc|whit|wi(g |nc|nw)|wmlb|wonu|x700|yas\-|your|zeto|zte\-/i
.test(a.substr(0, 4));
}
return false;
}
function isBrowser() {
return (typeof window !== 'undefined' && window.document != null) ||
//@ts-ignore
(typeof WorkerGlobalScope !== 'undefined');
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ENV$1 = env();
/**
* This file contains environment-related flag registrations.
*/
/** Whether to enable debug mode. */
ENV$1.registerFlag('DEBUG', () => false, debugValue => {
if (debugValue) {
console.warn('Debugging mode is ON. The output of every math call will ' +
'be downloaded to CPU and checked for NaNs. ' +
'This significantly impacts performance.');
}
});
/** Whether we are in a browser (as versus, say, node.js) environment. */
ENV$1.registerFlag('IS_BROWSER', () => isBrowser());
/** Whether we are in a browser (as versus, say, node.js) environment. */
ENV$1.registerFlag('IS_NODE', () => (typeof process !== 'undefined') &&
(typeof process.versions !== 'undefined') &&
(typeof process.versions.node !== 'undefined'));
/** Whether this browser is Chrome. */
ENV$1.registerFlag('IS_CHROME', () => typeof navigator !== 'undefined' && navigator != null &&
navigator.userAgent != null && /Chrome/.test(navigator.userAgent) &&
/Google Inc/.test(navigator.vendor));
/** Whether this browser is Safari. */
ENV$1.registerFlag('IS_SAFARI', () => typeof navigator !== 'undefined' && navigator != null &&
navigator.userAgent != null && /Safari/.test(navigator.userAgent) &&
/Apple/.test(navigator.vendor));
/**
* True when the environment is "production" where we disable safety checks
* to gain performance.
*/
ENV$1.registerFlag('PROD', () => false);
/**
* Whether to do sanity checks when inferring a shape from user-provided
* values, used when creating a new tensor.
*/
ENV$1.registerFlag('TENSORLIKE_CHECK_SHAPE_CONSISTENCY', () => ENV$1.getBool('DEBUG'));
/** Whether deprecation warnings are enabled. */
ENV$1.registerFlag('DEPRECATION_WARNINGS_ENABLED', () => true);
/** True if running unit tests. */
ENV$1.registerFlag('IS_TEST', () => false);
/** Whether to check computation result for errors. */
ENV$1.registerFlag('CHECK_COMPUTATION_FOR_ERRORS', () => ENV$1.getBool('DEBUG'));
/** Whether the backend needs to wrap input to imageBitmap. */
ENV$1.registerFlag('WRAP_TO_IMAGEBITMAP', () => false);
/** Whether to enable canvas2d willReadFrequently for GPU backends */
ENV$1.registerFlag('CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU', () => false);
/** Whether to use setTimeoutCustom */
ENV$1.registerFlag('USE_SETTIMEOUTCUSTOM', () => false);
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates an empty `tf.TensorBuffer` with the specified `shape` and `dtype`.
*
* The values are stored in CPU as `TypedArray`. Fill the buffer using
* `buffer.set()`, or by modifying directly `buffer.values`.
*
* When done, call `buffer.toTensor()` to get an immutable `tf.Tensor` with
* those values.
*
* ```js
* // Create a buffer and set values at particular indices.
* const buffer = tf.buffer([2, 2]);
* buffer.set(3, 0, 0);
* buffer.set(5, 1, 0);
*
* // Convert the buffer back to a tensor.
* buffer.toTensor().print();
* ```
*
* @param shape An array of integers defining the output tensor shape.
* @param dtype The dtype of the buffer. Defaults to 'float32'.
* @param values The values of the buffer as `TypedArray`. Defaults to
* zeros.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function buffer(shape, dtype = 'float32', values) {
dtype = dtype || 'float32';
assertNonNegativeIntegerDimensions(shape);
return new TensorBuffer(shape, dtype, values);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function inferShape(val, dtype) {
let firstElem = val;
if (isTypedArray(val)) {
return dtype === 'string' ? [] : [val.length];
}
if (isWebGLData(val)) {
const usedChannels = val.channels || 'RGBA';
return [val.height, val.width * usedChannels.length];
}
else if (isWebGPUData(val)) {
return [val.buffer.size / (dtype == null ? 4 : bytesPerElement(dtype))];
}
if (!Array.isArray(val)) {
return []; // Scalar.
}
const shape = [];
while (Array.isArray(firstElem) ||
isTypedArray(firstElem) && dtype !== 'string') {
shape.push(firstElem.length);
firstElem = firstElem[0];
}
if (Array.isArray(val) &&
env().getBool('TENSORLIKE_CHECK_SHAPE_CONSISTENCY')) {
deepAssertShapeConsistency(val, shape, []);
}
return shape;
}
function deepAssertShapeConsistency(val, shape, indices) {
indices = indices || [];
if (!(Array.isArray(val)) && !isTypedArray(val)) {
assert$1(shape.length === 0, () => `Element arr[${indices.join('][')}] is a primitive, ` +
`but should be an array/TypedArray of ${shape[0]} elements`);
return;
}
assert$1(shape.length > 0, () => `Element arr[${indices.join('][')}] should be a primitive, ` +
`but is an array of ${val.length} elements`);
assert$1(val.length === shape[0], () => `Element arr[${indices.join('][')}] should have ${shape[0]} ` +
`elements, but has ${val.length} elements`);
const subShape = shape.slice(1);
for (let i = 0; i < val.length; ++i) {
deepAssertShapeConsistency(val[i], subShape, indices.concat(i));
}
}
function assertDtype(expectedDtype, actualDType, argName, functionName) {
if (expectedDtype === 'string_or_numeric') {
return;
}
if (expectedDtype == null) {
throw new Error(`Expected dtype cannot be null.`);
}
if (expectedDtype !== 'numeric' && expectedDtype !== actualDType ||
expectedDtype === 'numeric' && actualDType === 'string') {
throw new Error(`Argument '${argName}' passed to '${functionName}' must ` +
`be ${expectedDtype} tensor, but got ${actualDType} tensor`);
}
}
function convertToTensor(x, argName, functionName, parseAsDtype = 'numeric') {
if (x instanceof getGlobalTensorClass()) {
assertDtype(parseAsDtype, x.dtype, argName, functionName);
return x;
}
let inferredDtype = inferDtype(x);
// If the user expects a bool/int/float, use that info to update the
// inferredDtype when it is not a string.
if (inferredDtype !== 'string' &&
['bool', 'int32', 'float32'].indexOf(parseAsDtype) >= 0) {
inferredDtype = parseAsDtype;
}
assertDtype(parseAsDtype, inferredDtype, argName, functionName);
if ((x == null) ||
(!isTypedArray(x) && !Array.isArray(x) && typeof x !== 'number' &&
typeof x !== 'boolean' && typeof x !== 'string')) {
const type = x == null ? 'null' : x.constructor.name;
throw new Error(`Argument '${argName}' passed to '${functionName}' must be a ` +
`Tensor or TensorLike, but got '${type}'`);
}
const inferredShape = inferShape(x, inferredDtype);
if (!isTypedArray(x) && !Array.isArray(x)) {
x = [x];
}
const skipTypedArray = true;
const values = inferredDtype !== 'string' ?
toTypedArray(x, inferredDtype) :
flatten$1(x, [], skipTypedArray);
return ENGINE.makeTensor(values, inferredShape, inferredDtype);
}
function convertToTensorArray(arg, argName, functionName, parseAsDtype = 'numeric') {
if (!Array.isArray(arg)) {
throw new Error(`Argument ${argName} passed to ${functionName} must be a ` +
'`Tensor[]` or `TensorLike[]`');
}
const tensors = arg;
return tensors.map((t, i) => convertToTensor(t, `${argName}[${i}]`, functionName, parseAsDtype));
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const OP_SCOPE_SUFFIX = '__op';
/**
* Used for wrapping functions that perform math operations on
* Tensors. The function will be wrapped in a named scope that cleans all
* memory usage after the function is done.
*/
function op(f) {
const keys = Object.keys(f);
if (keys.length !== 1) {
throw new Error(`Please provide an object with a single key ` +
`(operation name) mapping to a function. Got an object with ` +
`${keys.length} keys.`);
}
let opName = keys[0];
const fn = f[opName];
// Strip the underscore from the end of the function name.
if (opName.endsWith('_')) {
opName = opName.substring(0, opName.length - 1);
}
// add an __op suffix to distinguish ops from kernels in tf.profile
opName = opName + OP_SCOPE_SUFFIX;
// tslint:disable-next-line:no-any
const f2 = (...args) => {
ENGINE.startScope(opName);
try {
const result = fn(...args);
if (isPromise(result)) {
console.error('Cannot return a Promise inside of tidy.');
}
ENGINE.endScope(result);
return result;
}
catch (ex) {
ENGINE.endScope(null);
throw ex;
}
};
Object.defineProperty(f2, 'name', { value: opName, configurable: true });
// tslint:disable-next-line:no-any
return f2;
}
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Casts a `tf.Tensor` to a new dtype.
*
* ```js
* const x = tf.tensor1d([1.5, 2.5, 3]);
* tf.cast(x, 'int32').print();
* ```
* @param x The input tensor to be casted.
* @param dtype The dtype to cast the input tensor to.
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function cast_(x, dtype) {
const $x = convertToTensor(x, 'x', 'cast');
// Sanity checks.
if (!isValidDtype(dtype)) {
throw new Error(`Failed to cast to unknown dtype ${dtype}`);
}
if (dtype === 'string' && $x.dtype !== 'string' ||
dtype !== 'string' && $x.dtype === 'string') {
throw new Error('Only strings can be casted to strings');
}
const inputs = { x: $x };
const attrs = { dtype };
return ENGINE.runKernel(Cast, inputs, attrs);
}
const cast$3 = /* @__PURE__ */ op({ cast_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a new tensor with the same values and shape as the specified
* tensor.
*
* ```js
* const x = tf.tensor([1, 2]);
*
* x.clone().print();
* ```
*
* @param x The tensor to clone.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function clone_(x) {
const $x = convertToTensor(x, 'x', 'clone', 'string_or_numeric');
const inputs = { x: $x };
// Note this op is called tf.identity in python. Hence the kernel name used
// here.
return ENGINE.runKernel(Identity$1, inputs);
}
const clone = /* @__PURE__ */ op({ clone_ });
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Prints information about the `tf.Tensor` including its data.
*
* ```js
* const verbose = true;
* tf.tensor2d([1, 2, 3, 4], [2, 2]).print(verbose);
* ```
* @param x The tensor to be printed.
* @param verbose Whether to print verbose information about the ` Tensor`,
* including dtype and size.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function print(x, verbose = false) {
console.log(x.toString(verbose));
}
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Required side effectful code for tfjs-core
// Set up Engine and ENV
getOrMakeEngine();
const opHandler = {
buffer,
cast: cast$3,
clone,
print
};
setOpHandler(opHandler);
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Enables production mode which disables correctness checks in favor of
* performance.
*
* @doc {heading: 'Environment'}
*/
function enableProdMode() {
env().set('PROD', true);
}
/**
* It returns the global engine that keeps track of all tensors and backends.
*
* @doc {heading: 'Environment'}
*/
function engine() {
return ENGINE;
}
/**
* Returns memory info at the current time in the program. The result is an
* object with the following properties:
*
* - `numBytes`: Number of bytes allocated (undisposed) at this time.
* - `numTensors`: Number of unique tensors allocated.
* - `numDataBuffers`: Number of unique data buffers allocated
* (undisposed) at this time, which is ≤ the number of tensors
* (e.g. `a.reshape(newShape)` makes a new Tensor that shares the same
* data buffer with `a`).
* - `unreliable`: True if the memory usage is unreliable. See `reasons` when
* `unreliable` is true.
* - `reasons`: `string[]`, reasons why the memory is unreliable, present if
* `unreliable` is true.
*
* WebGL Properties:
* - `numBytesInGPU`: Number of bytes allocated (undisposed) in the GPU only at
* this time.
*
* @doc {heading: 'Performance', subheading: 'Memory'}
*/
function memory() {
return ENGINE.memory();
}
/**
* Executes the provided function `fn` and after it is executed, cleans up all
* intermediate tensors allocated by `fn` except those returned by `fn`.
* `fn` must not return a Promise (async functions not allowed). The returned
* result can be a complex object.
*
* Using this method helps avoid memory leaks. In general, wrap calls to
* operations in `tf.tidy` for automatic memory cleanup.
*
* NOTE: Variables do *not* get cleaned up when inside a tidy(). If you want to
* dispose variables, please use `tf.disposeVariables` or call dispose()
* directly on variables.
*
* ```js
* // y = 2 ^ 2 + 1
* const y = tf.tidy(() => {
* // a, b, and one will be cleaned up when the tidy ends.
* const one = tf.scalar(1);
* const a = tf.scalar(2);
* const b = a.square();
*
* console.log('numTensors (in tidy): ' + tf.memory().numTensors);
*
* // The value returned inside the tidy function will return
* // through the tidy, in this case to the variable y.
* return b.add(one);
* });
*
* console.log('numTensors (outside tidy): ' + tf.memory().numTensors);
* y.print();
* ```
*
* @param nameOrFn The name of the closure, or the function to execute.
* If a name is provided, the 2nd argument should be the function.
* If debug mode is on, the timing and the memory usage of the function
* will be tracked and displayed on the console using the provided name.
* @param fn The function to execute.
*
* @doc {heading: 'Performance', subheading: 'Memory'}
*/
function tidy(nameOrFn, fn) {
return ENGINE.tidy(nameOrFn, fn);
}
/**
* Disposes any `tf.Tensor`s found within the provided object.
*
* @param container an object that may be a `tf.Tensor` or may directly
* contain `tf.Tensor`s, such as a `Tensor[]` or `{key: Tensor, ...}`. If
* the object is not a `tf.Tensor` or does not contain `Tensors`, nothing
* happens. In general it is safe to pass any object here, except that
* `Promise`s are not supported.
*
* @doc {heading: 'Performance', subheading: 'Memory'}
*/
function dispose(container) {
const tensors = getTensorsInContainer(container);
tensors.forEach(tensor => tensor.dispose());
}
/**
* Keeps a `tf.Tensor` generated inside a `tf.tidy` from being disposed
* automatically.
*
* ```js
* let b;
* const y = tf.tidy(() => {
* const one = tf.scalar(1);
* const a = tf.scalar(2);
*
* // b will not be cleaned up by the tidy. a and one will be cleaned up
* // when the tidy ends.
* b = tf.keep(a.square());
*
* console.log('numTensors (in tidy): ' + tf.memory().numTensors);
*
* // The value returned inside the tidy function will return
* // through the tidy, in this case to the variable y.
* return b.add(one);
* });
*
* console.log('numTensors (outside tidy): ' + tf.memory().numTensors);
* console.log('y:');
* y.print();
* console.log('b:');
* b.print();
* ```
*
* @param result The tensor to keep from being disposed.
*
* @doc {heading: 'Performance', subheading: 'Memory'}
*/
function keep(result) {
return ENGINE.keep(result);
}
/**
* Registers a global backend. The registration should happen when importing
* a module file (e.g. when importing `backend_webgl.ts`), and is used for
* modular builds (e.g. custom tfjs bundle with only webgl support).
*
* @param factory The backend factory function. When called, it should
* return a backend instance, or a promise of an instance.
* @param priority The priority of the backend (higher = more important).
* In case multiple backends are registered, the priority is used to find
* the best backend. Defaults to 1.
* @return False if there is already a registered backend under this name, true
* if not.
*
* @doc {heading: 'Backends'}
*/
function registerBackend(name, factory, priority = 1) {
return ENGINE.registerBackend(name, factory, priority);
}
/**
* Gets the current backend. If no backends have been initialized, this will
* attempt to initialize the best backend. Will throw an error if the highest
* priority backend has async initialization, in which case you should call
* 'await tf.ready()' before running other code.
*
* @doc {heading: 'Backends'}
*/
function backend() {
return ENGINE.backend;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Adds two `tf.Tensor`s element-wise, A + B. Supports broadcasting.
*
*
* ```js
* const a = tf.tensor1d([1, 2, 3, 4]);
* const b = tf.tensor1d([10, 20, 30, 40]);
*
* a.add(b).print(); // or tf.add(a, b)
* ```
*
* ```js
* // Broadcast add a with b.
* const a = tf.scalar(5);
* const b = tf.tensor1d([10, 20, 30, 40]);
*
* a.add(b).print(); // or tf.add(a, b)
* ```
* @param a The first `tf.Tensor` to add.
* @param b The second `tf.Tensor` to add. Must have the same type as `a`.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function add_(a, b) {
let $a = convertToTensor(a, 'a', 'add');
let $b = convertToTensor(b, 'b', 'add');
[$a, $b] = makeTypesMatch($a, $b);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Add, inputs);
}
const add$1 = /* @__PURE__ */ op({ add_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Divides two `tf.Tensor`s element-wise, A / B. Supports broadcasting.
* The result is rounded with floor function.
*
*
* ```js
* const a = tf.tensor1d([1, 4, 9, 16]);
* const b = tf.tensor1d([1, 2, 3, 4]);
*
* a.floorDiv(b).print(); // or tf.div(a, b)
* ```
*
* ```js
* // Broadcast div a with b.
* const a = tf.tensor1d([2, 4, 6, 8]);
* const b = tf.scalar(2);
*
* a.floorDiv(b).print(); // or tf.floorDiv(a, b)
* ```
*
* @param a The first tensor as the numerator.
* @param b The second tensor as the denominator. Must have the same dtype as
* `a`.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function floorDiv_(a, b) {
let $a = convertToTensor(a, 'a', 'floorDiv');
let $b = convertToTensor(b, 'b', 'floorDiv');
[$a, $b] = makeTypesMatch($a, $b);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(FloorDiv, inputs);
}
const floorDiv$2 = /* @__PURE__ */ op({ floorDiv_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Divides two `tf.Tensor`s element-wise, A / B. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([1, 4, 9, 16]);
* const b = tf.tensor1d([1, 2, 3, 4]);
*
* a.div(b).print(); // or tf.div(a, b)
* ```
*
* ```js
* // Broadcast div a with b.
* const a = tf.tensor1d([2, 4, 6, 8]);
* const b = tf.scalar(2);
*
* a.div(b).print(); // or tf.div(a, b)
* ```
*
* @param a The first tensor as the numerator.
* @param b The second tensor as the denominator. Must have the same dtype as
* `a`.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function div_(a, b) {
let $a = convertToTensor(a, 'a', 'div');
let $b = convertToTensor(b, 'b', 'div');
[$a, $b] = makeTypesMatch($a, $b);
if ($a.dtype === 'int32' && $b.dtype === 'int32') {
return floorDiv$2($a, $b);
}
const inputs = { a: $a, b: $b };
const attrs = {};
// tslint:disable-next-line: no-unnecessary-type-assertion
return ENGINE.runKernel(RealDiv, inputs, attrs);
}
const div$1 = /* @__PURE__ */ op({ div_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Multiplies two `tf.Tensor`s element-wise, A * B. Supports broadcasting.
*
* We also expose `tf.mulStrict` which has the same signature as this op and
* asserts that `a` and `b` are the same shape (does not broadcast).
*
* ```js
* const a = tf.tensor1d([1, 2, 3, 4]);
* const b = tf.tensor1d([2, 3, 4, 5]);
*
* a.mul(b).print(); // or tf.mul(a, b)
* ```
*
* ```js
* // Broadcast mul a with b.
* const a = tf.tensor1d([1, 2, 3, 4]);
* const b = tf.scalar(5);
*
* a.mul(b).print(); // or tf.mul(a, b)
* ```
* @param a The first tensor to multiply.
* @param b The second tensor to multiply. Must have the same dtype as `a`.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function mul_(a, b) {
let $a = convertToTensor(a, 'a', 'mul');
let $b = convertToTensor(b, 'b', 'mul');
[$a, $b] = makeTypesMatch($a, $b);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Multiply, inputs);
}
const mul = /* @__PURE__ */ op({ mul_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes absolute value element-wise: `abs(x)`
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.abs().print(); // or tf.abs(x)
* ```
* @param x The input `tf.Tensor`.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function abs_(x) {
const $x = convertToTensor(x, 'x', 'abs');
if ($x.dtype === 'complex64') {
const inputs = { x: $x };
return ENGINE.runKernel(ComplexAbs, inputs);
}
else {
const inputs = { x: $x };
return ENGINE.runKernel(Abs, inputs);
}
}
const abs$2 = /* @__PURE__ */ op({ abs_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the logical or of elements across dimensions of a `tf.Tensor`.
*
* Reduces the input along the dimensions given in `axes`. Unless `keepDims`
* is true, the rank of the `tf.Tensor` is reduced by 1 for each entry in
* `axes`. If `keepDims` is true, the reduced dimensions are retained with
* length 1. If `axes` has no entries, all dimensions are reduced, and a
* `tf.Tensor` with a single element is returned.
*
* ```js
* const x = tf.tensor1d([1, 1, 1], 'bool');
*
* x.any().print(); // or tf.any(x)
* ```
*
* ```js
* const x = tf.tensor2d([1, 1, 0, 0], [2, 2], 'bool');
*
* const axis = 1;
* x.any(axis).print(); // or tf.any(x, axis)
* ```
*
* @param x The input tensor. Must be of dtype bool.
* @param axis The dimension(s) to reduce. By default it reduces
* all dimensions.
* @param keepDims If true, retains reduced dimensions with size 1.
*
* @doc {heading: 'Operations', subheading: 'Reduction'}
*/
function any_(x, axis = null, keepDims = false) {
const $x = convertToTensor(x, 'x', 'any', 'bool');
const inputs = { x: $x };
const attrs = { axis, keepDims };
return ENGINE.runKernel(Any, inputs, attrs);
}
// tslint:disable-next-line:variable-name
const any$2 = /* @__PURE__ */ op({ any_ });
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the indices of the maximum values along an `axis`.
*
* The result has the same shape as `input` with the dimension along `axis`
* removed.
*
* ```js
* const x = tf.tensor1d([1, 2, 3]);
*
* x.argMax().print(); // or tf.argMax(x)
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 4, 3], [2, 2]);
*
* const axis = 1;
* x.argMax(axis).print(); // or tf.argMax(x, axis)
* ```
*
* @param x The input tensor.
* @param axis The dimension to reduce. Defaults to 0 (outer-most dimension).
*
* @doc {heading: 'Operations', subheading: 'Reduction'}
*/
function argMax_(x, axis = 0) {
const $x = convertToTensor(x, 'x', 'argMax');
const inputs = { x: $x };
const attrs = { axis };
return ENGINE.runKernel(ArgMax, inputs, attrs);
}
const argMax$2 = /* @__PURE__ */ op({ argMax_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
*
* @param inputShape Input tensor shape is of the following dimensions:
* `[batch, height, width, inChannels]`.
* @param filterShape The filter shape is of the following dimensions:
* `[filterHeight, filterWidth, depth]`.
* @param strides The strides of the sliding window for each dimension of the
* input tensor: `[strideHeight, strideWidth]`.
* If `strides` is a single number,
* then `strideHeight == strideWidth`.
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1*1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_docs/python/tf/nn/convolution](
* https://www.tensorflow.org/api_docs/python/tf/nn/convolution)
* @param dataFormat The data format of the input and output data.
* Defaults to 'NHWC'.
* @param dilations The dilation rates: `[dilationHeight, dilationWidth]`.
* Defaults to `[1, 1]`. If `dilations` is a single number, then
* `dilationHeight == dilationWidth`.
*/
function computeDilation2DInfo(inputShape, filterShape, strides, pad, dataFormat = 'NHWC', dilations) {
// `computerConv2DInfo` require filterShape to be in the dimension of:
// `[filterHeight, filterWidth, depth, outDepth]`, dilation2d doesn't have
// outDepth, it should have the same depth as the input.
// Input shape: [batch, height, width, inChannels]
const inputChannels = inputShape[3];
const $filterShape = [...filterShape, inputChannels];
const $dataFormat = convertConv2DDataFormat(dataFormat);
return computeConv2DInfo(inputShape, $filterShape, strides, dilations, pad, null /* roundingMode */, null /* depthWise */, $dataFormat);
}
function computePool2DInfo(inShape, filterSize, strides, dilations, pad, roundingMode, dataFormat = 'channelsLast') {
const [filterHeight, filterWidth] = parseTupleParam(filterSize);
let filterShape;
if (dataFormat === 'channelsLast') {
filterShape = [filterHeight, filterWidth, inShape[3], inShape[3]];
}
else if (dataFormat === 'channelsFirst') {
filterShape = [filterHeight, filterWidth, inShape[1], inShape[1]];
}
else {
throw new Error(`Unknown dataFormat ${dataFormat}`);
}
return computeConv2DInfo(inShape, filterShape, strides, dilations, pad, roundingMode, false, dataFormat);
}
/**
* Computes the information for a forward pass of a pooling3D operation.
*/
function computePool3DInfo(inShape, filterSize, strides, dilations, pad, roundingMode, dataFormat = 'NDHWC') {
const [filterDepth, filterHeight, filterWidth] = parse3TupleParam(filterSize);
let filterShape;
let $dataFormat;
if (dataFormat === 'NDHWC') {
$dataFormat = 'channelsLast';
filterShape =
[filterDepth, filterHeight, filterWidth, inShape[4], inShape[4]];
}
else if (dataFormat === 'NCDHW') {
$dataFormat = 'channelsFirst';
filterShape =
[filterDepth, filterHeight, filterWidth, inShape[1], inShape[1]];
}
else {
throw new Error(`Unknown dataFormat ${dataFormat}`);
}
return computeConv3DInfo(inShape, filterShape, strides, dilations, pad, false, $dataFormat, roundingMode);
}
/**
* Computes the information for a forward pass of a convolution/pooling
* operation.
*/
function computeConv2DInfo(inShape, filterShape, strides, dilations, pad, roundingMode, depthwise = false, dataFormat = 'channelsLast') {
let [batchSize, inHeight, inWidth, inChannels] = [-1, -1, -1, -1];
if (dataFormat === 'channelsLast') {
[batchSize, inHeight, inWidth, inChannels] = inShape;
}
else if (dataFormat === 'channelsFirst') {
[batchSize, inChannels, inHeight, inWidth] = inShape;
}
else {
throw new Error(`Unknown dataFormat ${dataFormat}`);
}
const [filterHeight, filterWidth, , filterChannels] = filterShape;
const [strideHeight, strideWidth] = parseTupleParam(strides);
const [dilationHeight, dilationWidth] = parseTupleParam(dilations);
const effectiveFilterHeight = getEffectiveFilterSize(filterHeight, dilationHeight);
const effectiveFilterWidth = getEffectiveFilterSize(filterWidth, dilationWidth);
const { padInfo, outHeight, outWidth } = getPadAndOutInfo(pad, inHeight, inWidth, strideHeight, strideWidth, effectiveFilterHeight, effectiveFilterWidth, roundingMode, dataFormat);
const outChannels = depthwise ? filterChannels * inChannels : filterChannels;
let outShape;
if (dataFormat === 'channelsFirst') {
outShape = [batchSize, outChannels, outHeight, outWidth];
}
else if (dataFormat === 'channelsLast') {
outShape = [batchSize, outHeight, outWidth, outChannels];
}
return {
batchSize,
dataFormat,
inHeight,
inWidth,
inChannels,
outHeight,
outWidth,
outChannels,
padInfo,
strideHeight,
strideWidth,
filterHeight,
filterWidth,
effectiveFilterHeight,
effectiveFilterWidth,
dilationHeight,
dilationWidth,
inShape,
outShape,
filterShape
};
}
/**
* Computes the information for a forward pass of a 3D convolution/pooling
* operation.
*/
function computeConv3DInfo(inShape, filterShape, strides, dilations, pad, depthwise = false, dataFormat = 'channelsLast', roundingMode) {
let [batchSize, inDepth, inHeight, inWidth, inChannels] = [-1, -1, -1, -1, -1];
if (dataFormat === 'channelsLast') {
[batchSize, inDepth, inHeight, inWidth, inChannels] = inShape;
}
else if (dataFormat === 'channelsFirst') {
[batchSize, inChannels, inDepth, inHeight, inWidth] = inShape;
}
else {
throw new Error(`Unknown dataFormat ${dataFormat}`);
}
const [filterDepth, filterHeight, filterWidth, , filterChannels] = filterShape;
const [strideDepth, strideHeight, strideWidth] = parse3TupleParam(strides);
const [dilationDepth, dilationHeight, dilationWidth] = parse3TupleParam(dilations);
const effectiveFilterDepth = getEffectiveFilterSize(filterDepth, dilationDepth);
const effectiveFilterHeight = getEffectiveFilterSize(filterHeight, dilationHeight);
const effectiveFilterWidth = getEffectiveFilterSize(filterWidth, dilationWidth);
const { padInfo, outDepth, outHeight, outWidth } = get3DPadAndOutInfo(pad, inDepth, inHeight, inWidth, strideDepth, strideHeight, strideWidth, effectiveFilterDepth, effectiveFilterHeight, effectiveFilterWidth, roundingMode);
const outChannels = depthwise ? filterChannels * inChannels : filterChannels;
let outShape;
if (dataFormat === 'channelsFirst') {
outShape = [batchSize, outChannels, outDepth, outHeight, outWidth];
}
else if (dataFormat === 'channelsLast') {
outShape = [batchSize, outDepth, outHeight, outWidth, outChannels];
}
return {
batchSize,
dataFormat,
inDepth,
inHeight,
inWidth,
inChannels,
outDepth,
outHeight,
outWidth,
outChannels,
padInfo,
strideDepth,
strideHeight,
strideWidth,
filterDepth,
filterHeight,
filterWidth,
effectiveFilterDepth,
effectiveFilterHeight,
effectiveFilterWidth,
dilationDepth,
dilationHeight,
dilationWidth,
inShape,
outShape,
filterShape
};
}
function computeOutputShape2D(inShape, fieldSize, stride, zeroPad, roundingMode) {
if (zeroPad == null) {
zeroPad = computeDefaultPad(inShape, fieldSize, stride);
}
const inputRows = inShape[0];
const inputCols = inShape[1];
const outputRows = round$2((inputRows - fieldSize + 2 * zeroPad) / stride + 1, roundingMode);
const outputCols = round$2((inputCols - fieldSize + 2 * zeroPad) / stride + 1, roundingMode);
return [outputRows, outputCols];
}
function computeOutputShape4D(inShape, filterShape, outChannels, strides, zeroPad, roundingMode) {
if (zeroPad == null) {
zeroPad = computeDefaultPad(inShape, filterShape[0], strides[0]);
}
const outShape = [0, 0, 0, outChannels];
for (let index = 0; index < 3; index++) {
if (inShape[index] + 2 * zeroPad >= filterShape[index]) {
outShape[index] = round$2((inShape[index] - filterShape[index] + 2 * zeroPad) / strides[index] +
1, roundingMode);
}
}
return outShape;
}
function computeDefaultPad(inputShape, fieldSize, stride, dilation = 1) {
const effectiveFieldSize = getEffectiveFilterSize(fieldSize, dilation);
return Math.floor((inputShape[0] * (stride - 1) - stride + effectiveFieldSize) / 2);
}
function parseTupleParam(param) {
if (typeof param === 'number') {
return [param, param, param];
}
if (param.length === 2) {
return [param[0], param[1], 1];
}
return param;
}
function parse3TupleParam(param) {
return typeof param === 'number' ? [param, param, param] : param;
}
/* See https://www.tensorflow.org/api_docs/python/tf/nn/atrous_conv2d
* Atrous convolution is equivalent to standard convolution with upsampled
* filters with effective_filter_height =
* filter_height + (filter_height - 1) * (dilation - 1)
* and effective_filter_width =
* filter_width + (filter_width - 1) * (dilation - 1),
* produced by inserting dilation - 1 zeros along consecutive elements across
* the filters' spatial dimensions.
* When there is a dilation, this converts a filter dimension to the
* effective filter dimension, so it can be used in a standard convolution.
*/
function getEffectiveFilterSize(filterSize, dilation) {
if (dilation <= 1) {
return filterSize;
}
return filterSize + (filterSize - 1) * (dilation - 1);
}
function getPadAndOutInfo(pad, inHeight, inWidth, strideHeight, strideWidth, filterHeight, filterWidth, roundingMode, dataFormat) {
let padInfo;
let outHeight;
let outWidth;
if (typeof pad === 'number') {
const padType = (pad === 0) ? 'VALID' : 'NUMBER';
padInfo = { top: pad, bottom: pad, left: pad, right: pad, type: padType };
const outShape = computeOutputShape2D([inHeight, inWidth], filterHeight, strideHeight, pad, roundingMode);
outHeight = outShape[0];
outWidth = outShape[1];
}
else if (pad === 'same') {
outHeight = Math.ceil(inHeight / strideHeight);
outWidth = Math.ceil(inWidth / strideWidth);
const padAlongHeight = Math.max(0, (outHeight - 1) * strideHeight + filterHeight - inHeight);
const padAlongWidth = Math.max(0, (outWidth - 1) * strideWidth + filterWidth - inWidth);
const top = Math.floor(padAlongHeight / 2);
const bottom = padAlongHeight - top;
const left = Math.floor(padAlongWidth / 2);
const right = padAlongWidth - left;
padInfo = { top, bottom, left, right, type: 'SAME' };
}
else if (pad === 'valid') {
padInfo = { top: 0, bottom: 0, left: 0, right: 0, type: 'VALID' };
outHeight = Math.ceil((inHeight - filterHeight + 1) / strideHeight);
outWidth = Math.ceil((inWidth - filterWidth + 1) / strideWidth);
}
else if (typeof pad === 'object') {
const top = dataFormat === 'channelsLast' ? pad[1][0] : pad[2][0];
const bottom = dataFormat === 'channelsLast' ? pad[1][1] : pad[2][1];
const left = dataFormat === 'channelsLast' ? pad[2][0] : pad[3][0];
const right = dataFormat === 'channelsLast' ? pad[2][1] : pad[3][1];
const padType = (top === 0 && bottom === 0 && left === 0 && right === 0) ?
'VALID' :
'EXPLICIT';
padInfo = { top, bottom, left, right, type: padType };
outHeight = round$2((inHeight - filterHeight + top + bottom) / strideHeight + 1, roundingMode);
outWidth = round$2((inWidth - filterWidth + left + right) / strideWidth + 1, roundingMode);
}
else {
throw Error(`Unknown padding parameter: ${pad}`);
}
return { padInfo, outHeight, outWidth };
}
function get3DPadAndOutInfo(pad, inDepth, inHeight, inWidth, strideDepth, strideHeight, strideWidth, filterDepth, filterHeight, filterWidth, roundingMode) {
let padInfo;
let outDepth;
let outHeight;
let outWidth;
if (pad === 'valid') {
pad = 0;
}
if (typeof pad === 'number') {
const padType = (pad === 0) ? 'VALID' : 'NUMBER';
padInfo = {
top: pad,
bottom: pad,
left: pad,
right: pad,
front: pad,
back: pad,
type: padType
};
const outShape = computeOutputShape4D([inDepth, inHeight, inWidth, 1], [filterDepth, filterHeight, filterWidth], 1, [strideDepth, strideHeight, strideWidth], pad, roundingMode);
outDepth = outShape[0];
outHeight = outShape[1];
outWidth = outShape[2];
}
else if (pad === 'same') {
outDepth = Math.ceil(inDepth / strideDepth);
outHeight = Math.ceil(inHeight / strideHeight);
outWidth = Math.ceil(inWidth / strideWidth);
const padAlongDepth = (outDepth - 1) * strideDepth + filterDepth - inDepth;
const padAlongHeight = (outHeight - 1) * strideHeight + filterHeight - inHeight;
const padAlongWidth = (outWidth - 1) * strideWidth + filterWidth - inWidth;
const front = Math.floor(padAlongDepth / 2);
const back = padAlongDepth - front;
const top = Math.floor(padAlongHeight / 2);
const bottom = padAlongHeight - top;
const left = Math.floor(padAlongWidth / 2);
const right = padAlongWidth - left;
padInfo = { top, bottom, left, right, front, back, type: 'SAME' };
}
else {
throw Error(`Unknown padding parameter: ${pad}`);
}
return { padInfo, outDepth, outHeight, outWidth };
}
/**
* Rounds a value depending on the rounding mode
* @param value
* @param roundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*/
function round$2(value, roundingMode) {
if (!roundingMode) {
return Math.trunc(value);
}
switch (roundingMode) {
case 'round':
// used for Caffe Conv
return Math.round(value);
case 'ceil':
// used for Caffe Pool
return Math.ceil(value);
case 'floor':
return Math.floor(value);
default:
throw new Error(`Unknown roundingMode ${roundingMode}`);
}
}
function tupleValuesAreOne(param) {
const [dimA, dimB, dimC] = parseTupleParam(param);
return dimA === 1 && dimB === 1 && dimC === 1;
}
function eitherStridesOrDilationsAreOne(strides, dilations) {
return tupleValuesAreOne(strides) || tupleValuesAreOne(dilations);
}
function stridesOrDilationsArePositive(values) {
return parseTupleParam(values).every(value => value > 0);
}
/**
* Convert Conv2D dataFormat from 'NHWC'|'NCHW' to
* 'channelsLast'|'channelsFirst'
* @param dataFormat in 'NHWC'|'NCHW' mode
* @return dataFormat in 'channelsLast'|'channelsFirst' mode
* @throws unknown dataFormat
*/
function convertConv2DDataFormat(dataFormat) {
if (dataFormat === 'NHWC') {
return 'channelsLast';
}
else if (dataFormat === 'NCHW') {
return 'channelsFirst';
}
else {
throw new Error(`Unknown dataFormat ${dataFormat}`);
}
}
/**
* Check validity of pad when using dimRoundingMode.
* @param opDesc A string of op description
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid` output will be smaller than input if filter is larger
* than 1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_docs/python/tf/nn/convolution](
* https://www.tensorflow.org/api_docs/python/tf/nn/convolution)
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
* @throws unknown padding parameter
*/
function checkPadOnDimRoundingMode(opDesc, pad, dimRoundingMode) {
if (dimRoundingMode != null) {
if (typeof pad === 'string') {
throw Error(`Error in ${opDesc}: pad must be an integer when using ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
}
else if (typeof pad === 'number') {
assert$1(isInt(pad), () => `Error in ${opDesc}: pad must be an integer when using ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${pad}.`);
}
else if (typeof pad === 'object') {
pad.forEach(p => {
p.forEach(v => {
assert$1(isInt(v), () => `Error in ${opDesc}: pad must be an integer when using ` +
`dimRoundingMode ${dimRoundingMode} but got pad ${v}.`);
});
});
}
else {
throw Error(`Error in ${opDesc}: Unknown padding parameter: ${pad}`);
}
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Reshapes a `tf.Tensor` to a given shape.
*
* Given an input tensor, returns a new tensor with the same values as the
* input tensor with shape `shape`.
*
* If one component of shape is the special value -1, the size of that
* dimension is computed so that the total size remains constant. In
* particular, a shape of [-1] flattens into 1-D. At most one component of
* shape can be -1.
*
* If shape is 1-D or higher, then the operation returns a tensor with shape
* shape filled with the values of tensor. In this case, the number of
* elements implied by shape must be the same as the number of elements in
* tensor.
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
* x.reshape([2, 2]).print();
* ```
*
* @param x The input tensor to be reshaped.
* @param shape An array of integers defining the output tensor shape.
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function reshape_(x, shape) {
const $x = convertToTensor(x, 'x', 'reshape', 'string_or_numeric');
const inputs = { x: $x };
const attrs = { shape };
return ENGINE.runKernel(Reshape$1, inputs, attrs);
}
const reshape$2 = /* @__PURE__ */ op({ reshape_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Concatenates a list of `tf.Tensor`s along a given axis.
*
* The tensors ranks and types must match, and their sizes must match in all
* dimensions except `axis`.
*
* Also available are stricter rank-specific methods that assert that
* `tensors` are of the given rank:
* - `tf.concat1d`
* - `tf.concat2d`
* - `tf.concat3d`
* - `tf.concat4d`
*
* Except `tf.concat1d` (which does not have axis param), all methods have
* same signature as this method.
*
* ```js
* const a = tf.tensor1d([1, 2]);
* const b = tf.tensor1d([3, 4]);
* a.concat(b).print(); // or a.concat(b)
* ```
*
* ```js
* const a = tf.tensor1d([1, 2]);
* const b = tf.tensor1d([3, 4]);
* const c = tf.tensor1d([5, 6]);
* tf.concat([a, b, c]).print();
* ```
*
* ```js
* const a = tf.tensor2d([[1, 2], [10, 20]]);
* const b = tf.tensor2d([[3, 4], [30, 40]]);
* const axis = 1;
* tf.concat([a, b], axis).print();
* ```
* @param tensors A list of tensors to concatenate.
* @param axis The axis to concatenate along. Defaults to 0 (the first dim).
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function concat_(tensors, axis = 0) {
assert$1(tensors.length >= 1, () => 'Pass at least one tensor to concat');
const $tensors = convertToTensorArray(tensors, 'tensors', 'concat', 'string_or_numeric');
if ($tensors[0].dtype === 'complex64') {
$tensors.forEach(tensor => {
if (tensor.dtype !== 'complex64') {
throw new Error(`Cannot concatenate complex64 tensors with a tensor
with dtype ${tensor.dtype}. `);
}
});
}
if ($tensors.length === 1) {
return clone($tensors[0]);
}
const inputs = $tensors;
const attr = { axis };
return ENGINE.runKernel(Concat, inputs, attr);
}
const concat$2 = /* @__PURE__ */ op({ concat_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the dot product of two matrices, A * B. These must be matrices.
*
* ```js
* const a = tf.tensor2d([1, 2], [1, 2]);
* const b = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* a.matMul(b).print(); // or tf.matMul(a, b)
* ```
* @param a First matrix in dot product operation.
* @param b Second matrix in dot product operation.
* @param transposeA If true, `a` is transposed before multiplication.
* @param transposeB If true, `b` is transposed before multiplication.
*
* @doc {heading: 'Operations', subheading: 'Matrices'}
*/
function matMul_(a, b, transposeA = false, transposeB = false) {
let $a = convertToTensor(a, 'a', 'matMul');
let $b = convertToTensor(b, 'b', 'matMul');
[$a, $b] = makeTypesMatch($a, $b);
const inputs = { a: $a, b: $b };
const attrs = { transposeA, transposeB };
return ENGINE.runKernel(BatchMatMul, inputs, attrs);
}
const matMul$1 = /* @__PURE__ */ op({ matMul_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes sigmoid element-wise, `1 / (1 + exp(-x))`
*
* ```js
* const x = tf.tensor1d([0, -1, 2, -3]);
*
* x.sigmoid().print(); // or tf.sigmoid(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function sigmoid_(x) {
const $x = convertToTensor(x, 'x', 'sigmoid', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Sigmoid$1, inputs);
}
const sigmoid$2 = /* @__PURE__ */ op({ sigmoid_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Extracts a slice from a `tf.Tensor` starting at coordinates `begin`
* and is of size `size`.
*
* Also available are stricter rank-specific methods with the same signature
* as this method that assert that `x` is of the given rank:
* - `tf.slice1d`
* - `tf.slice2d`
* - `tf.slice3d`
* - `tf.slice4d`
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
*
* x.slice([1], [2]).print();
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* x.slice([1, 0], [1, 2]).print();
* ```
* @param x The input `tf.Tensor` to slice from.
* @param begin The coordinates to start the slice from. The length can be
* less than the rank of x - the rest of the axes will have implicit 0 as
* start. Can also be a single number, in which case it specifies the
* first axis.
* @param size The size of the slice. The length can be less than the rank of
* x - the rest of the axes will have implicit -1. A value of -1 requests
* the rest of the dimensions in the axis. Can also be a single number,
* in which case it specifies the size of the first axis.
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function slice_(x, begin, size) {
const $x = convertToTensor(x, 'x', 'slice', 'string_or_numeric');
if ($x.rank === 0) {
throw new Error('Slicing scalar is not possible');
}
const inputs = { x: $x };
const attrs = { begin, size };
return ENGINE.runKernel(Slice, inputs, attrs);
}
const slice$2 = /* @__PURE__ */ op({ slice_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes hyperbolic tangent of the input `tf.Tensor` element-wise: `tanh(x)`
*
* ```js
* const x = tf.tensor1d([0, 1, -1, 70]);
*
* x.tanh().print(); // or tf.tanh(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function tanh_(x) {
const $x = convertToTensor(x, 'x', 'tanh', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Tanh$1, inputs);
}
const tanh$2 = /* @__PURE__ */ op({ tanh_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* This operation reshapes the "batch" dimension 0 into `M + 1` dimensions of
* shape `blockShape + [batch]`, interleaves these blocks back into the grid
* defined by the spatial dimensions `[1, ..., M]`, to obtain a result with
* the same rank as the input. The spatial dimensions of this intermediate
* result are then optionally cropped according to `crops` to produce the
* output. This is the reverse of `tf.spaceToBatchND`. See below for a precise
* description.
*
* ```js
* const x = tf.tensor4d([1, 2, 3, 4], [4, 1, 1, 1]);
* const blockShape = [2, 2];
* const crops = [[0, 0], [0, 0]];
*
* x.batchToSpaceND(blockShape, crops).print();
* ```
*
* @param x A `tf.Tensor`. N-D with `x.shape` = `[batch] + spatialShape +
* remainingShape`, where spatialShape has `M` dimensions.
* @param blockShape A 1-D array. Must have shape `[M]`, all values must
* be >= 1.
* @param crops A 2-D array. Must have shape `[M, 2]`, all values must be >= 0.
* `crops[i] = [cropStart, cropEnd]` specifies the amount to crop from input
* dimension `i + 1`, which corresponds to spatial dimension `i`. It is required
* that `cropStart[i] + cropEnd[i] <= blockShape[i] * inputShape[i + 1]`
*
* This operation is equivalent to the following steps:
*
* 1. Reshape `x` to `reshaped` of shape: `[blockShape[0], ...,
* blockShape[M-1], batch / prod(blockShape), x.shape[1], ...,
* x.shape[N-1]]`
*
* 2. Permute dimensions of `reshaped` to produce `permuted` of shape `[batch /
* prod(blockShape),x.shape[1], blockShape[0], ..., x.shape[M],
* blockShape[M-1],x.shape[M+1], ..., x.shape[N-1]]`
*
* 3. Reshape `permuted` to produce `reshapedPermuted` of shape `[batch /
* prod(blockShape),x.shape[1] * blockShape[0], ..., x.shape[M] *
* blockShape[M-1],x.shape[M+1], ..., x.shape[N-1]]`
*
* 4. Crop the start and end of dimensions `[1, ..., M]` of `reshapedPermuted`
* according to `crops` to produce the output of shape: `[batch /
* prod(blockShape),x.shape[1] * blockShape[0] - crops[0,0] - crops[0,1],
* ..., x.shape[M] * blockShape[M-1] - crops[M-1,0] -
* crops[M-1,1],x.shape[M+1], ..., x.shape[N-1]]`
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function batchToSpaceND_(x, blockShape, crops) {
const $x = convertToTensor(x, 'x', 'batchToSpaceND');
const prod = blockShape.reduce((a, b) => a * b);
assert$1($x.rank >= 1 + blockShape.length, () => `input rank is ${$x.rank} but should be > than blockShape.length ${blockShape.length}`);
assert$1(crops.length === blockShape.length, () => `crops.length is ${crops.length} but should be equal to blockShape.length ${blockShape.length}`);
assert$1($x.shape[0] % prod === 0, () => `input tensor batch is ${$x.shape[0]} but is not divisible by the product of ` +
`the elements of blockShape ${blockShape.join(' * ')} === ${prod}`);
const inputs = { x: $x };
const attrs = { blockShape, crops };
return ENGINE.runKernel(BatchToSpaceND, inputs, attrs);
}
const batchToSpaceND$2 = /* @__PURE__ */ op({ batchToSpaceND_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Broadcast an array to a compatible shape NumPy-style.
*
* The tensor's shape is compared to the broadcast shape from end to beginning.
* Ones are prepended to the tensor's shape until it has the same length as
* the broadcast shape. If input.shape[i]==shape[i], the (i+1)-th axis is
* already broadcast-compatible. If input.shape[i]==1 and shape[i]==N, then
* the input tensor is tiled N times along that axis (using tf.tile).
*
* @param input The tensor that is to be broadcasted.
* @param shape The input is to be broadcast to this shape.
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function broadcastTo_(x, shape) {
let input = convertToTensor(x, 'broadcastTo', 'x');
const xShape = input.shape;
assertNonNegativeIntegerDimensions(shape);
if (shape.length < input.rank) {
throw new Error(`broadcastTo(): shape.length=${shape.length} < input.rank=${input.rank}.`);
}
if (shape.length > input.rank) {
const newShape = input.shape.slice();
while (newShape.length < shape.length) {
newShape.unshift(1);
}
input = reshape$2(input, newShape);
}
const inputShape = input.shape;
const reps = Array.from(shape);
for (let i = shape.length - 1; i >= 0; i--) {
if (inputShape[i] === shape[i]) {
reps[i] = 1;
}
else if (input.shape[i] !== 1) {
throw new Error(`broadcastTo(): [${xShape}] cannot be broadcast to [${shape}].`);
}
}
const axes = reps.map((n, i) => n > 1 ? i : -1).filter(i => i >= 0);
if (axes.length === 0) {
return clone(input);
}
// TODO call broadcastTo kernel directly once backends implement broadcstTo
const inputs = { x: input };
const attrs = { reps };
return ENGINE.runKernel(Tile, inputs, attrs);
}
const broadcastTo = /* @__PURE__ */ op({ broadcastTo_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` filled with a scalar value.
*
* ```js
* tf.fill([2, 2], 4).print();
* ```
*
* @param shape An array of integers defining the output tensor shape.
* @param value The scalar value to fill the tensor with.
* @param dtype The type of an element in the resulting tensor. Defaults to
* 'float32' if the given param value is a number, otherwise 'string'.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function fill$2(shape, value, dtype) {
assertNonNegativeIntegerDimensions(shape);
dtype = dtype || inferDtype(value);
const attrs = { shape, value, dtype };
return ENGINE.runKernel(Fill, {}, attrs);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Clips values element-wise. `max(min(x, clipValueMax), clipValueMin)`
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.clipByValue(-2, 3).print(); // or tf.clipByValue(x, -2, 3)
* ```
* @param x The input tensor.
* @param clipValueMin Lower bound of range to be clipped to.
* @param clipValueMax Upper bound of range to be clipped to.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function clipByValue_(x, clipValueMin, clipValueMax) {
const $x = convertToTensor(x, 'x', 'clipByValue');
assert$1((clipValueMin <= clipValueMax), () => `Error in clip: min (${clipValueMin}) must be ` +
`less than or equal to max (${clipValueMax}).`);
if (clipValueMin === clipValueMax) {
return fill$2($x.shape, clipValueMin, $x.dtype);
}
const inputs = { x: $x };
const attrs = { clipValueMin, clipValueMax };
return ENGINE.runKernel(ClipByValue, inputs, attrs);
}
const clipByValue$2 = /* @__PURE__ */ op({ clipByValue_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Converts two real numbers to a complex number.
*
* Given a tensor `real` representing the real part of a complex number, and a
* tensor `imag` representing the imaginary part of a complex number, this
* operation returns complex numbers elementwise of the form [r0, i0, r1, i1],
* where r represents the real part and i represents the imag part.
*
* The input tensors real and imag must have the same shape.
*
* ```js
* const real = tf.tensor1d([2.25, 3.25]);
* const imag = tf.tensor1d([4.75, 5.75]);
* const complex = tf.complex(real, imag);
*
* complex.print();
* ```
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function complex_(real, imag) {
const $real = convertToTensor(real, 'real', 'complex');
const $imag = convertToTensor(imag, 'imag', 'complex');
assertShapesMatch($real.shape, $imag.shape, `real and imag shapes, ${$real.shape} and ${$imag.shape}, ` +
`must match in call to tf.complex().`);
const inputs = { real: $real, imag: $imag };
return ENGINE.runKernel(Complex, inputs);
}
const complex$2 = /* @__PURE__ */ op({ complex_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes a 2D convolution over the input x.
*
* @param x The input tensor, of rank 4 or rank 3, of shape
* `[batch, height, width, inChannels]`. If rank 3, batch of 1 is
* assumed.
* @param filter The filter, rank 4, of shape
* `[filterHeight, filterWidth, inDepth, outDepth]`.
* @param strides The strides of the convolution: `[strideHeight,
* strideWidth]`.
* @param pad The type of padding algorithm.
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* - For more info, see this guide:
* [https://www.tensorflow.org/api_docs/python/tf/nn/convolution](
* https://www.tensorflow.org/api_docs/python/tf/nn/convolution)
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
* "NHWC". Specify the data format of the input and output data. With the
* default format "NHWC", the data is stored in the order of: [batch,
* height, width, channels].
* @param dilations The dilation rates: `[dilationHeight, dilationWidth]`
* in which we sample input values across the height and width dimensions
* in atrous convolution. Defaults to `[1, 1]`. If `dilations` is a single
* number, then `dilationHeight == dilationWidth`. If it is greater than
* 1, then all values of `strides` must be 1.
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*
* @doc {heading: 'Operations', subheading: 'Convolution'}
*/
function conv2d_(x, filter, strides, pad, dataFormat = 'NHWC', dilations = [1, 1], dimRoundingMode) {
const $x = convertToTensor(x, 'x', 'conv2d', 'float32');
const $filter = convertToTensor(filter, 'filter', 'conv2d', 'float32');
let x4D = $x;
let reshapedTo4D = false;
if ($x.rank === 3) {
reshapedTo4D = true;
x4D = reshape$2($x, [1, $x.shape[0], $x.shape[1], $x.shape[2]]);
}
assert$1(x4D.rank === 4, () => `Error in conv2d: input must be rank 4, but got rank ${x4D.rank}.`);
assert$1($filter.rank === 4, () => `Error in conv2d: filter must be rank 4, but got rank ` +
`${$filter.rank}.`);
checkPadOnDimRoundingMode('conv2d', pad, dimRoundingMode);
const inDepth = dataFormat === 'NHWC' ? x4D.shape[3] : x4D.shape[1];
assert$1(inDepth === $filter.shape[2], () => `Error in conv2d: depth of input (${inDepth}) must match ` +
`input depth for filter ${$filter.shape[2]}.`);
assert$1(eitherStridesOrDilationsAreOne(strides, dilations), () => 'Error in conv2D: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
assert$1(stridesOrDilationsArePositive(dilations), () => 'Error in conv2D: Dilated rates should be larger than 0.');
assert$1(stridesOrDilationsArePositive(strides), () => 'Error in conv2D: Strides should be larger than 0.');
const inputs = { x: x4D, filter: $filter };
const attrs = { strides, pad, dataFormat, dilations, dimRoundingMode };
// tslint:disable-next-line: no-unnecessary-type-assertion
const res = ENGINE.runKernel(Conv2D, inputs, attrs);
if (reshapedTo4D) {
return reshape$2(res, [res.shape[1], res.shape[2], res.shape[3]]);
}
return res;
}
const conv2d$1 = /* @__PURE__ */ op({ conv2d_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the derivative of the input of a 2D convolution.
*
* @param xShape The shape of the input: [batch, height, width, inDepth].
* If length of 3, batch of 1 is assumed.
* @param dy The derivative of the output, of rank 4 or rank 3 of shape
* `[batch, outHeight, outWidth, outDepth]`. If rank 3, batch of 1 is
* assumed.
* @param filter The filter, rank 4, of shape
* `[filterHeight, filterWidth, inDepth, outDepth]`.
* @param strides The strides of the convolution: `[strideHeight,
* strideWidth]`.
* @param pad The type of padding algorithm used:
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
* "NHWC". Specify the data format of the input and output data. With the
* default format "NHWC", the data is stored in the order of: [batch,
* height, width, channels].
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*/
function conv2DBackpropInput_(xShape, dy, filter, strides, pad, dataFormat = 'NHWC', dimRoundingMode) {
assert$1(xShape.length === dy.rank, () => `Length of inShape ` +
`(${xShape.length}) and rank of dy (${dy.rank}) must match`);
let xShape4D = xShape;
let dy4D = dy;
let reshapedTo4D = false;
if (dy.rank === 3) {
reshapedTo4D = true;
dy4D = reshape$2(dy, [1, dy.shape[0], dy.shape[1], dy.shape[2]]);
xShape4D = [1, xShape[0], xShape[1], xShape[2]];
}
assert$1(xShape4D.length === 4, () => `Error in conv2dDerInput: inShape must be length 4, but got length ` +
`${xShape4D.length}.`);
assert$1(dy4D.rank === 4, () => `Error in conv2dDerInput: dy must be rank 4, but got ` +
`rank ${dy4D.rank}`);
assert$1(filter.rank === 4, () => `Error in conv2dDerInput: filter must be rank 4, but got ` +
`rank ${filter.rank}`);
const inDepth = dataFormat === 'NHWC' ? xShape4D[3] : xShape4D[1];
const outDepth = dataFormat === 'NHWC' ? dy4D.shape[3] : dy4D.shape[1];
assert$1(inDepth === filter.shape[2], () => `Error in conv2dDerInput: depth of input (${inDepth}) must ` +
`match input depth for filter ${filter.shape[2]}.`);
assert$1(outDepth === filter.shape[3], () => `Error in conv2dDerInput: depth of output (${outDepth}) must ` +
`match output depth for filter ${filter.shape[3]}.`);
checkPadOnDimRoundingMode('conv2dDerInput', pad, dimRoundingMode);
const inputs = { dy: dy4D, filter };
const attrs = { strides, pad, dataFormat, dimRoundingMode, inputShape: xShape4D };
// tslint:disable-next-line: no-unnecessary-type-assertion
const res = ENGINE.runKernel(Conv2DBackpropInput, inputs, attrs);
if (reshapedTo4D) {
return reshape$2(res, [res.shape[1], res.shape[2], res.shape[3]]);
}
return res;
}
const conv2DBackpropInput$2 = /* @__PURE__ */ op({ conv2DBackpropInput_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the derivative of the input of a 3D convolution.
*
* @param xShape The shape of the input: [batch, depth, height, width,
* in_channels]. If length of 4, batch of 1 is assumed.
* @param dy The derivative of the output, of rank 5 or rank 4 of shape
* `[batch, outDepth, outHeight, outWidth, in_channels]`.
* If rank 4, batch of 1 is assumed.
* @param filter The filter, rank 5, of shape
* `[filterDepth, filterHeight, filterWidth, inDepth, outDepth]`.
* @param strides The strides of the convolution: `[strideDepth, strideHeight,
* strideWidth]`.
* @param pad The type of padding algorithm used:
* - `same` and stride 1: output will be of same size as input,
* regardless of filter size.
* - `valid`: output will be smaller than input if filter is larger
* than 1x1.
*/
function conv3DBackpropInput_(xShape, dy, filter, strides, pad) {
assert$1(xShape.length === dy.rank, () => `Length of inShape ` +
`(${xShape.length}) and rank of dy (${dy.rank}) must match`);
let xShape5D = xShape;
let dy5D = dy;
let reshapedTo5D = false;
if (dy.rank === 4) {
reshapedTo5D = true;
dy5D = reshape$2(dy, [1, dy.shape[0], dy.shape[1], dy.shape[2], dy.shape[3]]);
xShape5D = [1, xShape[0], xShape[1], xShape[2], xShape[3]];
}
const inDepth = xShape5D[4];
const outDepth = dy5D.shape[4];
assert$1(xShape5D.length === 5, () => `Error in conv3dDerInput: inShape must be length 5, but got length ` +
`${xShape5D.length}.`);
assert$1(dy5D.rank === 5, () => `Error in conv3dDerInput: dy must be rank 5, but got ` +
`rank ${dy5D.rank}`);
assert$1(filter.rank === 5, () => `Error in conv3dDerInput: filter must be rank 5, but got ` +
`rank ${filter.rank}`);
assert$1(inDepth === filter.shape[3], () => `Error in conv3dDerInput: depth of input (${inDepth}) must ` +
`match input depth for filter ${filter.shape[3]}.`);
assert$1(outDepth === filter.shape[4], () => `Error in conv3dDerInput: depth of output (${outDepth}) must ` +
`match output depth for filter ${filter.shape[4]}.`);
const inputs = { dy: dy5D, filter };
const attrs = { pad, strides, inputShape: xShape5D };
// tslint:disable-next-line: no-unnecessary-type-assertion
const res = ENGINE.runKernel(Conv3DBackpropInputV2, inputs, attrs);
if (reshapedTo5D) {
return reshape$2(res, [res.shape[1], res.shape[2], res.shape[3], res.shape[4]]);
}
return res;
}
const conv3DBackpropInput$1 = /* @__PURE__ */ op({ conv3DBackpropInput_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes cos of the input `tf.Tensor` element-wise: `cos(x)`
*
* ```js
* const x = tf.tensor1d([0, Math.PI / 2, Math.PI * 3 / 4]);
*
* x.cos().print(); // or tf.cos(x)
* ```
* @param x The input tensor. Must be float32 type.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function cos_(x) {
const $x = convertToTensor(x, 'x', 'cos', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Cos, inputs);
}
const cos$2 = /* @__PURE__ */ op({ cos_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes hyperbolic cos of the input `tf.Tensor` element-wise: `cosh(x)`
*
* ```js
* const x = tf.tensor1d([0, 1, -1, .7]);
*
* x.cosh().print(); // or tf.cosh(x)
* ```
* @param x The input tensor. Must be float32 type.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function cosh_(x) {
const $x = convertToTensor(x, 'x', 'cosh', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Cosh, inputs);
}
const cosh$2 = /* @__PURE__ */ op({ cosh_ });
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the 'License');
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an 'AS IS' BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the cumulative product of a `tf.Tensor` along `axis`.
*
* ```js
* const x = tf.tensor([1, 2, 3, 4]);
* x.cumprod().print();
* ```
* ```js
* const x = tf.tensor([[1, 2], [3, 4]]);
* x.cumprod().print();
* ```
*
* @param x The input tensor to cumulatively multiply.
* @param axis The axis along which to multiply. Optional. Defaults to 0.
* @param exclusive Whether to perform exclusive cumulative product. Optional.
* Defaults to false. If set to true then the product of each tensor entry
* does not include its own value, but only the values previous to it
* along the specified axis.
* @param reverse Whether to multiply in the opposite direction. Optional.
* Defaults to false.
*
* @doc {heading: 'Operations', subheading: 'Scan'}
*/
function cumprod_(x, axis = 0, exclusive = false, reverse = false) {
const $x = convertToTensor(x, 'x', 'cumprod');
const inputs = { x: $x };
const attrs = { axis, exclusive, reverse };
return ENGINE.runKernel(Cumprod, inputs, attrs);
}
const cumprod$2 = /* @__PURE__ */ op({ cumprod_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the cumulative sum of a `tf.Tensor` along `axis`.
*
* ```js
* const x = tf.tensor([1, 2, 3, 4]);
* x.cumsum().print();
* ```
* ```js
* const x = tf.tensor([[1, 2], [3, 4]]);
* x.cumsum().print();
* ```
*
* @param x The input tensor to be summed.
* @param axis The axis along which to sum. Optional. Defaults to 0.
* @param exclusive Whether to perform exclusive cumulative sum. Optional.
* Defaults to false. If set to true then the sum of each tensor entry
* does not include its own value, but only the values previous to it
* along the specified axis.
* @param reverse Whether to sum in the opposite direction. Optional.
* Defaults to false.
*
* @doc {heading: 'Operations', subheading: 'Scan'}
*/
function cumsum_(x, axis = 0, exclusive = false, reverse = false) {
const $x = convertToTensor(x, 'x', 'cumsum');
const inputs = { x: $x };
const attrs = { axis, exclusive, reverse };
return ENGINE.runKernel(Cumsum, inputs, attrs);
}
const cumsum$2 = /* @__PURE__ */ op({ cumsum_ });
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the dimensions in the input shape that are broadcasted to
* produce the provided output shape.
*
* The returned dimensions are 0-indexed and sorted. An example:
* inShape = [4, 1, 3]
* outShape = [5, 4, 3, 3]
* result = [1]. Dimension 1 (2nd dimension of input) gets broadcasted 1 => 3.
*/
function getBroadcastDims$1(inShape, outShape) {
const inRank = inShape.length;
const dims = [];
for (let i = 0; i < inRank; i++) {
const dim = inRank - 1 - i;
const a = inShape[dim] || 1;
const b = outShape[outShape.length - 1 - i] || 1;
if (b > 1 && a === 1) {
dims.unshift(dim);
}
}
return dims;
}
/**
* Returns the axes in the output space that should be reduced to produce
* the input space.
*/
function getReductionAxes(inShape, outShape) {
const result = [];
for (let i = 0; i < outShape.length; i++) {
const inDim = inShape[inShape.length - i - 1];
const outAxis = outShape.length - i - 1;
const outDim = outShape[outAxis];
if (inDim == null || (inDim === 1 && outDim > 1)) {
result.unshift(outAxis);
}
}
return result;
}
function assertAndGetBroadcastShape(shapeA, shapeB) {
const l = Math.max(shapeA.length, shapeB.length);
const result = new Array(l);
for (let i = 0; i < l; i++) {
let a = shapeA[shapeA.length - i - 1];
if (a == null) {
a = 1;
}
let b = shapeB[shapeB.length - i - 1];
if (b == null) {
b = 1;
}
if (a === 1) {
result[l - i - 1] = b;
}
else if (b === 1) {
result[l - i - 1] = a;
}
else if (a !== b) {
const errMsg = `Operands could not be broadcast together with shapes ` +
`${shapeA} and ${shapeB}.`;
throw Error(errMsg);
}
else {
result[l - i - 1] = a;
}
}
return result;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of (a == b) element-wise. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
* const b = tf.tensor1d([2, 2, 2]);
*
* a.equal(b).print();
* ```
*
* @param a The first input tensor.
* @param b The second input tensor. Must have the same dtype as `a`.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function equal_(a, b) {
let $a = convertToTensor(a, 'a', 'equal', 'string_or_numeric');
let $b = convertToTensor(b, 'b', 'equal', 'string_or_numeric');
[$a, $b] = makeTypesMatch($a, $b);
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Equal, inputs);
}
const equal$2 = /* @__PURE__ */ op({ equal_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the elements, either `a` or `b` depending on the `condition`.
*
* If the condition is true, select from `a`, otherwise select from `b`.
*
* ```js
* const cond = tf.tensor1d([false, false, true], 'bool');
* const a = tf.tensor1d([1 , 2, 3]);
* const b = tf.tensor1d([-1, -2, -3]);
*
* a.where(cond, b).print();
* ```
*
* @param condition The input condition. Must be of dtype bool.
* @param a If `condition` is rank 1, `a` may have a higher rank but
* its first dimension must match the size of `condition`.
* @param b A tensor with the same dtype as `a` and with shape that is
* compatible with `a`.
* @return A tensor with same dtype as `a` and `b`, and shape that is
* broadcastable from `a` and `b`.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function where_(condition, a, b) {
const $a = convertToTensor(a, 'a', 'where');
const $b = convertToTensor(b, 'b', 'where');
const $condition = convertToTensor(condition, 'condition', 'where', 'bool');
// TODO: move this logic to forward function when the broadcastTo op is
// implemented in WASM.
// Find the broadcastable shape for $condition, $a, and $b.
const broadcastShape = assertAndGetBroadcastShape(assertAndGetBroadcastShape($condition.shape, $a.shape), $b.shape);
const $broadcastedCondition = broadcastTo($condition, broadcastShape);
const $broadcastedA = broadcastTo($a, broadcastShape);
const $broadcastedB = broadcastTo($b, broadcastShape);
const inputs = {
condition: $broadcastedCondition,
t: $broadcastedA,
e: $broadcastedB
};
return ENGINE.runKernel(Select, inputs);
}
const where = /* @__PURE__ */ op({ where_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with all elements set to 0 with the same shape as the
* given tensor.
*
* ```js
* const x = tf.tensor([1, 2]);
* tf.zerosLike(x).print();
* ```
*
* @param x The tensor of required shape.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function zerosLike_(x) {
const $x = convertToTensor(x, 'x', 'zerosLike');
const inputs = { x: $x };
return ENGINE.runKernel(ZerosLike, inputs);
}
const zerosLike$2 = /* @__PURE__ */ op({ zerosLike_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes exponential linear element-wise: `x > 0 ? x : (e ^ x) - 1`.
*
* ```js
* const x = tf.tensor1d([-1, 1, -3, 2]);
*
* x.elu().print(); // or tf.elu(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function elu_(x) {
const $x = convertToTensor(x, 'x', 'elu', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Elu$1, inputs);
}
const elu$3 = /* @__PURE__ */ op({ elu_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes Gauss error function of the input `tf.Tensor` element-wise:
* `erf(x)`
*
* ```js
* const x = tf.tensor1d([0, .1, -.1, .7]);
*
* x.erf().print(); // or tf.erf(x);
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function erf_(x) {
let $x = convertToTensor(x, 'x', 'erf');
assert$1($x.dtype === 'int32' || $x.dtype === 'float32', () => 'Input dtype must be `int32` or `float32`.');
if ($x.dtype === 'int32') {
$x = cast$3($x, 'float32');
}
const inputs = { x: $x };
return ENGINE.runKernel(Erf, inputs);
}
const erf$2 = /* @__PURE__ */ op({ erf_ });
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns true if the axis specifies the inner most dimensions of the
* array.
*/
function axesAreInnerMostDims(axes, rank) {
for (let i = 0; i < axes.length; ++i) {
if (axes[axes.length - i - 1] !== rank - 1 - i) {
return false;
}
}
return true;
}
function combineLocations(outputLoc, reduceLoc, axes) {
const rank = outputLoc.length + reduceLoc.length;
const loc = [];
let outIdx = 0;
let reduceIdx = 0;
for (let dim = 0; dim < rank; dim++) {
if (axes.indexOf(dim) === -1) {
loc.push(outputLoc[outIdx++]);
}
else {
loc.push(reduceLoc[reduceIdx++]);
}
}
return loc;
}
function computeOutAndReduceShapes(aShape, axes) {
const outShape = [];
const rank = aShape.length;
for (let dim = 0; dim < rank; dim++) {
if (axes.indexOf(dim) === -1) {
outShape.push(aShape[dim]);
}
}
const reduceShape = axes.map(dim => aShape[dim]);
return [outShape, reduceShape];
}
function expandShapeToKeepDim(shape, axes) {
const reduceSubShape = axes.map(x => 1);
return combineLocations(shape, reduceSubShape, axes);
}
function assertAxesAreInnerMostDims(msg, axes, rank) {
assert$1(axesAreInnerMostDims(axes, rank), () => `${msg} supports only inner-most axes for now. ` +
`Got axes ${axes} and rank-${rank} input.`);
}
/**
* Returns the axes permutation to be used with `tf.transpose`, if such
* permutation is necessary. Otherwise it returns null. This method is used by
* operations that operate only on inner-most axes.
*/
function getAxesPermutation(axes, rank) {
if (axesAreInnerMostDims(axes, rank)) {
return null;
}
const result = [];
for (let i = 0; i < rank; ++i) {
if (axes.indexOf(i) === -1) {
result.push(i);
}
}
axes.forEach(axis => result.push(axis));
return result;
}
/** Returns the axes permutation that undoes the original permutation. */
function getUndoAxesPermutation(axes) {
return axes.map((axis, i) => [i, axis])
.sort((a, b) => a[1] - b[1])
.map(x => x[0]);
}
function getInnerMostAxes(numAxes, rank) {
const res = [];
for (let i = rank - numAxes; i < rank; ++i) {
res.push(i);
}
return res;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the maximum of elements across dimensions of a `tf.Tensor`.
*
* Reduces the input along the dimensions given in `axes`. Unless `keepDims`
* is true, the rank of the `tf.Tensor` is reduced by 1 for each entry in
* `axes`. If `keepDims` is true, the reduced dimensions are retained with
* length 1. If `axes` has no entries, all dimensions are reduced, and a
* `tf.Tensor` with a single element is returned.
*
* ```js
* const x = tf.tensor1d([1, 2, 3]);
*
* x.max().print(); // or tf.max(x)
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* const axis = 1;
* x.max(axis).print(); // or tf.max(x, axis)
* ```
*
* @param x The input tensor.
* @param axis The dimension(s) to reduce. By default it reduces
* all dimensions.
* @param keepDims If true, retains reduced dimensions with size 1.
*
* @doc {heading: 'Operations', subheading: 'Reduction'}
*/
function max_(x, axis = null, keepDims = false) {
const $x = convertToTensor(x, 'x', 'max');
const inputs = { x: $x };
const attrs = { reductionIndices: axis, keepDims };
return ENGINE.runKernel(Max, inputs, attrs);
}
const max$2 = /* @__PURE__ */ op({ max_ });
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the minimum value from the input.
*
* Reduces the input along the dimensions given in `axes`. Unless `keepDims`
* is true, the rank of the array is reduced by 1 for each entry in `axes`.
* If `keepDims` is true, the reduced dimensions are retained with length 1.
* If `axes` has no entries, all dimensions are reduced, and an array with a
* single element is returned.
*
* ```js
* const x = tf.tensor1d([1, 2, 3]);
*
* x.min().print(); // or tf.min(x)
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* const axis = 1;
* x.min(axis).print(); // or tf.min(x, axis)
* ```
*
* @param x The input Tensor.
* @param axis The dimension(s) to reduce. By default it reduces
* all dimensions.
* @param keepDims If true, retains reduced dimensions with size 1.
*
* @doc {heading: 'Operations', subheading: 'Reduction'}
*/
function min_(x, axis = null, keepDims = false) {
const $x = convertToTensor(x, 'x', 'min');
const inputs = { x: $x };
const attrs = { axis, keepDims };
// tslint:disable-next-line: no-unnecessary-type-assertion
return ENGINE.runKernel(Min, inputs, attrs);
}
const min$2 = /* @__PURE__ */ op({ min_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the power of one `tf.Tensor` to another. Supports broadcasting.
*
* Given a `tf.Tensor` x and a `tf.Tensor` y, this operation computes x^y for
* corresponding elements in x and y. The result's dtype will be the upcasted
* type of the `base` and `exp` dtypes.
*
* ```js
* const a = tf.tensor([[2, 3], [4, 5]])
* const b = tf.tensor([[1, 2], [3, 0]]).toInt();
*
* a.pow(b).print(); // or tf.pow(a, b)
* ```
*
* ```js
* const a = tf.tensor([[1, 2], [3, 4]])
* const b = tf.tensor(2).toInt();
*
* a.pow(b).print(); // or tf.pow(a, b)
* ```
* We also expose `powStrict` which has the same signature as this op and
* asserts that `base` and `exp` are the same shape (does not broadcast).
*
* @param base The base `tf.Tensor` to pow element-wise.
* @param exp The exponent `tf.Tensor` to pow element-wise.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function pow_(base, exp) {
let $base = convertToTensor(base, 'base', 'pow');
let $exp = convertToTensor(exp, 'exp', 'pow');
[$base, $exp] = makeTypesMatch($base, $exp);
const inputs = { a: $base, b: $exp };
return ENGINE.runKernel(Pow, inputs);
}
const pow$2 = /* @__PURE__ */ op({ pow_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** This is shared code across all tensor creation methods. */
function makeTensor(values, shape, inferredShape, dtype) {
if (dtype == null) {
dtype = inferDtype(values);
}
else if (dtype === 'complex64') {
throw new Error(`Cannot construct a complex64 tensor directly. ` +
`Please use tf.complex(real, imag).`);
}
if (isWebGPUData(values) || isWebGLData(values)) {
if (dtype !== 'float32' && dtype !== 'int32') {
throw new Error(`Creating tensor from GPU data only supports ` +
`'float32'|'int32' dtype, while the dtype is ${dtype}.`);
}
return ENGINE.backend.createTensorFromGPUData(values, shape || inferredShape, dtype);
}
if (!isTypedArray(values) && !Array.isArray(values) &&
typeof values !== 'number' && typeof values !== 'boolean' &&
typeof values !== 'string') {
throw new Error('values passed to tensor(values) must be a number/boolean/string or ' +
'an array of numbers/booleans/strings, or a TypedArray');
}
// Verify that the shape matches the inferred shape.
if (shape != null) {
assertNonNegativeIntegerDimensions(shape);
const providedSize = sizeFromShape(shape);
const inferredSize = sizeFromShape(inferredShape);
assert$1(providedSize === inferredSize, () => `Based on the provided shape, [${shape}], the tensor should have ` +
`${providedSize} values but has ${inferredSize}`);
for (let i = 0; i < inferredShape.length; ++i) {
const inferred = inferredShape[i];
const flatDimsDontMatch = i === inferredShape.length - 1 ?
inferred !== sizeFromShape(shape.slice(i)) :
true;
assert$1(inferredShape[i] === shape[i] || !flatDimsDontMatch, () => `Error creating a new Tensor. Inferred shape ` +
`(${inferredShape}) does not match the provided ` +
`shape (${shape}). `);
}
}
if (!isTypedArray(values) && !Array.isArray(values)) {
values = [values];
}
shape = shape || inferredShape;
values = dtype !== 'string' ?
toTypedArray(values, dtype) :
flatten$1(values, [], true);
return ENGINE.makeTensor(values, shape, dtype);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates rank-0 `tf.Tensor` (scalar) with the provided value and dtype.
*
* The same functionality can be achieved with `tf.tensor`, but in general
* we recommend using `tf.scalar` as it makes the code more readable.
*
* ```js
* tf.scalar(3.14).print();
* ```
*
* @param value The value of the scalar.
* @param dtype The data type.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function scalar(value, dtype) {
if (((isTypedArray(value) && dtype !== 'string') || Array.isArray(value)) &&
dtype !== 'complex64') {
throw new Error('Error creating a new Scalar: value must be a primitive ' +
'(number|boolean|string)');
}
if (dtype === 'string' && isTypedArray(value) &&
!(value instanceof Uint8Array)) {
throw new Error('When making a scalar from encoded string, ' +
'the value must be `Uint8Array`.');
}
const shape = [];
const inferredShape = [];
return makeTensor(value, shape, inferredShape, dtype);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes square root of the input `tf.Tensor` element-wise: `y = sqrt(x)`
*
* ```js
* const x = tf.tensor1d([1, 2, 4, -1]);
*
* x.sqrt().print(); // or tf.sqrt(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function sqrt_(x) {
const $x = convertToTensor(x, 'x', 'sqrt', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Sqrt, inputs);
}
const sqrt$2 = /* @__PURE__ */ op({ sqrt_ });
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes square of `x` element-wise: `x ^ 2`
*
* ```js
* const x = tf.tensor1d([1, 2, Math.sqrt(2), -1]);
*
* x.square().print(); // or tf.square(x)
* ```
* @param x The input Tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function square_(x) {
const $x = convertToTensor(x, 'x', 'square');
const attrs = {};
return ENGINE.runKernel('Square', { x: $x }, attrs);
}
const square$2 = /* @__PURE__ */ op({ square_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the sum of elements across dimensions of a `tf.Tensor`.
*
* Reduces the input along the dimensions given in `axes`. Unless `keepDims`
* is true, the rank of the `tf.Tensor` is reduced by 1 for each entry in
* `axes`. If `keepDims` is true, the reduced dimensions are retained with
* length 1. If axes has no entries, all dimensions are reduced, and a
* `tf.Tensor` with a single element is returned.
*
* ```js
* const x = tf.tensor1d([1, 2, 3]);
*
* x.sum().print(); // or tf.sum(x)
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* const axis = 1;
* x.sum(axis).print(); // or tf.sum(x, axis)
* ```
*
* @param x The input tensor to compute the sum over. If the dtype is `bool`
* it will be converted to `int32` and the output dtype will be `int32`.
* @param axis The dimension(s) to reduce. By default it reduces
* all dimensions.
* @param keepDims If true, retains reduced dimensions with size 1.
*
* @doc {heading: 'Operations', subheading: 'Reduction'}
*/
function sum_(x, axis = null, keepDims = false) {
let $x = convertToTensor(x, 'x', 'sum');
if ($x.dtype === 'bool') {
$x = cast$3($x, 'int32');
}
const inputs = { x: $x };
const attrs = { axis, keepDims };
return ENGINE.runKernel(Sum, inputs, attrs);
}
const sum$2 = /* @__PURE__ */ op({ sum_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the norm of scalar, vectors, and matrices.
* This function can compute several different vector norms (the 1-norm, the
* Euclidean or 2-norm, the inf-norm, and in general the p-norm for p > 0)
* and matrix norms (Frobenius, 1-norm, and inf-norm).
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
*
* x.norm().print(); // or tf.norm(x)
* ```
*
* @param x The input array.
* @param ord Optional. Order of the norm. Supported norm types are
* following:
*
* | ord | norm for matrices | norm for vectors
* |------------|---------------------------|---------------------
* |'euclidean' |Frobenius norm |2-norm
* |'fro' |Frobenius norm |
* |Infinity |max(sum(abs(x), axis=1)) |max(abs(x))
* |-Infinity |min(sum(abs(x), axis=1)) |min(abs(x))
* |1 |max(sum(abs(x), axis=0)) |sum(abs(x))
* |2 | |sum(abs(x)^2)^(1/2)
*
* @param axis Optional. If axis is null (the default), the input is
* considered a vector and a single vector norm is computed over the entire
* set of values in the Tensor, i.e. norm(x, ord) is equivalent
* to norm(x.reshape([-1]), ord). If axis is an integer, the input
* is considered a batch of vectors, and axis determines the axis in x
* over which to compute vector norms. If axis is a 2-tuple of integer it is
* considered a batch of matrices and axis determines the axes in NDArray
* over which to compute a matrix norm.
* @param keepDims Optional. If true, the norm has the same dimensionality
* as the input.
*
* @doc {heading: 'Operations', subheading: 'Matrices'}
*/
function norm_(x, ord = 'euclidean', axis = null, keepDims = false) {
x = convertToTensor(x, 'x', 'norm');
const norm = normImpl(x, ord, axis);
let keepDimsShape = norm.shape;
if (keepDims) {
const axes = parseAxisParam(axis, x.shape);
keepDimsShape = expandShapeToKeepDim(norm.shape, axes);
}
return reshape$2(norm, keepDimsShape);
}
function normImpl(x, p, axis = null) {
if (x.rank === 0) {
return abs$2(x);
}
// consider vector when no axis is specified
if (x.rank !== 1 && axis === null) {
return normImpl(reshape$2(x, [-1]), p, axis);
}
// vector
if (x.rank === 1 || typeof axis === 'number' ||
Array.isArray(axis) && axis.length === 1) {
if (p === 1) {
return sum$2(abs$2(x), axis);
}
if (p === Infinity) {
return max$2(abs$2(x), axis);
}
if (p === -Infinity) {
return min$2(abs$2(x), axis);
}
if (p === 'euclidean' || p === 2) {
// norm(x, 2) = sum(abs(xi) ^ 2) ^ 1/2
return sqrt$2(sum$2(pow$2(abs$2(x), scalar(2, 'int32')), axis));
}
throw new Error(`Error in norm: invalid ord value: ${p}`);
}
// matrix (assumption axis[0] < axis[1])
if (Array.isArray(axis) && axis.length === 2) {
if (p === 1) {
return max$2(sum$2(abs$2(x), axis[0]), axis[1] - 1);
}
if (p === Infinity) {
return max$2(sum$2(abs$2(x), axis[1]), axis[0]);
}
if (p === -Infinity) {
return min$2(sum$2(abs$2(x), axis[1]), axis[0]);
}
if (p === 'fro' || p === 'euclidean') {
// norm(x) = sqrt(sum(pow(x, 2)))
return sqrt$2(sum$2(square$2(x), axis));
}
throw new Error(`Error in norm: invalid ord value: ${p}`);
}
throw new Error(`Error in norm: invalid axis: ${axis}`);
}
const norm = /* @__PURE__ */ op({ norm_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes exponential of the input `tf.Tensor` element-wise. `e ^ x`
*
* ```js
* const x = tf.tensor1d([1, 2, -3]);
*
* x.exp().print(); // or tf.exp(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function exp_(x) {
const $x = convertToTensor(x, 'x', 'exp');
const inputs = { x: $x };
return ENGINE.runKernel(Exp, inputs);
}
const exp$2 = /* @__PURE__ */ op({ exp_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns a `tf.Tensor` that has expanded rank, by inserting a dimension
* into the tensor's shape.
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
* const axis = 1;
* x.expandDims(axis).print();
* ```
*
* @param x The input tensor whose dimensions are to be expanded.
* @param axis The dimension index at which to insert shape of `1`. Defaults
* to 0 (the first dimension).
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function expandDims_(x, axis = 0) {
const $x = convertToTensor(x, 'x', 'expandDims', 'string_or_numeric');
assert$1(axis <= $x.rank, () => 'Axis must be <= rank of the tensor');
const inputs = { input: $x };
const attrs = { dim: axis };
return ENGINE.runKernel(ExpandDims, inputs, attrs);
}
const expandDims$3 = /* @__PURE__ */ op({ expandDims_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Construct a tensor by repeating it the number of times given by reps.
*
* This operation creates a new tensor by replicating `input` `reps`
* times. The output tensor's `i`th dimension has `input.shape[i] *
* reps[i]` elements, and the values of `input` are replicated
* `reps[i]` times along the `i`th dimension. For example, tiling
* `[a, b, c, d]` by `[2]` produces `[a, b, c, d, a, b, c, d]`.
*
* ```js
* const a = tf.tensor1d([1, 2]);
*
* a.tile([2]).print(); // or tf.tile(a, [2])
* ```
*
* ```js
* const a = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* a.tile([1, 2]).print(); // or tf.tile(a, [1,2])
* ```
* @param x The tensor to tile.
* @param reps Determines the number of replications per dimension.
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function tile_(x, reps) {
const $x = convertToTensor(x, 'x', 'tile', 'string_or_numeric');
assert$1($x.rank === reps.length, () => `Error in transpose: rank of input ${$x.rank} ` +
`must match length of reps ${reps}.`);
const inputs = { x: $x };
const attrs = { reps };
return ENGINE.runKernel(Tile, inputs, attrs);
}
const tile$3 = /* @__PURE__ */ op({ tile_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Create an identity matrix.
*
* @param numRows Number of rows.
* @param numColumns Number of columns. Defaults to `numRows`.
* @param batchShape If provided, will add the batch shape to the beginning
* of the shape of the returned `tf.Tensor` by repeating the identity
* matrix.
* @param dtype Data type.
* @returns Identity matrix of the specified size and data type, possibly
* with batch repetition if `batchShape` is specified.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function eye_(numRows, numColumns, batchShape, dtype = 'float32') {
if (numColumns == null) {
numColumns = numRows;
}
const buff = buffer([numRows, numColumns], dtype);
const n = numRows <= numColumns ? numRows : numColumns;
for (let i = 0; i < n; ++i) {
buff.set(1, i, i);
}
const out = reshape$2(buff.toTensor(), [numRows, numColumns]);
if (batchShape == null) {
return out;
}
else {
if (batchShape.length === 1) {
return tile$3(expandDims$3(out, 0), [batchShape[0], 1, 1]);
}
else if (batchShape.length === 2) {
// tslint:disable-next-line:no-unnecessary-type-assertion
return tile$3(expandDims$3(expandDims$3(out, 0), 0), [batchShape[0], batchShape[1], 1, 1]);
}
else if (batchShape.length === 3) {
// tslint:disable-next-line:no-unnecessary-type-assertion
return tile$3(expandDims$3(expandDims$3(expandDims$3(out, 0), 0), 0), [
batchShape[0], batchShape[1], batchShape[2], 1, 1
]);
}
else {
throw new Error(`eye() currently supports only 1D and 2D ` +
// tslint:disable-next-line:no-any
`batchShapes, but received ${batchShape.length}D.`);
}
}
}
const eye = /* @__PURE__ */ op({ eye_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes floor of input `tf.Tensor` element-wise: `floor(x)`.
*
* ```js
* const x = tf.tensor1d([.6, 1.1, -3.3]);
*
* x.floor().print(); // or tf.floor(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function floor_(x) {
const $x = convertToTensor(x, 'x', 'floor', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Floor, inputs);
}
const floor$2 = /* @__PURE__ */ op({ floor_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Gather slices from tensor `x`'s axis `axis` according to `indices`.
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
* const indices = tf.tensor1d([1, 3, 3], 'int32');
*
* x.gather(indices).print();
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
* const indices = tf.tensor1d([1, 1, 0], 'int32');
*
* x.gather(indices).print();
* ```
* @param x The input tensor whose slices are to be gathered.
* @param indices The indices of the values to extract.
* @param axis The axis over which to select values. Defaults to 0.
* @param batchDims Optional. The number of batch dimensions. It must be less
* than or equal to rank(indices). Defaults to 0.
* The output tensor will have shape of
* `x.shape[:axis] + indices.shape[batchDims:] + x.shape[axis + 1:]`
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function gather_(x, indices, axis = 0, batchDims = 0) {
const $x = convertToTensor(x, 'x', 'gather');
const $indices = convertToTensor(indices, 'indices', 'gather', 'int32');
const inputs = { x: $x, indices: $indices };
const attrs = { axis, batchDims };
return ENGINE.runKernel(GatherV2, inputs, attrs);
}
const gather$1 = /* @__PURE__ */ op({ gather_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of (a > b) element-wise. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
* const b = tf.tensor1d([2, 2, 2]);
*
* a.greater(b).print();
* ```
*
* @param a The first input tensor.
* @param b The second input tensor. Must have the same dtype as `a`.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function greater_(a, b) {
let $a = convertToTensor(a, 'a', 'greater', 'string_or_numeric');
let $b = convertToTensor(b, 'b', 'greater', 'string_or_numeric');
[$a, $b] = makeTypesMatch($a, $b);
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Greater, inputs);
}
const greater$2 = /* @__PURE__ */ op({ greater_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of (a >= b) element-wise. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
* const b = tf.tensor1d([2, 2, 2]);
*
* a.greaterEqual(b).print();
* ```
*
* @param a The first input tensor.
* @param b The second input tensor. Must have the same dtype as `a`.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function greaterEqual_(a, b) {
let $a = convertToTensor(a, 'a', 'greaterEqual', 'string_or_numeric');
let $b = convertToTensor(b, 'b', 'greaterEqual', 'string_or_numeric');
[$a, $b] = makeTypesMatch($a, $b);
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(GreaterEqual, inputs);
}
const greaterEqual$2 = /* @__PURE__ */ op({ greaterEqual_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the imaginary part of a complex (or real) tensor.
*
* Given a tensor input, this operation returns a tensor of type float that is
* the imaginary part of each element in input considered as a complex number.
* If input is real, a tensor of all zeros is returned.
*
* ```js
* const x = tf.complex([-2.25, 3.25], [4.75, 5.75]);
* tf.imag(x).print();
* ```
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function imag_(input) {
const $input = convertToTensor(input, 'input', 'imag');
const inputs = { input: $input };
return ENGINE.runKernel(Imag, inputs);
}
const imag$2 = /* @__PURE__ */ op({ imag_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes leaky rectified linear element-wise.
*
* See
* [http://web.stanford.edu/~awni/papers/relu_hybrid_icml2013_final.pdf](
* http://web.stanford.edu/~awni/papers/relu_hybrid_icml2013_final.pdf)
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.leakyRelu(0.1).print(); // or tf.leakyRelu(x, 0.1)
* ```
* @param x The input tensor.
* @param alpha The scaling factor for negative values, defaults to 0.2.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function leakyRelu_(x, alpha = 0.2) {
const $x = convertToTensor(x, 'x', 'leakyRelu');
const inputs = { x: $x };
const attrs = { alpha };
return ENGINE.runKernel(LeakyRelu, inputs, attrs);
}
const leakyRelu$2 = /* @__PURE__ */ op({ leakyRelu_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of (a < b) element-wise. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
* const b = tf.tensor1d([2, 2, 2]);
*
* a.less(b).print();
* ```
* @param a The first input tensor.
* @param b The second input tensor. Must have the same dtype as `a`.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function less_(a, b) {
let $a = convertToTensor(a, 'a', 'less', 'string_or_numeric');
let $b = convertToTensor(b, 'b', 'less', 'string_or_numeric');
[$a, $b] = makeTypesMatch($a, $b);
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Less, inputs);
}
const less$2 = /* @__PURE__ */ op({ less_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of (a <= b) element-wise. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
* const b = tf.tensor1d([2, 2, 2]);
*
* a.lessEqual(b).print();
* ```
*
* @param a The first input tensor.
* @param b The second input tensor. Must have the same dtype as `a`.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function lessEqual_(a, b) {
let $a = convertToTensor(a, 'a', 'lessEqual', 'string_or_numeric');
let $b = convertToTensor(b, 'b', 'lessEqual', 'string_or_numeric');
[$a, $b] = makeTypesMatch($a, $b);
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(LessEqual, inputs);
}
const lessEqual$2 = /* @__PURE__ */ op({ lessEqual_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes natural logarithm of the input `tf.Tensor` element-wise: `ln(x)`
*
* ```js
* const x = tf.tensor1d([1, 2, Math.E]);
*
* x.log().print(); // or tf.log(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function log_(x) {
const $x = convertToTensor(x, 'x', 'log', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Log, inputs);
}
const log$2 = /* @__PURE__ */ op({ log_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes natural logarithm of the input `tf.Tensor` plus one
* element-wise: `ln(1 + x)`
*
* ```js
* const x = tf.tensor1d([1, 2, Math.E - 1]);
*
* x.log1p().print(); // or tf.log1p(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function log1p_(x) {
const $x = convertToTensor(x, 'x', 'log1p');
const inputs = { x: $x };
return ENGINE.runKernel(Log1p, inputs);
}
const log1p$2 = /* @__PURE__ */ op({ log1p_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes and returns the gradient of f(x) with respect to the list of
* trainable variables provided by `varList`. If no list is provided, it
* defaults to all trainable variables.
*
* ```js
* const a = tf.variable(tf.tensor1d([3, 4]));
* const b = tf.variable(tf.tensor1d([5, 6]));
* const x = tf.tensor1d([1, 2]);
*
* // f(a, b) = a * x ^ 2 + b * x
* const f = () => a.mul(x.square()).add(b.mul(x)).sum();
* // df/da = x ^ 2, df/db = x
* const {value, grads} = tf.variableGrads(f);
*
* Object.keys(grads).forEach(varName => grads[varName].print());
* ```
*
* @param f The function to execute. f() should return a scalar.
* @param varList The list of variables to compute the gradients with respect
* to. Defaults to all trainable variables.
* @returns An object with the following keys and values:
* - `value`: The value of the function `f`.
* - `grads`: A map from the names of the variables to the gradients.
* If the `varList` argument is provided explicitly and contains a subset of
* non-trainable variables, this map in the return value will contain keys
* that map the names of the non-trainable variables to `null`.
*
* @doc {heading: 'Training', subheading: 'Gradients'}
*/
function variableGrads(f, varList) {
assert$1(isFunction(f), () => 'The f passed in variableGrads(f) must be a function');
assert$1(varList == null ||
Array.isArray(varList) && varList.every(v => v instanceof Variable), () => 'The varList passed in variableGrads(f, varList) must be an array ' +
'of variables');
const specifiedVarList = varList != null;
if (!specifiedVarList) {
// Get all of the trainable variables.
varList = [];
for (const varName in ENGINE.registeredVariables) {
varList.push(ENGINE.registeredVariables[varName]);
}
}
const specifiedNonTrainable = specifiedVarList ? varList.filter(variable => !variable.trainable) : null;
// Prune non-trainable variables.
const originalVarCount = varList.length;
varList = varList.filter(variable => variable.trainable);
assert$1(varList.length > 0, () => `variableGrads() expects at least one of the input variables to ` +
`be trainable, but none of the ${originalVarCount} variables is ` +
`trainable.`);
const allowNoGradients = true;
const { value, grads } = ENGINE.gradients(f, varList, null, allowNoGradients);
assert$1(grads.some(g => g != null), () => 'Cannot find a connection between any variable and the result of ' +
'the loss function y=f(x). Please make sure the operations that ' +
'use variables are inside the function f passed to minimize().');
assert$1(value.rank === 0, () => `The f passed in variableGrads(f) must return a scalar, but it ` +
`returned a rank-${value.rank} tensor`);
const namedGrads = {};
varList.forEach((v, i) => {
if (grads[i] != null) {
namedGrads[v.name] = grads[i];
}
});
if (specifiedNonTrainable != null) {
// If varList is explicitly provided and contains non-trainable values,
// add them to the returned gradients with `null` values.
specifiedNonTrainable.forEach(v => namedGrads[v.name] = null);
}
return { value, grads: namedGrads };
}
/**
* Overrides the gradient computation of a function `f`.
*
* Takes a function
* `f(...inputs, save) => {value: Tensor, gradFunc: (dy, saved) => Tensor[]}`
* and returns another function `g(...inputs)` which takes the same inputs as
* `f`. When called, `g` returns `f().value`. In backward mode, custom gradients
* with respect to each input of `f` are computed using `f().gradFunc`.
*
* The `save` function passed to `f` should be used for saving tensors needed
* in the gradient. And the `saved` passed to the `gradFunc` is a
* `NamedTensorMap`, which contains those saved tensors.
*
* ```js
* const customOp = tf.customGrad((x, save) => {
* // Save x to make sure it's available later for the gradient.
* save([x]);
* // Override gradient of our custom x ^ 2 op to be dy * abs(x);
* return {
* value: x.square(),
* // Note `saved.x` which points to the `x` we saved earlier.
* gradFunc: (dy, saved) => [dy.mul(saved[0].abs())]
* };
* });
*
* const x = tf.tensor1d([-1, -2, 3]);
* const dx = tf.grad(x => customOp(x));
*
* console.log(`f(x):`);
* customOp(x).print();
* console.log(`f'(x):`);
* dx(x).print();
* ```
*
* @param f The function to evaluate in forward mode, which should return
* `{value: Tensor, gradFunc: (dy, saved) => Tensor[]}`, where `gradFunc`
* returns the custom gradients of `f` with respect to its inputs.
*
* @doc {heading: 'Training', subheading: 'Gradients'}
*/
function customGrad(f) {
return ENGINE.customGrad(f);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes `-1 * x` element-wise.
*
* ```js
* const x = tf.tensor2d([1, 2, -2, 0], [2, 2]);
*
* x.neg().print(); // or tf.neg(x)
* ```
*
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function neg_(x) {
const $x = convertToTensor(x, 'x', 'neg');
const inputs = { x: $x };
return ENGINE.runKernel(Neg, inputs);
}
const neg$2 = /* @__PURE__ */ op({ neg_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes softplus of the input `tf.Tensor` element-wise: `log(exp(x) + 1)`
*
* ```js
* const x = tf.tensor1d([0, 1, -1, .7]);
*
* x.softplus().print(); // or tf.softplus(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function softplus_(x) {
const $x = convertToTensor(x, 'x', 'softplus');
const inputs = { x: $x };
return ENGINE.runKernel(Softplus$1, inputs);
}
const softplus$2 = /* @__PURE__ */ op({ softplus_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Subtracts two `tf.Tensor`s element-wise, A - B. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([10, 20, 30, 40]);
* const b = tf.tensor1d([1, 2, 3, 4]);
*
* a.sub(b).print(); // or tf.sub(a, b)
* ```
*
* ```js
* // Broadcast subtract a with b.
* const a = tf.tensor1d([10, 20, 30, 40]);
* const b = tf.scalar(5);
*
* a.sub(b).print(); // or tf.sub(a, b)
* ```
* @param a The first `tf.Tensor` to subtract from.
* @param b The second `tf.Tensor` to be subtracted. Must have the same dtype as
* `a`.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function sub_(a, b) {
let $a = convertToTensor(a, 'a', 'sub');
let $b = convertToTensor(b, 'b', 'sub');
[$a, $b] = makeTypesMatch($a, $b);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Sub, inputs);
}
const sub$2 = /* @__PURE__ */ op({ sub_ });
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the log softmax.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
*
* a.logSoftmax().print(); // or tf.logSoftmax(a)
* ```
*
* ```js
* const a = tf.tensor2d([2, 4, 6, 1, 2, 3], [2, 3]);
*
* a.logSoftmax().print(); // or tf.logSoftmax(a)
* ```
*
* @param logits The logits array.
* @param axis The dimension softmax would be performed on. Defaults to `-1`
* which indicates the last dimension.
*
* @doc {heading: 'Operations', subheading: 'Normalization'}
*/
function logSoftmax_(logits, axis = -1) {
const $logits = convertToTensor(logits, 'logits', 'logSoftmax');
if (axis === -1) {
axis = $logits.rank - 1;
}
if (axis !== $logits.rank - 1) {
throw Error('Log Softmax along a non-last dimension is not yet supported. ' +
`Logits was rank ${$logits.rank} and axis was ${axis}`);
}
// const forward: ForwardFunc<Tensor> = (backend, save) => {
// const keepDims = true;
// const xMax = max(logits, axis, true);
// const shifted = sub(logits, xMax);
// const value =
// sub(cast(shifted, 'float32'), log(sum(exp(shifted), axis,
// keepDims)));
// save([value]);
// return value;
// };
// Use a custom gradient for numerical stability.
const customOp = customGrad((logits, save) => {
const keepDims = true;
const xMax = max$2(logits, axis, true);
const shifted = sub$2(logits, xMax);
const value = sub$2(cast$3(shifted, 'float32'), log$2(sum$2(exp$2(shifted), axis, keepDims)));
save([value]);
const gradFunc = (dy, saved) => {
const [value] = saved;
const keepDims = true;
const softmax = exp$2(value);
return sub$2(dy, mul(sum$2(dy, axis, keepDims), softmax));
};
return { value, gradFunc };
});
return customOp($logits);
// TODO Use Engine.runKernel when CPU/WebGL/WASM backends implement this.
// const inputs: LogSoftmaxInputs = {logits: $logits};
// const attrs: LogSoftmaxAttrs = {axis};
// return ENGINE.runKernel(
// LogSoftmax, inputs as unknown as NamedTensorMap,
// attrs as unknown as NamedAttrMap);
}
const logSoftmax = /* @__PURE__ */ op({ logSoftmax_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of `a AND b` element-wise. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([false, false, true, true], 'bool');
* const b = tf.tensor1d([false, true, false, true], 'bool');
*
* a.logicalAnd(b).print();
* ```
*
* @param a The first input tensor. Must be of dtype bool.
* @param b The second input tensor. Must be of dtype bool.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function logicalAnd_(a, b) {
const $a = convertToTensor(a, 'a', 'logicalAnd', 'bool');
const $b = convertToTensor(b, 'b', 'logicalAnd', 'bool');
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(LogicalAnd, inputs);
}
const logicalAnd$2 = /* @__PURE__ */ op({ logicalAnd_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of `NOT x` element-wise.
*
* ```js
* const a = tf.tensor1d([false, true], 'bool');
*
* a.logicalNot().print();
* ```
*
* @param x The input tensor. Must be of dtype 'bool'.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function logicalNot_(x) {
const $x = convertToTensor(x, 'x', 'logicalNot', 'bool');
const inputs = { x: $x };
return ENGINE.runKernel(LogicalNot, inputs);
}
const logicalNot$2 = /* @__PURE__ */ op({ logicalNot_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the max of a and b (`a > b ? a : b`) element-wise.
* Supports broadcasting.
*
* We also expose `tf.maximumStrict` which has the same signature as this op and
* asserts that `a` and `b` are the same shape (does not broadcast).
*
* ```js
* const a = tf.tensor1d([1, 4, 3, 16]);
* const b = tf.tensor1d([1, 2, 9, 4]);
*
* a.maximum(b).print(); // or tf.maximum(a, b)
* ```
*
* ```js
* // Broadcast maximum a with b.
* const a = tf.tensor1d([2, 4, 6, 8]);
* const b = tf.scalar(5);
*
* a.maximum(b).print(); // or tf.maximum(a, b)
* ```
*
* @param a The first tensor.
* @param b The second tensor. Must have the same type as `a`.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function maximum_(a, b) {
let $a = convertToTensor(a, 'a', 'maximum');
let $b = convertToTensor(b, 'b', 'maximum');
[$a, $b] = makeTypesMatch($a, $b);
if ($a.dtype === 'bool') {
$a = cast$3($a, 'int32');
$b = cast$3($b, 'int32');
}
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Maximum, inputs);
}
const maximum$2 = /* @__PURE__ */ op({ maximum_ });
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the mean of elements across dimensions of a `tf.Tensor`.
*
* Reduces `x` along the dimensions given in `axis`. Unless `keepDims` is
* true, the rank of the `tf.Tensor` is reduced by 1 for each entry in `axis`.
* If `keepDims` is true, the reduced dimensions are retained with length 1.
* If `axis` has no entries, all dimensions are reduced, and a `tf.Tensor` with
* a single element is returned.
*
* ```js
* const x = tf.tensor1d([1, 2, 3]);
*
* x.mean().print(); // or tf.mean(a)
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* const axis = 1;
* x.mean(axis).print(); // or tf.mean(x, axis)
* ```
*
* @param x The input tensor.
* @param axis The dimension(s) to reduce. By default it reduces
* all dimensions.
* @param keepDims If true, retains reduced dimensions with size 1.
*
* @doc {heading: 'Operations', subheading: 'Reduction'}
*/
function mean_(x, axis = null, keepDims = false) {
const $x = convertToTensor(x, 'x', 'mean');
const inputs = { x: $x };
const attrs = { axis, keepDims };
return ENGINE.runKernel(Mean, inputs, attrs);
}
const mean$1 = /* @__PURE__ */ op({ mean_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with all elements set to 0.
*
* ```js
* tf.zeros([2, 2]).print();
* ```
*
* @param shape An array of integers defining the output tensor shape.
* @param dtype The type of an element in the resulting tensor. Can
* be 'float32', 'int32' or 'bool'. Defaults to 'float'.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function zeros$1(shape, dtype = 'float32') {
assertNonNegativeIntegerDimensions(shape);
if (dtype === 'complex64') {
const real = zeros$1(shape, 'float32');
const imag = zeros$1(shape, 'float32');
return complex$2(real, imag);
}
const values = makeZerosTypedArray(sizeFromShape(shape), dtype);
return ENGINE.makeTensor(values, shape, dtype);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with all elements set to 1.
*
* ```js
* tf.ones([2, 2]).print();
* ```
*
* @param shape An array of integers defining the output tensor shape.
* @param dtype The type of an element in the resulting tensor. Defaults to
* 'float'.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function ones(shape, dtype = 'float32') {
assertNonNegativeIntegerDimensions(shape);
if (dtype === 'complex64') {
const real = ones(shape, 'float32');
const imag = zeros$1(shape, 'float32');
return complex$2(real, imag);
}
const values = makeOnesTypedArray(sizeFromShape(shape), dtype);
return ENGINE.makeTensor(values, shape, dtype);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the min of a and b (`a < b ? a : b`) element-wise.
* Supports broadcasting.
*
* We also expose `minimumStrict` which has the same signature as this op and
* asserts that `a` and `b` are the same shape (does not broadcast).
*
* ```js
* const a = tf.tensor1d([1, 4, 3, 16]);
* const b = tf.tensor1d([1, 2, 9, 4]);
*
* a.minimum(b).print(); // or tf.minimum(a, b)
* ```
*
* ```js
* // Broadcast minimum a with b.
* const a = tf.tensor1d([2, 4, 6, 8]);
* const b = tf.scalar(5);
*
* a.minimum(b).print(); // or tf.minimum(a, b)
* ```
*
* @param a The first tensor.
* @param b The second tensor. Must have the same type as `a`.
*
* @doc {heading: 'Operations', subheading: 'Arithmetic'}
*/
function minimum_(a, b) {
let $a = convertToTensor(a, 'a', 'minimum');
let $b = convertToTensor(b, 'b', 'minimum');
[$a, $b] = makeTypesMatch($a, $b);
if ($a.dtype === 'bool') {
$a = cast$3($a, 'int32');
$b = cast$3($b, 'int32');
}
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(Minimum, inputs);
}
const minimum$2 = /* @__PURE__ */ op({ minimum_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the truth value of (a != b) element-wise. Supports broadcasting.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
* const b = tf.tensor1d([0, 2, 3]);
*
* a.notEqual(b).print();
* ```
* @param a The first input tensor.
* @param b The second input tensor. Must have the same dtype as `a`.
*
* @doc {heading: 'Operations', subheading: 'Logical'}
*/
function notEqual_(a, b) {
let $a = convertToTensor(a, 'a', 'notEqual', 'string_or_numeric');
let $b = convertToTensor(b, 'b', 'notEqual', 'string_or_numeric');
[$a, $b] = makeTypesMatch($a, $b);
assertAndGetBroadcastShape($a.shape, $b.shape);
const inputs = { a: $a, b: $b };
return ENGINE.runKernel(NotEqual, inputs);
}
const notEqual$2 = /* @__PURE__ */ op({ notEqual_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a one-hot `tf.Tensor`. The locations represented by `indices` take
* value `onValue` (defaults to 1), while all other locations take value
* `offValue` (defaults to 0). If `indices` is rank `R`, the output has rank
* `R+1` with the last axis of size `depth`.
* `indices` used to encode prediction class must start from 0. For example,
* if you have 3 classes of data, class 1 should be encoded as 0, class 2
* should be 1, and class 3 should be 2.
*
* ```js
* tf.oneHot(tf.tensor1d([0, 1], 'int32'), 3).print();
* ```
*
* @param indices `tf.Tensor` of indices with dtype `int32`. Indices must
* start from 0.
* @param depth The depth of the one hot dimension.
* @param onValue A number used to fill in the output when the index matches
* the location.
* @param offValue A number used to fill in the output when the index does
* not match the location.
* @param dtype The dtype of the output tensor, default to 'int32'.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function oneHot_(indices, depth, onValue = 1, offValue = 0, dtype = 'int32') {
if (depth < 2) {
throw new Error(`Error in oneHot: depth must be >=2, but it is ${depth}`);
}
const $indices = convertToTensor(indices, 'indices', 'oneHot', 'int32');
const inputs = { indices: $indices };
const attrs = { dtype, depth, onValue, offValue };
return ENGINE.runKernel(OneHot, inputs, attrs);
}
const oneHot$2 = /* @__PURE__ */ op({ oneHot_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with all elements set to 1 with the same shape as the
* given tensor.
*
* ```js
* const x = tf.tensor([1, 2]);
* tf.onesLike(x).print();
* ```
* @param x A tensor.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function onesLike_(x) {
const $x = convertToTensor(x, 'x', 'onesLike');
const inputs = { x: $x };
return ENGINE.runKernel(OnesLike, inputs);
}
const onesLike$2 = /* @__PURE__ */ op({ onesLike_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Pads a `tf.Tensor` with a given value and paddings.
*
* This operation implements `CONSTANT` mode. For `REFLECT` and `SYMMETRIC`,
* refer to `tf.mirrorPad`.
*
* Also available are stricter rank-specific methods with the same signature
* as this method that assert that `paddings` is of given length.
* - `tf.pad1d`
* - `tf.pad2d`
* - `tf.pad3d`
* - `tf.pad4d`
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
* x.pad([[1, 2]]).print();
* ```
* @param x The tensor to pad.
* @param paddings An array of length `R` (the rank of the tensor), where
* each element is a length-2 tuple of ints `[padBefore, padAfter]`,
* specifying how much to pad along each dimension of the tensor.
* @param constantValue The pad value to use. Defaults to 0.
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function pad_(x, paddings, constantValue = 0) {
const $x = convertToTensor(x, 'x', 'pad');
if ($x.rank === 0) {
throw new Error('pad(scalar) is not defined. Pass non-scalar to pad');
}
const attrs = { paddings, constantValue };
const inputs = { x: $x };
return ENGINE.runKernel(PadV2, inputs, attrs);
}
const pad = /* @__PURE__ */ op({ pad_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* This operation divides "spatial" dimensions `[1, ..., M]` of the input into
* a grid of blocks of shape `blockShape`, and interleaves these blocks with
* the "batch" dimension (0) such that in the output, the spatial
* dimensions `[1, ..., M]` correspond to the position within the grid,
* and the batch dimension combines both the position within a spatial block
* and the original batch position. Prior to division into blocks,
* the spatial dimensions of the input are optionally zero padded
* according to `paddings`. See below for a precise description.
*
* ```js
* const x = tf.tensor4d([1, 2, 3, 4], [1, 2, 2, 1]);
* const blockShape = [2, 2];
* const paddings = [[0, 0], [0, 0]];
*
* x.spaceToBatchND(blockShape, paddings).print();
* ```
*
* @param x A `tf.Tensor`. N-D with `x.shape` = `[batch] + spatialShape +
* remainingShape`, where spatialShape has `M` dimensions.
* @param blockShape A 1-D array. Must have shape `[M]`, all values must
* be >= 1.
* @param paddings A 2-D array. Must have shape `[M, 2]`, all values must be >=
* 0. `paddings[i] = [padStart, padEnd]` specifies the amount to zero-pad
* from input dimension `i + 1`, which corresponds to spatial dimension `i`. It
* is required that
* `(inputShape[i + 1] + padStart + padEnd) % blockShape[i] === 0`
*
* This operation is equivalent to the following steps:
*
* 1. Zero-pad the start and end of dimensions `[1, ..., M]` of the input
* according to `paddings` to produce `padded` of shape paddedShape.
*
* 2. Reshape `padded` to `reshapedPadded` of shape:
* `[batch] + [paddedShape[1] / blockShape[0], blockShape[0], ...,
* paddedShape[M] / blockShape[M-1], blockShape[M-1]] + remainingShape`
*
* 3. Permute dimensions of `reshapedPadded` to produce `permutedReshapedPadded`
* of shape: `blockShape + [batch] + [paddedShape[1] / blockShape[0], ...,
* paddedShape[M] / blockShape[M-1]] + remainingShape`
*
* 4. Reshape `permutedReshapedPadded` to flatten `blockShape` into the
* batch dimension, producing an output tensor of shape:
* `[batch * prod(blockShape)] + [paddedShape[1] / blockShape[0], ...,
* paddedShape[M] / blockShape[M-1]] + remainingShape`
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function spaceToBatchND_(x, blockShape, paddings) {
const $x = convertToTensor(x, 'x', 'spaceToBatchND');
assert$1($x.rank >= 1 + blockShape.length, () => `input rank ${$x.rank} should be > than [blockShape] ${blockShape.length}`);
assert$1(paddings.length === blockShape.length, () => `paddings.shape[0] ${paddings.length} must be equal to [blockShape] ${blockShape.length}`);
assert$1($x.shape.reduce((a, b, i) => {
if (i > 0 && i <= blockShape.length) {
return a &&
((b + paddings[i - 1][0] + paddings[i - 1][1]) %
blockShape[i - 1] ===
0);
}
return a;
}, true), () => `input spatial dimensions ${$x.shape.slice(1)} with paddings ${paddings.toString()} must be divisible by blockShapes ${blockShape.toString()}`);
const inputs = { x: $x };
const attrs = { blockShape, paddings };
return ENGINE.runKernel(SpaceToBatchND, inputs, attrs);
}
const spaceToBatchND$2 = /* @__PURE__ */ op({ spaceToBatchND_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes leaky rectified linear element-wise with parametric alphas.
*
* `x < 0 ? alpha * x : f(x) = x`
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
* const alpha = tf.scalar(0.1);
*
* x.prelu(alpha).print(); // or tf.prelu(x, alpha)
* ```
* @param x The input tensor.
* @param alpha Scaling factor for negative values.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function prelu_(x, alpha) {
const $x = convertToTensor(x, 'x', 'prelu');
const $alpha = convertToTensor(alpha, 'alpha', 'prelu');
const inputs = { x: $x, alpha: $alpha };
return ENGINE.runKernel(Prelu, inputs);
}
const prelu$2 = /* @__PURE__ */ op({ prelu_ });
var alea$1 = {exports: {}};
(function (module) {
// A port of an algorithm by Johannes Baagøe <baagoe@baagoe.com>, 2010
// http://baagoe.com/en/RandomMusings/javascript/
// https://github.com/nquinlan/better-random-numbers-for-javascript-mirror
// Original work is under MIT license -
// Copyright (C) 2010 by Johannes Baagøe <baagoe@baagoe.org>
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
(function(global, module, define) {
function Alea(seed) {
var me = this, mash = Mash();
me.next = function() {
var t = 2091639 * me.s0 + me.c * 2.3283064365386963e-10; // 2^-32
me.s0 = me.s1;
me.s1 = me.s2;
return me.s2 = t - (me.c = t | 0);
};
// Apply the seeding algorithm from Baagoe.
me.c = 1;
me.s0 = mash(' ');
me.s1 = mash(' ');
me.s2 = mash(' ');
me.s0 -= mash(seed);
if (me.s0 < 0) { me.s0 += 1; }
me.s1 -= mash(seed);
if (me.s1 < 0) { me.s1 += 1; }
me.s2 -= mash(seed);
if (me.s2 < 0) { me.s2 += 1; }
mash = null;
}
function copy(f, t) {
t.c = f.c;
t.s0 = f.s0;
t.s1 = f.s1;
t.s2 = f.s2;
return t;
}
function impl(seed, opts) {
var xg = new Alea(seed),
state = opts && opts.state,
prng = xg.next;
prng.int32 = function() { return (xg.next() * 0x100000000) | 0; };
prng.double = function() {
return prng() + (prng() * 0x200000 | 0) * 1.1102230246251565e-16; // 2^-53
};
prng.quick = prng;
if (state) {
if (typeof(state) == 'object') copy(state, xg);
prng.state = function() { return copy(xg, {}); };
}
return prng;
}
function Mash() {
var n = 0xefc8249d;
var mash = function(data) {
data = String(data);
for (var i = 0; i < data.length; i++) {
n += data.charCodeAt(i);
var h = 0.02519603282416938 * n;
n = h >>> 0;
h -= n;
h *= n;
n = h >>> 0;
h -= n;
n += h * 0x100000000; // 2^32
}
return (n >>> 0) * 2.3283064365386963e-10; // 2^-32
};
return mash;
}
if (module && module.exports) {
module.exports = impl;
} else {
this.alea = impl;
}
})(
commonjsGlobal,
module);
} (alea$1));
var aleaExports = alea$1.exports;
var xor128$1 = {exports: {}};
(function (module) {
// A Javascript implementaion of the "xor128" prng algorithm by
// George Marsaglia. See http://www.jstatsoft.org/v08/i14/paper
(function(global, module, define) {
function XorGen(seed) {
var me = this, strseed = '';
me.x = 0;
me.y = 0;
me.z = 0;
me.w = 0;
// Set up generator function.
me.next = function() {
var t = me.x ^ (me.x << 11);
me.x = me.y;
me.y = me.z;
me.z = me.w;
return me.w ^= (me.w >>> 19) ^ t ^ (t >>> 8);
};
if (seed === (seed | 0)) {
// Integer seed.
me.x = seed;
} else {
// String seed.
strseed += seed;
}
// Mix in string seed, then discard an initial batch of 64 values.
for (var k = 0; k < strseed.length + 64; k++) {
me.x ^= strseed.charCodeAt(k) | 0;
me.next();
}
}
function copy(f, t) {
t.x = f.x;
t.y = f.y;
t.z = f.z;
t.w = f.w;
return t;
}
function impl(seed, opts) {
var xg = new XorGen(seed),
state = opts && opts.state,
prng = function() { return (xg.next() >>> 0) / 0x100000000; };
prng.double = function() {
do {
var top = xg.next() >>> 11,
bot = (xg.next() >>> 0) / 0x100000000,
result = (top + bot) / (1 << 21);
} while (result === 0);
return result;
};
prng.int32 = xg.next;
prng.quick = prng;
if (state) {
if (typeof(state) == 'object') copy(state, xg);
prng.state = function() { return copy(xg, {}); };
}
return prng;
}
if (module && module.exports) {
module.exports = impl;
} else {
this.xor128 = impl;
}
})(
commonjsGlobal,
module);
} (xor128$1));
var xor128Exports = xor128$1.exports;
var xorwow$1 = {exports: {}};
(function (module) {
// A Javascript implementaion of the "xorwow" prng algorithm by
// George Marsaglia. See http://www.jstatsoft.org/v08/i14/paper
(function(global, module, define) {
function XorGen(seed) {
var me = this, strseed = '';
// Set up generator function.
me.next = function() {
var t = (me.x ^ (me.x >>> 2));
me.x = me.y; me.y = me.z; me.z = me.w; me.w = me.v;
return (me.d = (me.d + 362437 | 0)) +
(me.v = (me.v ^ (me.v << 4)) ^ (t ^ (t << 1))) | 0;
};
me.x = 0;
me.y = 0;
me.z = 0;
me.w = 0;
me.v = 0;
if (seed === (seed | 0)) {
// Integer seed.
me.x = seed;
} else {
// String seed.
strseed += seed;
}
// Mix in string seed, then discard an initial batch of 64 values.
for (var k = 0; k < strseed.length + 64; k++) {
me.x ^= strseed.charCodeAt(k) | 0;
if (k == strseed.length) {
me.d = me.x << 10 ^ me.x >>> 4;
}
me.next();
}
}
function copy(f, t) {
t.x = f.x;
t.y = f.y;
t.z = f.z;
t.w = f.w;
t.v = f.v;
t.d = f.d;
return t;
}
function impl(seed, opts) {
var xg = new XorGen(seed),
state = opts && opts.state,
prng = function() { return (xg.next() >>> 0) / 0x100000000; };
prng.double = function() {
do {
var top = xg.next() >>> 11,
bot = (xg.next() >>> 0) / 0x100000000,
result = (top + bot) / (1 << 21);
} while (result === 0);
return result;
};
prng.int32 = xg.next;
prng.quick = prng;
if (state) {
if (typeof(state) == 'object') copy(state, xg);
prng.state = function() { return copy(xg, {}); };
}
return prng;
}
if (module && module.exports) {
module.exports = impl;
} else {
this.xorwow = impl;
}
})(
commonjsGlobal,
module);
} (xorwow$1));
var xorwowExports = xorwow$1.exports;
var xorshift7$1 = {exports: {}};
(function (module) {
// A Javascript implementaion of the "xorshift7" algorithm by
// François Panneton and Pierre L'ecuyer:
// "On the Xorgshift Random Number Generators"
// http://saluc.engr.uconn.edu/refs/crypto/rng/panneton05onthexorshift.pdf
(function(global, module, define) {
function XorGen(seed) {
var me = this;
// Set up generator function.
me.next = function() {
// Update xor generator.
var X = me.x, i = me.i, t, v;
t = X[i]; t ^= (t >>> 7); v = t ^ (t << 24);
t = X[(i + 1) & 7]; v ^= t ^ (t >>> 10);
t = X[(i + 3) & 7]; v ^= t ^ (t >>> 3);
t = X[(i + 4) & 7]; v ^= t ^ (t << 7);
t = X[(i + 7) & 7]; t = t ^ (t << 13); v ^= t ^ (t << 9);
X[i] = v;
me.i = (i + 1) & 7;
return v;
};
function init(me, seed) {
var j, X = [];
if (seed === (seed | 0)) {
// Seed state array using a 32-bit integer.
X[0] = seed;
} else {
// Seed state using a string.
seed = '' + seed;
for (j = 0; j < seed.length; ++j) {
X[j & 7] = (X[j & 7] << 15) ^
(seed.charCodeAt(j) + X[(j + 1) & 7] << 13);
}
}
// Enforce an array length of 8, not all zeroes.
while (X.length < 8) X.push(0);
for (j = 0; j < 8 && X[j] === 0; ++j);
if (j == 8) X[7] = -1;
me.x = X;
me.i = 0;
// Discard an initial 256 values.
for (j = 256; j > 0; --j) {
me.next();
}
}
init(me, seed);
}
function copy(f, t) {
t.x = f.x.slice();
t.i = f.i;
return t;
}
function impl(seed, opts) {
if (seed == null) seed = +(new Date);
var xg = new XorGen(seed),
state = opts && opts.state,
prng = function() { return (xg.next() >>> 0) / 0x100000000; };
prng.double = function() {
do {
var top = xg.next() >>> 11,
bot = (xg.next() >>> 0) / 0x100000000,
result = (top + bot) / (1 << 21);
} while (result === 0);
return result;
};
prng.int32 = xg.next;
prng.quick = prng;
if (state) {
if (state.x) copy(state, xg);
prng.state = function() { return copy(xg, {}); };
}
return prng;
}
if (module && module.exports) {
module.exports = impl;
} else {
this.xorshift7 = impl;
}
})(
commonjsGlobal,
module);
} (xorshift7$1));
var xorshift7Exports = xorshift7$1.exports;
var xor4096$1 = {exports: {}};
(function (module) {
// A Javascript implementaion of Richard Brent's Xorgens xor4096 algorithm.
//
// This fast non-cryptographic random number generator is designed for
// use in Monte-Carlo algorithms. It combines a long-period xorshift
// generator with a Weyl generator, and it passes all common batteries
// of stasticial tests for randomness while consuming only a few nanoseconds
// for each prng generated. For background on the generator, see Brent's
// paper: "Some long-period random number generators using shifts and xors."
// http://arxiv.org/pdf/1004.3115v1.pdf
//
// Usage:
//
// var xor4096 = require('xor4096');
// random = xor4096(1); // Seed with int32 or string.
// assert.equal(random(), 0.1520436450538547); // (0, 1) range, 53 bits.
// assert.equal(random.int32(), 1806534897); // signed int32, 32 bits.
//
// For nonzero numeric keys, this impelementation provides a sequence
// identical to that by Brent's xorgens 3 implementaion in C. This
// implementation also provides for initalizing the generator with
// string seeds, or for saving and restoring the state of the generator.
//
// On Chrome, this prng benchmarks about 2.1 times slower than
// Javascript's built-in Math.random().
(function(global, module, define) {
function XorGen(seed) {
var me = this;
// Set up generator function.
me.next = function() {
var w = me.w,
X = me.X, i = me.i, t, v;
// Update Weyl generator.
me.w = w = (w + 0x61c88647) | 0;
// Update xor generator.
v = X[(i + 34) & 127];
t = X[i = ((i + 1) & 127)];
v ^= v << 13;
t ^= t << 17;
v ^= v >>> 15;
t ^= t >>> 12;
// Update Xor generator array state.
v = X[i] = v ^ t;
me.i = i;
// Result is the combination.
return (v + (w ^ (w >>> 16))) | 0;
};
function init(me, seed) {
var t, v, i, j, w, X = [], limit = 128;
if (seed === (seed | 0)) {
// Numeric seeds initialize v, which is used to generates X.
v = seed;
seed = null;
} else {
// String seeds are mixed into v and X one character at a time.
seed = seed + '\0';
v = 0;
limit = Math.max(limit, seed.length);
}
// Initialize circular array and weyl value.
for (i = 0, j = -32; j < limit; ++j) {
// Put the unicode characters into the array, and shuffle them.
if (seed) v ^= seed.charCodeAt((j + 32) % seed.length);
// After 32 shuffles, take v as the starting w value.
if (j === 0) w = v;
v ^= v << 10;
v ^= v >>> 15;
v ^= v << 4;
v ^= v >>> 13;
if (j >= 0) {
w = (w + 0x61c88647) | 0; // Weyl.
t = (X[j & 127] ^= (v + w)); // Combine xor and weyl to init array.
i = (0 == t) ? i + 1 : 0; // Count zeroes.
}
}
// We have detected all zeroes; make the key nonzero.
if (i >= 128) {
X[(seed && seed.length || 0) & 127] = -1;
}
// Run the generator 512 times to further mix the state before using it.
// Factoring this as a function slows the main generator, so it is just
// unrolled here. The weyl generator is not advanced while warming up.
i = 127;
for (j = 4 * 128; j > 0; --j) {
v = X[(i + 34) & 127];
t = X[i = ((i + 1) & 127)];
v ^= v << 13;
t ^= t << 17;
v ^= v >>> 15;
t ^= t >>> 12;
X[i] = v ^ t;
}
// Storing state as object members is faster than using closure variables.
me.w = w;
me.X = X;
me.i = i;
}
init(me, seed);
}
function copy(f, t) {
t.i = f.i;
t.w = f.w;
t.X = f.X.slice();
return t;
}
function impl(seed, opts) {
if (seed == null) seed = +(new Date);
var xg = new XorGen(seed),
state = opts && opts.state,
prng = function() { return (xg.next() >>> 0) / 0x100000000; };
prng.double = function() {
do {
var top = xg.next() >>> 11,
bot = (xg.next() >>> 0) / 0x100000000,
result = (top + bot) / (1 << 21);
} while (result === 0);
return result;
};
prng.int32 = xg.next;
prng.quick = prng;
if (state) {
if (state.X) copy(state, xg);
prng.state = function() { return copy(xg, {}); };
}
return prng;
}
if (module && module.exports) {
module.exports = impl;
} else {
this.xor4096 = impl;
}
})(
commonjsGlobal, // window object or global
module);
} (xor4096$1));
var xor4096Exports = xor4096$1.exports;
var tychei$1 = {exports: {}};
(function (module) {
// A Javascript implementaion of the "Tyche-i" prng algorithm by
// Samuel Neves and Filipe Araujo.
// See https://eden.dei.uc.pt/~sneves/pubs/2011-snfa2.pdf
(function(global, module, define) {
function XorGen(seed) {
var me = this, strseed = '';
// Set up generator function.
me.next = function() {
var b = me.b, c = me.c, d = me.d, a = me.a;
b = (b << 25) ^ (b >>> 7) ^ c;
c = (c - d) | 0;
d = (d << 24) ^ (d >>> 8) ^ a;
a = (a - b) | 0;
me.b = b = (b << 20) ^ (b >>> 12) ^ c;
me.c = c = (c - d) | 0;
me.d = (d << 16) ^ (c >>> 16) ^ a;
return me.a = (a - b) | 0;
};
/* The following is non-inverted tyche, which has better internal
* bit diffusion, but which is about 25% slower than tyche-i in JS.
me.next = function() {
var a = me.a, b = me.b, c = me.c, d = me.d;
a = (me.a + me.b | 0) >>> 0;
d = me.d ^ a; d = d << 16 ^ d >>> 16;
c = me.c + d | 0;
b = me.b ^ c; b = b << 12 ^ d >>> 20;
me.a = a = a + b | 0;
d = d ^ a; me.d = d = d << 8 ^ d >>> 24;
me.c = c = c + d | 0;
b = b ^ c;
return me.b = (b << 7 ^ b >>> 25);
}
*/
me.a = 0;
me.b = 0;
me.c = 2654435769 | 0;
me.d = 1367130551;
if (seed === Math.floor(seed)) {
// Integer seed.
me.a = (seed / 0x100000000) | 0;
me.b = seed | 0;
} else {
// String seed.
strseed += seed;
}
// Mix in string seed, then discard an initial batch of 64 values.
for (var k = 0; k < strseed.length + 20; k++) {
me.b ^= strseed.charCodeAt(k) | 0;
me.next();
}
}
function copy(f, t) {
t.a = f.a;
t.b = f.b;
t.c = f.c;
t.d = f.d;
return t;
}
function impl(seed, opts) {
var xg = new XorGen(seed),
state = opts && opts.state,
prng = function() { return (xg.next() >>> 0) / 0x100000000; };
prng.double = function() {
do {
var top = xg.next() >>> 11,
bot = (xg.next() >>> 0) / 0x100000000,
result = (top + bot) / (1 << 21);
} while (result === 0);
return result;
};
prng.int32 = xg.next;
prng.quick = prng;
if (state) {
if (typeof(state) == 'object') copy(state, xg);
prng.state = function() { return copy(xg, {}); };
}
return prng;
}
if (module && module.exports) {
module.exports = impl;
} else {
this.tychei = impl;
}
})(
commonjsGlobal,
module);
} (tychei$1));
var tycheiExports = tychei$1.exports;
var seedrandom$1 = {exports: {}};
/*
Copyright 2019 David Bau.
Permission is hereby granted, free of charge, to any person obtaining
a copy of this software and associated documentation files (the
"Software"), to deal in the Software without restriction, including
without limitation the rights to use, copy, modify, merge, publish,
distribute, sublicense, and/or sell copies of the Software, and to
permit persons to whom the Software is furnished to do so, subject to
the following conditions:
The above copyright notice and this permission notice shall be
included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
(function (module) {
(function (global, pool, math) {
//
// The following constants are related to IEEE 754 limits.
//
var width = 256, // each RC4 output is 0 <= x < 256
chunks = 6, // at least six RC4 outputs for each double
digits = 52, // there are 52 significant digits in a double
rngname = 'random', // rngname: name for Math.random and Math.seedrandom
startdenom = math.pow(width, chunks),
significance = math.pow(2, digits),
overflow = significance * 2,
mask = width - 1,
nodecrypto; // node.js crypto module, initialized at the bottom.
//
// seedrandom()
// This is the seedrandom function described above.
//
function seedrandom(seed, options, callback) {
var key = [];
options = (options == true) ? { entropy: true } : (options || {});
// Flatten the seed string or build one from local entropy if needed.
var shortseed = mixkey(flatten(
options.entropy ? [seed, tostring(pool)] :
(seed == null) ? autoseed() : seed, 3), key);
// Use the seed to initialize an ARC4 generator.
var arc4 = new ARC4(key);
// This function returns a random double in [0, 1) that contains
// randomness in every bit of the mantissa of the IEEE 754 value.
var prng = function() {
var n = arc4.g(chunks), // Start with a numerator n < 2 ^ 48
d = startdenom, // and denominator d = 2 ^ 48.
x = 0; // and no 'extra last byte'.
while (n < significance) { // Fill up all significant digits by
n = (n + x) * width; // shifting numerator and
d *= width; // denominator and generating a
x = arc4.g(1); // new least-significant-byte.
}
while (n >= overflow) { // To avoid rounding up, before adding
n /= 2; // last byte, shift everything
d /= 2; // right using integer math until
x >>>= 1; // we have exactly the desired bits.
}
return (n + x) / d; // Form the number within [0, 1).
};
prng.int32 = function() { return arc4.g(4) | 0; };
prng.quick = function() { return arc4.g(4) / 0x100000000; };
prng.double = prng;
// Mix the randomness into accumulated entropy.
mixkey(tostring(arc4.S), pool);
// Calling convention: what to return as a function of prng, seed, is_math.
return (options.pass || callback ||
function(prng, seed, is_math_call, state) {
if (state) {
// Load the arc4 state from the given state if it has an S array.
if (state.S) { copy(state, arc4); }
// Only provide the .state method if requested via options.state.
prng.state = function() { return copy(arc4, {}); };
}
// If called as a method of Math (Math.seedrandom()), mutate
// Math.random because that is how seedrandom.js has worked since v1.0.
if (is_math_call) { math[rngname] = prng; return seed; }
// Otherwise, it is a newer calling convention, so return the
// prng directly.
else return prng;
})(
prng,
shortseed,
'global' in options ? options.global : (this == math),
options.state);
}
//
// ARC4
//
// An ARC4 implementation. The constructor takes a key in the form of
// an array of at most (width) integers that should be 0 <= x < (width).
//
// The g(count) method returns a pseudorandom integer that concatenates
// the next (count) outputs from ARC4. Its return value is a number x
// that is in the range 0 <= x < (width ^ count).
//
function ARC4(key) {
var t, keylen = key.length,
me = this, i = 0, j = me.i = me.j = 0, s = me.S = [];
// The empty key [] is treated as [0].
if (!keylen) { key = [keylen++]; }
// Set up S using the standard key scheduling algorithm.
while (i < width) {
s[i] = i++;
}
for (i = 0; i < width; i++) {
s[i] = s[j = mask & (j + key[i % keylen] + (t = s[i]))];
s[j] = t;
}
// The "g" method returns the next (count) outputs as one number.
(me.g = function(count) {
// Using instance members instead of closure state nearly doubles speed.
var t, r = 0,
i = me.i, j = me.j, s = me.S;
while (count--) {
t = s[i = mask & (i + 1)];
r = r * width + s[mask & ((s[i] = s[j = mask & (j + t)]) + (s[j] = t))];
}
me.i = i; me.j = j;
return r;
// For robust unpredictability, the function call below automatically
// discards an initial batch of values. This is called RC4-drop[256].
// See http://google.com/search?q=rsa+fluhrer+response&btnI
})(width);
}
//
// copy()
// Copies internal state of ARC4 to or from a plain object.
//
function copy(f, t) {
t.i = f.i;
t.j = f.j;
t.S = f.S.slice();
return t;
}
//
// flatten()
// Converts an object tree to nested arrays of strings.
//
function flatten(obj, depth) {
var result = [], typ = (typeof obj), prop;
if (depth && typ == 'object') {
for (prop in obj) {
try { result.push(flatten(obj[prop], depth - 1)); } catch (e) {}
}
}
return (result.length ? result : typ == 'string' ? obj : obj + '\0');
}
//
// mixkey()
// Mixes a string seed into a key that is an array of integers, and
// returns a shortened string seed that is equivalent to the result key.
//
function mixkey(seed, key) {
var stringseed = seed + '', smear, j = 0;
while (j < stringseed.length) {
key[mask & j] =
mask & ((smear ^= key[mask & j] * 19) + stringseed.charCodeAt(j++));
}
return tostring(key);
}
//
// autoseed()
// Returns an object for autoseeding, using window.crypto and Node crypto
// module if available.
//
function autoseed() {
try {
var out;
if (nodecrypto && (out = nodecrypto.randomBytes)) {
// The use of 'out' to remember randomBytes makes tight minified code.
out = out(width);
} else {
out = new Uint8Array(width);
(global.crypto || global.msCrypto).getRandomValues(out);
}
return tostring(out);
} catch (e) {
var browser = global.navigator,
plugins = browser && browser.plugins;
return [+new Date, global, plugins, global.screen, tostring(pool)];
}
}
//
// tostring()
// Converts an array of charcodes to a string
//
function tostring(a) {
return String.fromCharCode.apply(0, a);
}
//
// When seedrandom.js is loaded, we immediately mix a few bits
// from the built-in RNG into the entropy pool. Because we do
// not want to interfere with deterministic PRNG state later,
// seedrandom will not call math.random on its own again after
// initialization.
//
mixkey(math.random(), pool);
//
// Nodejs and AMD support: export the implementation as a module using
// either convention.
//
if (module.exports) {
module.exports = seedrandom;
// When in node.js, try using crypto package for autoseeding.
try {
nodecrypto = require('crypto');
} catch (ex) {}
} else {
// When included as a plain script, set up Math.seedrandom global.
math['seed' + rngname] = seedrandom;
}
// End anonymous scope, and pass initial values.
})(
// global: `self` in browsers (including strict mode and web workers),
// otherwise `this` in Node and other environments
(typeof self !== 'undefined') ? self : commonjsGlobal,
[], // pool: entropy pool starts empty
Math // math: package containing random, pow, and seedrandom
);
} (seedrandom$1));
var seedrandomExports = seedrandom$1.exports;
// A library of seedable RNGs implemented in Javascript.
//
// Usage:
//
// var seedrandom = require('seedrandom');
// var random = seedrandom(1); // or any seed.
// var x = random(); // 0 <= x < 1. Every bit is random.
// var x = random.quick(); // 0 <= x < 1. 32 bits of randomness.
// alea, a 53-bit multiply-with-carry generator by Johannes Baagøe.
// Period: ~2^116
// Reported to pass all BigCrush tests.
var alea = aleaExports;
// xor128, a pure xor-shift generator by George Marsaglia.
// Period: 2^128-1.
// Reported to fail: MatrixRank and LinearComp.
var xor128 = xor128Exports;
// xorwow, George Marsaglia's 160-bit xor-shift combined plus weyl.
// Period: 2^192-2^32
// Reported to fail: CollisionOver, SimpPoker, and LinearComp.
var xorwow = xorwowExports;
// xorshift7, by François Panneton and Pierre L'ecuyer, takes
// a different approach: it adds robustness by allowing more shifts
// than Marsaglia's original three. It is a 7-shift generator
// with 256 bits, that passes BigCrush with no systmatic failures.
// Period 2^256-1.
// No systematic BigCrush failures reported.
var xorshift7 = xorshift7Exports;
// xor4096, by Richard Brent, is a 4096-bit xor-shift with a
// very long period that also adds a Weyl generator. It also passes
// BigCrush with no systematic failures. Its long period may
// be useful if you have many generators and need to avoid
// collisions.
// Period: 2^4128-2^32.
// No systematic BigCrush failures reported.
var xor4096 = xor4096Exports;
// Tyche-i, by Samuel Neves and Filipe Araujo, is a bit-shifting random
// number generator derived from ChaCha, a modern stream cipher.
// https://eden.dei.uc.pt/~sneves/pubs/2011-snfa2.pdf
// Period: ~2^127
// No systematic BigCrush failures reported.
var tychei = tycheiExports;
// The original ARC4-based prng included in this library.
// Period: ~2^1600
var sr = seedrandomExports;
sr.alea = alea;
sr.xor128 = xor128;
sr.xorwow = xorwow;
sr.xorshift7 = xorshift7;
sr.xor4096 = xor4096;
sr.tychei = tychei;
var seedrandom = sr;
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// https://en.wikipedia.org/wiki/Marsaglia_polar_method
class MPRandGauss {
constructor(mean, stdDeviation, dtype, truncated, seed) {
this.mean = mean;
this.stdDev = stdDeviation;
this.dtype = dtype;
this.nextVal = NaN;
this.truncated = truncated;
if (this.truncated) {
this.upper = this.mean + this.stdDev * 2;
this.lower = this.mean - this.stdDev * 2;
}
const seedValue = seed ? seed : Math.random();
this.random = seedrandom.alea(seedValue.toString());
}
/** Returns next sample from a Gaussian distribution. */
nextValue() {
if (!isNaN(this.nextVal)) {
const value = this.nextVal;
this.nextVal = NaN;
return value;
}
let resultX, resultY;
let isValid = false;
while (!isValid) {
let v1, v2, s;
do {
v1 = 2 * this.random() - 1;
v2 = 2 * this.random() - 1;
s = v1 * v1 + v2 * v2;
} while (s >= 1 || s === 0);
const mul = Math.sqrt(-2 * Math.log(s) / s);
resultX = this.mean + this.stdDev * v1 * mul;
resultY = this.mean + this.stdDev * v2 * mul;
if (!this.truncated || this.isValidTruncated(resultX)) {
isValid = true;
}
}
if (!this.truncated || this.isValidTruncated(resultY)) {
this.nextVal = this.convertValue(resultY);
}
return this.convertValue(resultX);
}
/** Handles proper rounding for non-floating-point numbers. */
convertValue(value) {
if (this.dtype == null || this.dtype === 'float32') {
return value;
}
return Math.round(value);
}
/** Returns true if less than 2-standard-deviations from the mean. */
isValidTruncated(value) {
return value <= this.upper && value >= this.lower;
}
}
class UniformRandom {
constructor(min = 0, max = 1, dtype, seed) {
/** Handles proper rounding for non floating point numbers. */
this.canReturnFloat = () => (this.dtype == null || this.dtype === 'float32');
this.min = min;
this.range = max - min;
this.dtype = dtype;
if (seed == null) {
seed = Math.random();
}
if (typeof seed === 'number') {
seed = seed.toString();
}
if (!this.canReturnFloat() && this.range <= 1) {
throw new Error(`The difference between ${min} - ${max} <= 1 and dtype is not float`);
}
this.random = seedrandom.alea(seed);
}
convertValue(value) {
if (this.canReturnFloat()) {
return value;
}
return Math.round(value);
}
nextValue() {
return this.convertValue(this.min + this.range * this.random());
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with values sampled from a normal distribution.
*
* ```js
* tf.randomNormal([2, 2]).print();
* ```
*
* @param shape An array of integers defining the output tensor shape.
* @param mean The mean of the normal distribution.
* @param stdDev The standard deviation of the normal distribution.
* @param dtype The data type of the output.
* @param seed The seed for the random number generator.
*
* @doc {heading: 'Tensors', subheading: 'Random'}
*/
function randomNormal_(shape, mean = 0, stdDev = 1, dtype, seed) {
assertNonNegativeIntegerDimensions(shape);
if (dtype != null && dtype === 'bool') {
throw new Error(`Unsupported data type ${dtype}`);
}
const randGauss = new MPRandGauss(mean, stdDev, dtype, false /* truncated */, seed);
const res = buffer(shape, dtype);
for (let i = 0; i < res.values.length; i++) {
res.values[i] = randGauss.nextValue();
}
return res.toTensor();
}
const randomNormal$1 = /* @__PURE__ */ op({ randomNormal_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with values sampled from a uniform distribution.
*
* The generated values follow a uniform distribution in the range [minval,
* maxval). The lower bound minval is included in the range, while the upper
* bound maxval is excluded.
*
* ```js
* tf.randomUniform([2, 2]).print();
* ```
*
* @param shape An array of integers defining the output tensor shape.
* @param minval The lower bound on the range of random values to generate.
* Defaults to 0.
* @param maxval The upper bound on the range of random values to generate.
* Defaults to 1.
* @param dtype The data type of the output tensor. Defaults to 'float32'.
* @param seed An optional int. Defaults to 0. If seed is set to be non-zero,
* the random number generator is seeded by the given seed. Otherwise, it is
* seeded by a random seed.
*
* @doc {heading: 'Tensors', subheading: 'Random'}
*/
function randomUniform_(shape, minval = 0, maxval = 1, dtype = 'float32', seed) {
assertNonNegativeIntegerDimensions(shape);
const res = buffer(shape, dtype);
const random = new UniformRandom(minval, maxval, null, seed);
for (let i = 0; i < res.values.length; i++) {
res.values[i] = random.nextValue();
}
return res.toTensor();
}
const randomUniform = /* @__PURE__ */ op({ randomUniform_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a new `tf.Tensor1D` filled with the numbers in the range provided.
*
* The tensor is a half-open interval meaning it includes start, but
* excludes stop. Decrementing ranges and negative step values are also
* supported.
*
*
* ```js
* tf.range(0, 9, 2).print();
* ```
*
* @param start An integer start value
* @param stop An integer stop value
* @param step An integer increment (will default to 1 or -1)
* @param dtype The data type of the output tensor. Defaults to 'float32'.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function range$3(start, stop, step = 1, dtype = 'float32') {
if (step === 0) {
throw new Error('Cannot have a step of zero');
}
const attrs = { start, stop, step, dtype };
return ENGINE.runKernel(Range, {} /* inputs */, attrs);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Returns the real part of a complex (or real) tensor.
*
* Given a tensor input, this operation returns a tensor of type float that is
* the real part of each element in input considered as a complex number.
*
* If the input is real, it simply makes a clone.
*
* ```js
* const x = tf.complex([-2.25, 3.25], [4.75, 5.75]);
* tf.real(x).print();
* ```
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function real_(input) {
const $input = convertToTensor(input, 'input', 'real');
const inputs = { input: $input };
return ENGINE.runKernel(Real, inputs);
}
const real$2 = /* @__PURE__ */ op({ real_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes rectified linear element-wise: `max(x, 0)`.
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.relu().print(); // or tf.relu(x)
* ```
* @param x The input tensor. If the dtype is `bool`, the output dtype will be
* `int32`.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function relu_(x) {
const $x = convertToTensor(x, 'x', 'relu');
const inputs = { x: $x };
return ENGINE.runKernel(Relu$1, inputs);
}
const relu$2 = /* @__PURE__ */ op({ relu_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes rectified linear 6 element-wise: `min(max(x, 0), 6)`.
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 8]);
*
* x.relu6().print(); // or tf.relu6(x)
* ```
* @param x The input tensor. If the dtype is `bool`, the output dtype will be
* `int32`.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function relu6_(x) {
const $x = convertToTensor(x, 'x', 'relu6');
const inputs = { x: $x };
return ENGINE.runKernel(Relu6$1, inputs);
}
const relu6$2 = /* @__PURE__ */ op({ relu6_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Reverses a `tf.Tensor` along a specified axis.
*
* Also available are stricter rank-specific methods that assert that `x` is
* of the given rank:
* - `tf.reverse1d`
* - `tf.reverse2d`
* - `tf.reverse3d`
* - `tf.reverse4d`
*
* Except `tf.reverse1d` (which does not have axis param), all methods have
* same signature as this method.
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
*
* x.reverse().print();
* ```
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* const axis = 1;
* x.reverse(axis).print();
* ```
* @param x The input tensor to be reversed.
* @param axis The set of dimensions to reverse. Must be in the
* range [-rank(x), rank(x)). Defaults to all axes.
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function reverse_(x, axis) {
const $x = convertToTensor(x, 'x', 'reverse');
const inputs = { x: $x };
const attrs = { dims: axis };
return ENGINE.runKernel(Reverse, inputs, attrs);
}
const reverse$2 = /* @__PURE__ */ op({ reverse_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes reciprocal of square root of the input `tf.Tensor` element-wise:
* `y = 1 / sqrt(x)`
*
* ```js
* const x = tf.tensor1d([1, 2, 4, -1]);
*
* x.rsqrt().print(); // or tf.rsqrt(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function rsqrt_(x) {
const $x = convertToTensor(x, 'x', 'rsqrt', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Rsqrt, inputs);
}
const rsqrt$2 = /* @__PURE__ */ op({ rsqrt_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes scaled exponential linear element-wise.
*
* `x < 0 ? scale * alpha * (exp(x) - 1) : scale * x`
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.selu().print(); // or tf.selu(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function selu_(x) {
const $x = convertToTensor(x, 'x', 'selu');
const inputs = { x: $x };
return ENGINE.runKernel(Selu$1, inputs);
}
const selu$2 = /* @__PURE__ */ op({ selu_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes sin of the input Tensor element-wise: `sin(x)`
*
* ```js
* const x = tf.tensor1d([0, Math.PI / 2, Math.PI * 3 / 4]);
*
* x.sin().print(); // or tf.sin(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function sin_(x) {
const $x = convertToTensor(x, 'x', 'sin', 'float32');
const inputs = { x: $x };
return ENGINE.runKernel(Sin, inputs);
}
const sin$2 = /* @__PURE__ */ op({ sin_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes hyperbolic sin of the input `tf.Tensor` element-wise: `sinh(x)`
*
* ```js
* const x = tf.tensor1d([0, 1, -1, .7]);
*
* x.sinh().print(); // or tf.sinh(x)
* ```
* @param x The input tensor.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function sinh_(x) {
const $x = convertToTensor(x, 'x', 'sinh');
const inputs = { x: $x };
return ENGINE.runKernel(Sinh, inputs);
}
const sinh$2 = /* @__PURE__ */ op({ sinh_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Extracts a 1D slice from 1D array starting at coordinates `begin` and is
* of length `size`. See `slice` for details.
*/
function slice1d_(x, begin, size) {
const $x = convertToTensor(x, 'x', 'slice1d');
assert$1($x.rank === 1, () => `slice1d expects a rank-1 tensor, but got a rank-${$x.rank} tensor`);
return slice$2($x, [begin], [size]);
}
const slice1d = /* @__PURE__ */ op({ slice1d_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Extracts a 2D slice from a 2D array starting at coordinates `begin` and
* is of size `size`. See `slice` for details.
*/
function slice2d_(x, begin, size) {
const $x = convertToTensor(x, 'x', 'slice2d');
assert$1($x.rank === 2, () => `slice2d expects a rank-2 tensor, but got a rank-${$x.rank} tensor`);
return slice$2($x, begin, size);
}
const slice2d = /* @__PURE__ */ op({ slice2d_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Extracts a 3D slice from a 3D array starting at coordinates `begin` and
* is of size `size`. See `slice` for details.
*/
function slice3d_(x, begin, size) {
const $x = convertToTensor(x, 'x', 'slice3d');
assert$1($x.rank === 3, () => `slice3d expects a rank-3 tensor, but got a rank-${$x.rank} tensor`);
return slice$2($x, begin, size);
}
const slice3d = /* @__PURE__ */ op({ slice3d_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Extracts a 4D slice from a 4D array starting at coordinates `begin` and
* is of size `size`. See `slice` for details.
*/
function slice4d_(x, begin, size) {
const $x = convertToTensor(x, 'x', 'slice4d');
assert$1($x.rank === 4, () => `slice4d expects a rank-4 tensor, but got a rank-${$x.rank} tensor`);
return slice$2($x, begin, size);
}
const slice4d = /* @__PURE__ */ op({ slice4d_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the softmax normalized vector given the logits.
*
* ```js
* const a = tf.tensor1d([1, 2, 3]);
*
* a.softmax().print(); // or tf.softmax(a)
* ```
*
* ```js
* const a = tf.tensor2d([2, 4, 6, 1, 2, 3], [2, 3]);
*
* a.softmax().print(); // or tf.softmax(a)
* ```
*
* @param logits The logits array.
* @param dim The dimension softmax would be performed on. Defaults to `-1`
* which indicates the last dimension.
*
* @doc {heading: 'Operations', subheading: 'Normalization'}
*/
function softmax_(logits, dim = -1) {
const $logits = convertToTensor(logits, 'logits', 'softmax', 'float32');
if (dim === -1) {
dim = $logits.rank - 1;
}
if (dim !== $logits.rank - 1) {
throw Error('Softmax along a non-last dimension is not yet supported. ' +
`Logits was rank ${$logits.rank} and dim was ${dim}`);
}
const inputs = { logits: $logits };
const attrs = { dim };
return ENGINE.runKernel(Softmax$1, inputs, attrs);
}
const softmax$2 = /* @__PURE__ */ op({ softmax_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Splits a `tf.Tensor` into sub tensors.
*
* If `numOrSizeSplits` is a number, splits `x` along dimension `axis`
* into `numOrSizeSplits` smaller tensors.
* Requires that `numOrSizeSplits` evenly divides `x.shape[axis]`.
*
* If `numOrSizeSplits` is a number array, splits `x` into
* `numOrSizeSplits.length` pieces. The shape of the `i`-th piece has the
* same size as `x` except along dimension `axis` where the size is
* `numOrSizeSplits[i]`.
*
* ```js
* const x = tf.tensor2d([1, 2, 3, 4, 5, 6, 7, 8], [2, 4]);
* const [a, b] = tf.split(x, 2, 1);
* a.print();
* b.print();
*
* const [c, d, e] = tf.split(x, [1, 2, 1], 1);
* c.print();
* d.print();
* e.print();
* ```
*
* @param x The input tensor to split.
* @param numOrSizeSplits Either an integer indicating the number of
* splits along the axis or an array of integers containing the sizes of
* each output tensor along the axis. If a number then it must evenly divide
* `x.shape[axis]`; otherwise the sum of sizes must match `x.shape[axis]`.
* Can contain one -1 indicating that dimension is to be inferred.
* @param axis The dimension along which to split. Defaults to 0 (the first
* dim).
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function split_(x, numOrSizeSplits, axis = 0) {
const $x = convertToTensor(x, 'x', 'split');
const inputs = { x: $x };
const attr = { numOrSizeSplits, axis };
return ENGINE.runKernel(SplitV, inputs, attr);
}
const split$1 = /* @__PURE__ */ op({ split_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Removes dimensions of size 1 from the shape of a `tf.Tensor`.
*
* ```js
* const x = tf.tensor([1, 2, 3, 4], [1, 1, 4]);
* x.squeeze().print();
* ```
*
* @param x The input tensor to be squeezed.
* @param axis An optional list of numbers. If specified, only
* squeezes the dimensions listed. The dimension index starts at 0. It
* is an error to squeeze a dimension that is not 1.
*
* @doc {heading: 'Tensors', subheading: 'Transformations'}
*/
function squeeze_(x, axis) {
const $x = convertToTensor(x, 'x', 'squeeze', 'string_or_numeric');
return reshape$2($x, squeezeShape($x.shape, axis).newShape);
}
const squeeze = /* @__PURE__ */ op({ squeeze_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Stacks a list of rank-`R` `tf.Tensor`s into one rank-`(R+1)` `tf.Tensor`.
*
* ```js
* const a = tf.tensor1d([1, 2]);
* const b = tf.tensor1d([3, 4]);
* const c = tf.tensor1d([5, 6]);
* tf.stack([a, b, c]).print();
* ```
*
* @param tensors A list of tensor objects with the same shape and dtype.
* @param axis The axis to stack along. Defaults to 0 (the first dim).
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function stack_(tensors, axis = 0) {
const $tensors = convertToTensorArray(tensors, 'tensors', 'stack', 'string_or_numeric');
assert$1($tensors.length >= 1, () => 'Pass at least one tensor to tf.stack');
if ($tensors.length > 0) {
assert$1(axis <= $tensors[0].rank, () => 'Axis must be <= rank of the tensor');
}
const inputs = $tensors;
const attrs = { axis };
return ENGINE.runKernel(Pack, inputs, attrs);
}
const stack = /* @__PURE__ */ op({ stack_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes step of the input `tf.Tensor` element-wise: `x > 0 ? 1 : alpha`
*
* ```js
* const x = tf.tensor1d([0, 2, -1, -3]);
*
* x.step(.5).print(); // or tf.step(x, .5)
* ```
* @param x The input tensor.
* @param alpha The gradient when input is negative. Defaults to 0.
*
* @doc {heading: 'Operations', subheading: 'Basic math'}
*/
function step_(x, alpha = 0.0) {
const $x = convertToTensor(x, 'x', 'step');
const inputs = { x: $x };
const attrs = { alpha };
return ENGINE.runKernel(Step, inputs, attrs);
}
const step$2 = /* @__PURE__ */ op({ step_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with the provided values, shape and dtype.
*
* ```js
* // Pass an array of values to create a vector.
* tf.tensor([1, 2, 3, 4]).print();
* ```
*
* ```js
* // Pass a nested array of values to make a matrix or a higher
* // dimensional tensor.
* tf.tensor([[1, 2], [3, 4]]).print();
* ```
*
* ```js
* // Pass a flat array and specify a shape yourself.
* tf.tensor([1, 2, 3, 4], [2, 2]).print();
* ```
*
* ```js
* // Pass a `WebGLData` object and specify a shape yourself.
*
* // This makes it possible for TF.js applications to avoid GPU / CPU sync.
* // For example, if your application includes a preprocessing step on the GPU,
* // you could upload the GPU output directly to TF.js, rather than first
* // downloading the values.
*
* // Example for WebGL2:
* if (tf.findBackend('custom-webgl') == null) {
* const customCanvas = document.createElement('canvas');
* const customBackend = new tf.MathBackendWebGL(customCanvas);
* tf.registerBackend('custom-webgl', () => customBackend);
* }
* const savedBackend = tf.getBackend();
* await tf.setBackend('custom-webgl');
* const gl = tf.backend().gpgpu.gl;
* const texture = gl.createTexture();
* const tex2d = gl.TEXTURE_2D;
* const width = 2;
* const height = 2;
*
* gl.bindTexture(tex2d, texture);
* gl.texParameteri(tex2d, gl.TEXTURE_WRAP_S, gl.CLAMP_TO_EDGE);
* gl.texParameteri(tex2d, gl.TEXTURE_WRAP_T, gl.CLAMP_TO_EDGE);
* gl.texParameteri(tex2d, gl.TEXTURE_MIN_FILTER, gl.NEAREST);
* gl.texParameteri(tex2d, gl.TEXTURE_MAG_FILTER, gl.NEAREST);
* gl.texImage2D(
* tex2d, 0, gl.RGBA32F, // internalFormat
* width, height, 0,
* gl.RGBA, // textureFormat
* gl.FLOAT, // textureType
* new Float32Array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15])
* );
*
* // Currently, the `texture` has 4 pixels:
* // Pixel0 is {R:0, G:1, B:2, A:3}
* // Pixel1 is {R:4, G:5, B:6, A:7}
* // Pixel2 is {R:8, G:9, B:10, A:11}
* // Pixel3 is {R:12, G:13, B:14, A:15}
*
* const logicalShape = [height * width * 2];
* const a = tf.tensor({texture, height, width, channels: 'BR'}, logicalShape);
* a.print();
* // Tensor value will be [2, 0, 6, 4, 10, 8, 14, 12], since [2, 0] is the
* // values of 'B' and 'R' channels of Pixel0, [6, 4] is the values of 'B' and
* 'R'
* // channels of Pixel1...
*
* // For postprocessing on the GPU, it's possible to retrieve the texture
* // backing any tensor by calling the tensor's `dataToGPU` method like
* // so:
*
* const tex = a.dataToGPU();
* await tf.setBackend(savedBackend);
* ```
*
* ```js
* // Pass a `WebGPUData` object and specify a shape yourself.
*
* // This makes it possible for TF.js applications to avoid GPU / CPU sync.
* // For example, if your application includes a preprocessing step on the GPU,
* // you could upload the GPU output directly to TF.js, rather than first
* // downloading the values. Unlike WebGL, this optionally supports zero copy
* // by WebGPUData.zeroCopy. When zeroCopy is false or undefined(default), this
* // passing GPUBuffer can be destroyed after tensor is created. When zeroCopy
* // is true, this GPUBuffer is bound directly by the tensor, so do not destroy
* // this GPUBuffer until all access is done.
*
* // Example for WebGPU:
* function createGPUBufferFromData(device, data, dtype) {
* const bytesPerElement = 4;
* const sizeInBytes = data.length * bytesPerElement;
*
* const gpuWriteBuffer = device.createBuffer({
* mappedAtCreation: true,
* size: sizeInBytes,
* usage: GPUBufferUsage.MAP_WRITE | GPUBufferUsage.COPY_SRC
* });
* const arrayBuffer = gpuWriteBuffer.getMappedRange();
* if (dtype === 'float32') {
* new Float32Array(arrayBuffer).set(data);
* } else if (dtype === 'int32') {
* new Int32Array(arrayBuffer).set(data);
* } else {
* throw new Error(
* `Creating tensor from GPUBuffer only supports` +
* `'float32'|'int32' dtype, while the dtype is ${dtype}.`);
* }
* gpuWriteBuffer.unmap();
*
* const gpuReadBuffer = device.createBuffer({
* mappedAtCreation: false,
* size: sizeInBytes,
* usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.STORAGE |
* GPUBufferUsage.COPY_SRC
* });
*
* const copyEncoder = device.createCommandEncoder();
* copyEncoder.copyBufferToBuffer(
* gpuWriteBuffer, 0, gpuReadBuffer, 0, sizeInBytes);
* const copyCommands = copyEncoder.finish();
* device.queue.submit([copyCommands]);
* gpuWriteBuffer.destroy();
* return gpuReadBuffer;
* }
*
* const savedBackend = tf.getBackend();
* await tf.setBackend('webgpu').catch(
* () => {throw new Error(
* 'Failed to use WebGPU backend. Please use Chrome Canary to run.')});
* const dtype = 'float32';
* const device = tf.backend().device;
* const aData = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16];
* const bData = [1, 2, 3, 4, 1, 2, 3, 4, 1, 2, 3, 4, 1, 2, 3, 4];
* const expected = [2, 4, 6, 8, 6, 8, 10, 12, 10, 12, 14, 16, 14, 16, 18, 20];
* const aBuffer = createGPUBufferFromData(device, aData, dtype);
* const shape = [aData.length];
* // To use zeroCopy, use {buffer: aBuffer, zeroCopy: true} instead and destroy
* // aBuffer untill all access is done.
* const a = tf.tensor({buffer: aBuffer}, shape, dtype);
* const b = tf.tensor(bData, shape, dtype);
* const result = tf.add(a, b);
* result.print();
* a.dispose();
* b.dispose();
* result.dispose();
* aBuffer.destroy();
* await tf.setBackend(savedBackend);
* ```
* @param values The values of the tensor. Can be nested array of numbers,
* or a flat array, or a `TypedArray`(At the moment it supports Uint8Array,
* Uint8ClampedArray, Int32Array, Float32Array) data types, or a `WebGLData`
* object, or a `WebGPUData` object. If the values are strings, they will be
* encoded as utf-8 and kept as `Uint8Array[]`. If the values is a `WebGLData`
* object, the dtype could only be 'float32' or 'int32' and the object has to
* have: 1. texture, a `WebGLTexture`, the texture must share the same
* `WebGLRenderingContext` with TFJS's WebGL backend (you could create a custom
* WebGL backend from your texture's canvas) and the internal texture format
* for the input texture must be floating point or normalized integer; 2.
* height, the height of the texture; 3. width, the width of the texture; 4.
* channels, a non-empty subset of 'RGBA', indicating the values of which
* channels will be passed to the tensor, such as 'R' or 'BR' (The order of the
* channels affect the order of tensor values. ). (If the values passed from
* texture is less than the tensor size, zeros will be padded at the rear.). If
* the values is a `WebGPUData` object, the dtype could only be 'float32' or
* 'int32 and the object has to have: buffer, a `GPUBuffer`. The buffer must:
* 1. share the same `GPUDevice` with TFJS's WebGPU backend; 2. buffer.usage
* should at least support GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC; 3.
* buffer.size should not be smaller than the byte size of tensor shape.
* WebGPUData optionally supports zero copy by flag zeroCopy. When zeroCopy is
* false or undefined(default),this passing GPUBuffer can be destroyed after
* tensor is created. When zeroCopy is true, this GPUBuffer is bound directly
* by the tensor, so do not destroy this GPUBuffer until all access is done.
* @param shape The shape of the tensor. Optional. If not provided,
* it is inferred from `values`.
* @param dtype The data type.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function tensor(values, shape, dtype) {
const inferredShape = inferShape(values, dtype);
return makeTensor(values, shape, inferredShape, dtype);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates rank-1 `tf.Tensor` with the provided values, shape and dtype.
*
* The same functionality can be achieved with `tf.tensor`, but in general
* we recommend using `tf.tensor1d` as it makes the code more readable.
*
* ```js
* tf.tensor1d([1, 2, 3]).print();
* ```
*
* @param values The values of the tensor. Can be array of numbers,
* or a `TypedArray`.
* @param dtype The data type.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function tensor1d(values, dtype) {
assertNonNull(values);
const inferredShape = inferShape(values, dtype);
if (inferredShape.length !== 1) {
throw new Error('tensor1d() requires values to be a flat/TypedArray');
}
const shape = null;
return makeTensor(values, shape, inferredShape, dtype);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates rank-2 `tf.Tensor` with the provided values, shape and dtype.
*
* The same functionality can be achieved with `tf.tensor`, but in general
* we recommend using `tf.tensor2d` as it makes the code more readable.
*
* ```js
* // Pass a nested array.
* tf.tensor2d([[1, 2], [3, 4]]).print();
* ```
* ```js
* // Pass a flat array and specify a shape.
* tf.tensor2d([1, 2, 3, 4], [2, 2]).print();
* ```
*
* @param values The values of the tensor. Can be nested array of numbers,
* or a flat array, or a `TypedArray`.
* @param shape The shape of the tensor. If not provided, it is inferred from
* `values`.
* @param dtype The data type.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function tensor2d(values, shape, dtype) {
assertNonNull(values);
if (shape != null && shape.length !== 2) {
throw new Error('tensor2d() requires shape to have two numbers');
}
const inferredShape = inferShape(values, dtype);
if (inferredShape.length !== 2 && inferredShape.length !== 1) {
throw new Error('tensor2d() requires values to be number[][] or flat/TypedArray');
}
if (inferredShape.length === 1 && shape == null) {
throw new Error('tensor2d() requires shape to be provided when `values` ' +
'are a flat/TypedArray');
}
return makeTensor(values, shape, inferredShape, dtype);
}
/**
* Check whether updates.shape = indices.shape[:batchDim] +
* shape[sliceDim:]
*
* @param x The input tensor.
*/
function validateUpdateShape(shape, indices, updates) {
const sliceDim = (indices.rank > 1) ? indices.shape[indices.rank - 1] : 1;
const batchDim = (indices.rank > 1) ? indices.rank - 1 : 1;
const shapeError = 'Must have updates.shape = indices.shape[:batchDim] + ' +
`shape[sliceDim:], got updates.shape: ${updates.shape}` +
`, indices.shape: ${indices.shape}, shape: ${shape}` +
`, sliceDim: ${sliceDim}, and batchDim: ${batchDim}.`;
if (updates.rank < batchDim) {
throw new Error(shapeError + ` update.rank < ${batchDim}. `);
}
if (shape.length < sliceDim + (updates.rank - batchDim)) {
throw new Error(shapeError +
` Output shape length < ${sliceDim + (updates.rank - batchDim)}`);
}
if (updates.rank !== batchDim + shape.length - sliceDim) {
throw new Error(shapeError + ` update.rank != ${batchDim + shape.length - sliceDim}`);
}
for (let d = 0; d < batchDim; ++d) {
if (updates.shape[d] !== indices.shape[d]) {
throw new Error(shapeError +
` updates.shape[${d}] (${updates.shape[d]}) != indices.shape[${d}] (${indices.shape[d]}).`);
}
}
for (let d = 0; d < updates.rank - batchDim; ++d) {
if (updates.shape[d + batchDim] !== shape[d + sliceDim]) {
throw new Error(shapeError +
` updates.shape[${d + batchDim}] (${updates.shape[d + batchDim]}) != shape[${d + batchDim}] (${shape[d + batchDim]})`);
}
}
}
/**
* Validate scatter nd inputs.
*
* @param update The tensor contains the update values.
* @param indices The tensor contains the indices for the update values.
* @param shape The shape of the output tensor.
*/
function validateInput(updates, indices, shape) {
if (indices.rank < 1) {
throw new Error('tf.scatterND() expects the indices to be rank 1 or higher,' +
` but the rank was ${indices.rank}.`);
}
if (updates.rank < 1) {
throw new Error('tf.scatterND() expects the updates to be rank 1 or higher,' +
` but the rank was ${updates.rank}.`);
}
if (indices.dtype !== 'int32') {
throw new Error(`The dtype of 'indices' should be int32, but got dtype: ${indices.dtype}`);
}
if (shape.length < 1) {
throw new Error(`Output rank must be greater or equal to 1, but got shape: ${shape}`);
}
if (shape.length === 0) {
if (indices.size === 0) {
throw new Error(`Indices specified for empty output. indices shape: ${indices.shape}`);
}
if (updates.size === 0) {
throw new Error(`Updates specified for empty output. updates shape: ${updates.shape}`);
}
}
validateUpdateShape(shape, indices, updates);
}
/**
* Calculate the shape information for the output.
*
* @param update The tensor contains the update values.
* @param indices The tensor contains the indices for the update values.
* @param shape The shape of the output tensor.
*
* @returns ScatterShapeInfo
*/
function calculateShapes(updates, indices, shape) {
// Calculate the number of dimensions in indices
const indicesRank = indices.shape.length;
const sliceRank = (indicesRank > 1) ? indices.shape[indicesRank - 1] : 1;
// Calculate the number of elements that make up each slice of our updated
// tensor. This allows us to work with flattened tensors and copy over whole
// slices at a time.
const totalNd = shape.length;
let sliceSize = 1;
for (let i = sliceRank; i < totalNd; ++i) {
sliceSize *= shape[i];
}
const safeSliceDim = (sliceRank < 1) ? 1 : sliceRank;
const numUpdates = sizeFromShape(indices.shape) / safeSliceDim;
const strides = [...computeStrides(shape.slice(0, sliceRank)), 1];
const outputSize = sizeFromShape(shape);
return { sliceRank, numUpdates, sliceSize, strides, outputSize };
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a `tf.Tensor` with values sampled from a truncated normal
* distribution.
*
* ```js
* tf.truncatedNormal([2, 2]).print();
* ```
*
* The generated values follow a normal distribution with specified mean and
* standard deviation, except that values whose magnitude is more than 2
* standard deviations from the mean are dropped and re-picked.
*
* @param shape An array of integers defining the output tensor shape.
* @param mean The mean of the normal distribution.
* @param stdDev The standard deviation of the normal distribution.
* @param dtype The data type of the output tensor.
* @param seed The seed for the random number generator.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function truncatedNormal_(shape, mean = 0, stdDev = 1, dtype, seed) {
assertNonNegativeIntegerDimensions(shape);
if (dtype != null && dtype === 'bool') {
throw new Error(`Unsupported data type $ { dtype }`);
}
const randGauss = new MPRandGauss(mean, stdDev, dtype, true /* truncated */, seed);
const res = buffer(shape, dtype);
for (let i = 0; i < res.values.length; i++) {
res.values[i] = randGauss.nextValue();
}
return res.toTensor();
}
const truncatedNormal = /* @__PURE__ */ op({ truncatedNormal_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the sum along segments of a `tf.Tensor`.
*
* ```js
* const x = tf.tensor1d([1, 2, 3, 4]);
* const segmentIds = tf.tensor1d([1, 2, 0, 1], 'int32');
* const numSegments = 3;
*
* x.unsortedSegmentSum(segmentIds, numSegments).print()
* //or tf.unsortedSegmentSum(x, segmentIds, numSegments)
* ```
* @param x The `tf.Tensor` that will be summed along its segments.
* @param segmentIds A `tf.Tensor1D` whose rank is equal to the rank of `x`'s
* dimension along the `axis`. Maps each element of `x` to a segment.
* @param numSegments The number of distinct `segmentIds`.
*
* @doc {heading: 'Operations', subheading: 'Segment'}
*/
function unsortedSegmentSum_(x, segmentIds, numSegments) {
const $x = convertToTensor(x, 'x', 'unsortedSegmentSum');
const $segmentIds = convertToTensor(segmentIds, 'segmentIds', 'unsortedSegmentSum', 'int32');
assert$1(isInt(numSegments), () => 'numSegments must be of dtype int');
const inputs = { x: $x, segmentIds: $segmentIds };
const attrs = { numSegments };
return ENGINE.runKernel(UnsortedSegmentSum, inputs, attrs);
}
const unsortedSegmentSum$2 = /* @__PURE__ */ op({ unsortedSegmentSum_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Unstacks a `tf.Tensor` of rank-`R` into a list of rank-`(R-1)` `tf.Tensor`s.
*
* ```js
* const a = tf.tensor2d([1, 2, 3, 4], [2, 2]);
*
* tf.unstack(a).forEach(tensor => tensor.print());
* ```
*
* @param x A tensor object.
* @param axis The axis to unstack along. Defaults to 0 (the first dim).
*
* @doc {heading: 'Tensors', subheading: 'Slicing and Joining'}
*/
function unstack_(x, axis = 0) {
const $x = convertToTensor(x, 'x', 'unstack', 'string_or_numeric');
assert$1(axis >= -$x.shape.length && axis < $x.shape.length, () => `Axis = ${axis} is not in [-${$x.shape.length}, ${$x.shape.length})`);
const inputs = { value: $x };
const attrs = { axis };
return ENGINE.runKernel(Unpack, inputs, attrs);
}
const unstack = /* @__PURE__ */ op({ unstack_ });
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Creates a new variable with the provided initial value.
* ```js
* const x = tf.variable(tf.tensor([1, 2, 3]));
* x.assign(tf.tensor([4, 5, 6]));
*
* x.print();
* ```
*
* @param initialValue Initial value for the tensor.
* @param trainable If true, optimizers are allowed to update it.
* @param name Name of the variable. Defaults to a unique id.
* @param dtype If set, initialValue will be converted to the given type.
*
* @doc {heading: 'Tensors', subheading: 'Creation'}
*/
function variable(initialValue, trainable = true, name, dtype) {
return ENGINE.makeVariable(initialValue, trainable, name, dtype);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** An implementation of the Where kernel shared between cpu and webgl */
function whereImpl$2(condShape, condVals) {
const indices = [];
for (let i = 0; i < condVals.length; i++) {
if (condVals[i]) {
indices.push(i);
}
}
const inBuffer = buffer(condShape, 'int32');
const out = buffer([indices.length, condShape.length], 'int32');
for (let i = 0; i < indices.length; i++) {
const loc = inBuffer.indexToLoc(indices[i]);
const offset = i * condShape.length;
out.values.set(loc, offset);
}
return out.toTensor();
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Transposes the `tf.Tensor`. Permutes the dimensions according to `perm`.
*
* The returned `tf.Tensor`'s dimension `i` will correspond to the input
* dimension `perm[i]`. If `perm` is not given, it is set to `[n-1...0]`,
* where `n` is the rank of the input `tf.Tensor`. Hence by default, this
* operation performs a regular matrix transpose on 2-D input `tf.Tensor`s.
*
* ```js
* const a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
*
* a.transpose().print(); // or tf.transpose(a)
* ```
*
* @param x The tensor to transpose.
* @param perm The permutation of the dimensions of a.
* @param conjugate Will conjugate complex input if true.
*
* @doc {heading: 'Operations', subheading: 'Matrices'}
*/
function transpose_(x, perm, conjugate) {
const $x = convertToTensor(x, 'x', 'transpose');
if (perm == null) {
perm = $x.shape.map((s, i) => i).reverse();
}
assert$1($x.rank === perm.length, () => `Error in transpose: rank of input ${$x.rank} ` +
`must match length of perm ${perm}.`);
perm.forEach(axis => {
assert$1(axis >= 0 && axis < $x.rank, () => `All entries in 'perm' must be between 0 and ${$x.rank - 1}` +
` but got ${perm}`);
});
if ($x.rank <= 1) {
return $x.clone();
}
const inputs = { x: $x };
const attrs = { perm };
if ($x.dtype === 'complex64') {
return tidy(() => {
let $real = real$2($x);
let $imag = imag$2($x);
$real = ENGINE.runKernel(Transpose, { x: $real }, attrs);
$imag = ENGINE.runKernel(Transpose, { x: $imag }, attrs);
if (conjugate) {
$imag = neg$2($imag);
}
return complex$2($real, $imag);
});
}
return ENGINE.runKernel(Transpose, inputs, attrs);
}
const transpose$2 = /* @__PURE__ */ op({ transpose_ });
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Normalize noise shape based on provided tensor and noise shape.
*
* @param x Tensor.
* @param noiseShape The shape for the randomly generated keep/drop flags, as
* an array of numbers. Optional.
* @returns Normalized noise shape.
*/
function getNoiseShape(x, noiseShape) {
if (noiseShape == null) {
return x.shape.slice();
}
if (arraysEqual(x.shape, noiseShape)) {
return noiseShape;
}
if (x.shape.length === noiseShape.length) {
const newDimension = [];
for (let i = 0; i < x.shape.length; i++) {
if (noiseShape[i] == null && x.shape[i] != null) {
newDimension.push(x.shape[i]);
}
else {
newDimension.push(noiseShape[i]);
}
}
return newDimension;
}
return noiseShape;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes dropout.
*
* ```js
* const x = tf.tensor1d([1, 2, 2, 1]);
* const rate = 0.75;
* const output = tf.dropout(x, rate);
* output.print();
* ```
*
* @param x A floating point Tensor or TensorLike.
* @param rate A float in the range [0, 1). The probability that each element
* of x is discarded.
* @param noiseShape An array of numbers of type int32, representing the
* shape for randomly generated keep/drop flags. If the noiseShape has null
* value, it will be automatically replaced with the x's relative dimension
* size. Optional.
* @param seed Used to create random seeds. Optional.
* @returns A Tensor of the same shape of x.
*
* @doc {heading: 'Operations', subheading: 'Dropout'}
*/
function dropout_(x, rate, noiseShape, seed) {
const $x = convertToTensor(x, 'x', 'dropout');
assert$1($x.dtype === 'float32', () => `x has to be a floating point tensor since it's going to be ` +
`scaled, but got a ${$x.dtype} tensor instead.`);
assert$1(rate >= 0 && rate < 1, () => `rate must be a float in the range [0, 1), but got ${rate}.`);
if (rate === 0) {
return x instanceof Tensor ? $x.clone() : $x;
}
const $noiseShape = getNoiseShape($x, noiseShape);
const keepProb = 1 - rate;
const multiplier = div$1(floor$2(add$1(randomUniform($noiseShape, 0, 1, 'float32', seed), keepProb)), keepProb);
return mul($x, multiplier);
}
const dropout$2 = /* @__PURE__ */ op({ dropout_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the derivative of the filter of a 2D convolution.
*
* @param x The input tensor, of rank 4 or rank 3 of shape
* [batch, height, width, inChannels]. If rank 3, batch of 1 is assumed.
* @param dy The dy image, of rank 4 or rank 3, of shape
* [batch, height, width, outDepth]. If rank 3, batch of 1 is assumed.
* @param filterShape The shape of the filter, length 4,
* [filterHeight, filterWidth, inDepth, outDepth].
* @param strides The strides of the convolution: [strideHeight,
* strideWidth].
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
* used in the forward prop of the op.
* @param dataFormat: An optional string from: "NHWC", "NCHW". Defaults to
* "NHWC". Specify the data format of the input and output data. With the
* default format "NHWC", the data is stored in the order of: [batch,
* height, width, channels].
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*/
function conv2DBackpropFilter_(x, dy, filterShape, strides, pad, dataFormat = 'NHWC', dimRoundingMode) {
let x4D = x;
if (x.rank === 3) {
x4D = reshape$2(x, [1, x.shape[0], x.shape[1], x.shape[2]]);
}
let dy4D = dy;
if (dy4D.rank === 3) {
dy4D = reshape$2(dy, [1, dy.shape[0], dy.shape[1], dy.shape[2]]);
}
assert$1(x4D.rank === 4, () => `Error in conv2dDerFilter: input must be rank 4, but got shape ` +
`${x4D.shape}.`);
assert$1(dy4D.rank === 4, () => `Error in conv2dDerFilter: dy must be rank 4, but got shape ` +
`${dy4D.shape}.`);
assert$1(filterShape.length === 4, () => `Error in conv2dDerFilter: filterShape must be length 4, but got ` +
`${filterShape}.`);
const inDepth = dataFormat === 'NHWC' ? x4D.shape[3] : x4D.shape[1];
const outDepth = dataFormat === 'NHWC' ? dy4D.shape[3] : dy4D.shape[1];
assert$1(inDepth === filterShape[2], () => `Error in conv2dDerFilter: depth of input ${inDepth}) must ` +
`match input depth in filter (${filterShape[2]}.`);
assert$1(outDepth === filterShape[3], () => `Error in conv2dDerFilter: depth of dy (${outDepth}) must ` +
`match output depth for filter (${filterShape[3]}).`);
checkPadOnDimRoundingMode('conv2dDerFilter', pad, dimRoundingMode);
const inputs = { x: x4D, dy: dy4D };
const attrs = { strides, pad, dataFormat, dimRoundingMode, filterShape };
// tslint:disable-next-line: no-unnecessary-type-assertion
return ENGINE.runKernel(Conv2DBackpropFilter, inputs, attrs);
}
const conv2DBackpropFilter$2 = /* @__PURE__ */ op({ conv2DBackpropFilter_ });
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Returns gradient for fused activation.
function getFusedDyActivation(dy, y, activation) {
if (activation == null || activation === 'linear') {
return dy;
}
if (activation === 'relu') {
return mul(dy, step$2(y));
}
throw new Error(`Cannot compute gradient for fused activation ${activation}.`);
}
// Returns gradient for fused bias.
function getFusedBiasGradient(bias, dyActivation) {
let res = dyActivation;
const reduceAxes = getReductionAxes(bias.shape, dyActivation.shape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, bias.shape);
}
function applyActivation$1(x, activation, preluActivationWeights, leakyreluAlpha) {
if (activation === 'linear') {
return x;
}
else if (activation === 'relu') {
return relu$2(x);
}
else if (activation === 'elu') {
return elu$3(x);
}
else if (activation === 'relu6') {
return relu6$2(x);
}
else if (activation === 'prelu') {
return prelu$2(x, preluActivationWeights);
}
else if (activation === 'leakyrelu') {
return leakyRelu$2(x, leakyreluAlpha);
}
else if (activation === 'sigmoid') {
return sigmoid$2(x);
}
throw new Error(`Unknown fused activation ${activation}.`);
}
// Whether we should call fused ops.
const shouldFuse = (gradientDepth, activation) => {
const gradientMode = gradientDepth > 0;
return !gradientMode || activation === 'linear';
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNativeBackpropFilter_(x, dy, filterShape, strides, pad, dilations = [1, 1], dimRoundingMode) {
let x4D = x;
if (x.rank === 3) {
x4D = reshape$2(x, [1, x.shape[0], x.shape[1], x.shape[2]]);
}
let dy4D = dy;
if (dy4D.rank === 3) {
dy4D = reshape$2(dy, [1, dy.shape[0], dy.shape[1], dy.shape[2]]);
}
const inputs = { x: x4D, dy: dy4D };
const attrs = { strides, pad, dimRoundingMode, dilations, filterShape };
// tslint:disable-next-line: no-unnecessary-type-assertion
return ENGINE.runKernel(DepthwiseConv2dNativeBackpropFilter, inputs, attrs);
}
const depthwiseConv2dNativeBackpropFilter$2 = op({ depthwiseConv2dNativeBackpropFilter_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNativeBackpropInput_(xShape, dy, filter, strides, pad, dilations = [1, 1], dimRoundingMode) {
let dy4D = dy;
let reshapedTo4D = false;
if (dy.rank === 3) {
reshapedTo4D = true;
dy4D = reshape$2(dy, [1, dy.shape[0], dy.shape[1], dy.shape[2]]);
}
const inputs = { dy: dy4D, filter };
const attrs = { strides, pad, dimRoundingMode, dilations, inputShape: xShape };
const res =
// tslint:disable-next-line: no-unnecessary-type-assertion
ENGINE.runKernel(DepthwiseConv2dNativeBackpropInput, inputs, attrs);
if (reshapedTo4D) {
return reshape$2(res, [res.shape[1], res.shape[2], res.shape[3]]);
}
return res;
}
const depthwiseConv2dNativeBackpropInput$2 = op({ depthwiseConv2dNativeBackpropInput_ });
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the dot product of two matrices with optional activation and bias.
*
* ```js
* const a = tf.tensor2d([-1, -2], [1, 2]);
* const b = tf.tensor2d([1, 2, 3, 4], [2, 2]);
* const bias = tf.tensor2d([1, 2], [1, 2]);
*
* tf.fused.matMul({a, b, bias, activation: 'relu'}).print();
* ```
*
* @param obj An object with the following properties:
* - `a` First matrix in dot product operation.
* - `b` Second matrix in dot product operation.
* - `transposeA` If true, `a` is transposed before multiplication.
* - `transposeB` If true, `b` is transposed before multiplication.
* - `bias` Matrix to be added to the result.
* - `activation` Name of activation kernel (defaults to `linear`).
* - `preluActivationWeights` Tensor of prelu weights.
* - `leakyreluAlpha` Alpha of leakyrelu.
*/
function fusedMatMul_({ a, b, transposeA = false, transposeB = false, bias, activation = 'linear', preluActivationWeights, leakyreluAlpha = 0.2, }) {
if (shouldFuse(ENGINE.state.gradientDepth, activation) === false) {
let result = matMul$1(a, b, transposeA, transposeB);
if (bias != null) {
result = add$1(result, bias);
}
return applyActivation$1(result, activation, preluActivationWeights, leakyreluAlpha);
}
let $a = convertToTensor(a, 'a', 'fused matMul');
let $b = convertToTensor(b, 'b', 'fused matMul');
[$a, $b] = makeTypesMatch($a, $b);
const innerShapeA = transposeA ? $a.shape[$a.rank - 2] : $a.shape[$a.rank - 1];
const innerShapeB = transposeB ? $b.shape[$b.rank - 1] : $b.shape[$b.rank - 2];
const outerShapeA = transposeA ? $a.shape[$a.rank - 1] : $a.shape[$a.rank - 2];
const outerShapeB = transposeB ? $b.shape[$b.rank - 2] : $b.shape[$b.rank - 1];
const outerDimsA = $a.shape.slice(0, -2);
const outerDimsB = $b.shape.slice(0, -2);
const batchDimA = sizeFromShape(outerDimsA);
const batchDimB = sizeFromShape(outerDimsB);
assert$1(innerShapeA === innerShapeB, () => `Error in fused matMul: inner shapes (${innerShapeA}) and (` +
`${innerShapeB}) of Tensors with shapes ${$a.shape} and ` +
`${$b.shape} and transposeA=${transposeA}` +
` and transposeB=${transposeB} must match.`);
const outShapeOuterDims = assertAndGetBroadcastShape($a.shape.slice(0, -2), $b.shape.slice(0, -2));
const outShape = outShapeOuterDims.concat([outerShapeA, outerShapeB]);
const a3D = transposeA ?
reshape$2($a, [batchDimA, innerShapeA, outerShapeA]) :
reshape$2($a, [batchDimA, outerShapeA, innerShapeA]);
const b3D = transposeB ?
reshape$2($b, [batchDimB, outerShapeB, innerShapeB]) :
reshape$2($b, [batchDimB, innerShapeB, outerShapeB]);
let $bias;
if (bias != null) {
$bias = convertToTensor(bias, 'bias', 'fused matMul');
[$bias] = makeTypesMatch($bias, $a);
assertAndGetBroadcastShape(outShape, $bias.shape);
}
let $preluActivationWeights;
if (preluActivationWeights != null) {
$preluActivationWeights = convertToTensor(preluActivationWeights, 'prelu weights', 'fused matMul');
}
const grad = (dy, saved) => {
const [a3D, b3D, y, $bias] = saved;
// we reshape dy because the result of the forward is not
// necessarily going to be a 3d tensor due to a reshape done at the end of
// the customOp.
const dyActivation = getFusedDyActivation(reshape$2(dy, y.shape), y, activation);
let aDer;
let bDer;
if (!transposeA && !transposeB) {
aDer = matMul$1(dyActivation, b3D, false, true);
bDer = matMul$1(a3D, dyActivation, true, false);
}
else if (!transposeA && transposeB) {
aDer = matMul$1(dyActivation, b3D, false, false);
bDer = matMul$1(dyActivation, a3D, true, false);
}
else if (transposeA && !transposeB) {
aDer = matMul$1(b3D, dyActivation, false, true);
bDer = matMul$1(a3D, dyActivation, false, false);
}
else {
aDer = matMul$1(b3D, dyActivation, true, true);
bDer = matMul$1(dyActivation, a3D, true, true);
}
if (bias != null) {
const biasDer = getFusedBiasGradient($bias, dyActivation);
return [aDer, bDer, biasDer];
}
else {
return [aDer, bDer];
}
};
const inputs = {
a: a3D,
b: b3D,
bias: $bias,
preluActivationWeights: $preluActivationWeights
};
const attrs = { transposeA, transposeB, activation, leakyreluAlpha };
// Depending on the the params passed in we will have different number of
// inputs and thus a a different number of elements in the gradient.
if (bias == null) {
const customOp = customGrad((a3D, b3D, save) => {
const res =
// tslint:disable-next-line: no-unnecessary-type-assertion
ENGINE.runKernel(_FusedMatMul, inputs, attrs);
save([a3D, b3D, res]);
return { value: reshape$2(res, outShape), gradFunc: grad };
});
return customOp(a3D, b3D);
}
else {
const customOpWithBias = customGrad((a3D, b3D, $bias, save) => {
const res =
// tslint:disable-next-line: no-unnecessary-type-assertion
ENGINE.runKernel(_FusedMatMul, inputs, attrs);
save([a3D, b3D, res, $bias]);
return { value: reshape$2(res, outShape), gradFunc: grad };
});
return customOpWithBias(a3D, b3D, $bias);
}
}
const matMul = /* @__PURE__ */ op({ fusedMatMul_ });
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Inserts a value into a sorted array. This method allows duplicate, meaning it
* allows inserting duplicate value, in which case, the element will be inserted
* at the lowest index of the value.
* @param arr The array to modify.
* @param element The element to insert.
* @param comparator Optional. If no comparator is specified, elements are
* compared using array_util.defaultComparator, which is suitable for Strings
* and Numbers in ascending arrays. If the array contains multiple instances of
* the target value, the left-most instance will be returned. To provide a
* comparator, it should take 2 arguments to compare and return a negative,
* zero, or a positive number.
*/
function binaryInsert(arr, element, comparator) {
const index = binarySearch(arr, element, comparator);
const insertionPoint = index < 0 ? -(index + 1) : index;
arr.splice(insertionPoint, 0, element);
}
/**
* Searches the array for the target using binary search, returns the index
* of the found element, or position to insert if element not found. If no
* comparator is specified, elements are compared using array_
* util.defaultComparator, which is suitable for Strings and Numbers in
* ascending arrays. If the array contains multiple instances of the target
* value, the left-most instance will be returned.
* @param arr The array to be searched in.
* @param target The target to be searched for.
* @param comparator Should take 2 arguments to compare and return a negative,
* zero, or a positive number.
* @return Lowest index of the target value if found, otherwise the insertion
* point where the target should be inserted, in the form of
* (-insertionPoint - 1).
*/
function binarySearch(arr, target, comparator) {
return binarySearch_(arr, target, comparator || defaultComparator);
}
/**
* Compares its two arguments for order.
* @param a The first element to be compared.
* @param b The second element to be compared.
* @return A negative number, zero, or a positive number as the first
* argument is less than, equal to, or greater than the second.
*/
function defaultComparator(a, b) {
return a > b ? 1 : a < b ? -1 : 0;
}
function binarySearch_(arr, target, comparator) {
let left = 0;
let right = arr.length;
let middle = 0;
let found = false;
while (left < right) {
middle = left + ((right - left) >>> 1);
const compareResult = comparator(target, arr[middle]);
if (compareResult > 0) {
left = middle + 1;
}
else {
right = middle;
// If compareResult is 0, the value is found. We record it is found,
// and then keep looking because there may be duplicate.
found = !compareResult;
}
}
return found ? left : -left - 1;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function nonMaxSuppressionV3Impl$2(boxes, scores, maxOutputSize, iouThreshold, scoreThreshold) {
return nonMaxSuppressionImpl_(boxes, scores, maxOutputSize, iouThreshold, scoreThreshold, 0 /* softNmsSigma */);
}
function nonMaxSuppressionV4Impl$2(boxes, scores, maxOutputSize, iouThreshold, scoreThreshold, padToMaxOutputSize) {
return nonMaxSuppressionImpl_(boxes, scores, maxOutputSize, iouThreshold, scoreThreshold, 0 /* softNmsSigma */, false /* returnScoresTensor */, padToMaxOutputSize /* padToMaxOutputSize */, true
/* returnValidOutputs */ );
}
function nonMaxSuppressionV5Impl$2(boxes, scores, maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma) {
return nonMaxSuppressionImpl_(boxes, scores, maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma, true /* returnScoresTensor */);
}
function nonMaxSuppressionImpl_(boxes, scores, maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma, returnScoresTensor = false, padToMaxOutputSize = false, returnValidOutputs = false) {
// The list is sorted in ascending order, so that we can always pop the
// candidate with the largest score in O(1) time.
const candidates = [];
for (let i = 0; i < scores.length; i++) {
if (scores[i] > scoreThreshold) {
candidates.push({ score: scores[i], boxIndex: i, suppressBeginIndex: 0 });
}
}
candidates.sort(ascendingComparator);
// If softNmsSigma is 0, the outcome of this algorithm is exactly same as
// before.
const scale = softNmsSigma > 0 ? (-0.5 / softNmsSigma) : 0.0;
const selectedIndices = [];
const selectedScores = [];
while (selectedIndices.length < maxOutputSize && candidates.length > 0) {
const candidate = candidates.pop();
const { score: originalScore, boxIndex, suppressBeginIndex } = candidate;
if (originalScore < scoreThreshold) {
break;
}
// Overlapping boxes are likely to have similar scores, therefore we
// iterate through the previously selected boxes backwards in order to
// see if candidate's score should be suppressed. We use
// suppressBeginIndex to track and ensure a candidate can be suppressed
// by a selected box no more than once. Also, if the overlap exceeds
// iouThreshold, we simply ignore the candidate.
let ignoreCandidate = false;
for (let j = selectedIndices.length - 1; j >= suppressBeginIndex; --j) {
const iou = intersectionOverUnion(boxes, boxIndex, selectedIndices[j]);
if (iou >= iouThreshold) {
ignoreCandidate = true;
break;
}
candidate.score =
candidate.score * suppressWeight(iouThreshold, scale, iou);
if (candidate.score <= scoreThreshold) {
break;
}
}
// At this point, if `candidate.score` has not dropped below
// `scoreThreshold`, then we know that we went through all of the
// previous selections and can safely update `suppressBeginIndex` to the
// end of the selected array. Then we can re-insert the candidate with
// the updated score and suppressBeginIndex back in the candidate list.
// If on the other hand, `candidate.score` has dropped below the score
// threshold, we will not add it back to the candidates list.
candidate.suppressBeginIndex = selectedIndices.length;
if (!ignoreCandidate) {
// Candidate has passed all the tests, and is not suppressed, so
// select the candidate.
if (candidate.score === originalScore) {
selectedIndices.push(boxIndex);
selectedScores.push(candidate.score);
}
else if (candidate.score > scoreThreshold) {
// Candidate's score is suppressed but is still high enough to be
// considered, so add back to the candidates list.
binaryInsert(candidates, candidate, ascendingComparator);
}
}
}
// NonMaxSuppressionV4 feature: padding output to maxOutputSize.
const validOutputs = selectedIndices.length;
const elemsToPad = maxOutputSize - validOutputs;
if (padToMaxOutputSize && elemsToPad > 0) {
selectedIndices.push(...new Array(elemsToPad).fill(0));
selectedScores.push(...new Array(elemsToPad).fill(0.0));
}
const result = { selectedIndices };
if (returnScoresTensor) {
result['selectedScores'] = selectedScores;
}
if (returnValidOutputs) {
result['validOutputs'] = validOutputs;
}
return result;
}
function intersectionOverUnion(boxes, i, j) {
const iCoord = boxes.subarray(i * 4, i * 4 + 4);
const jCoord = boxes.subarray(j * 4, j * 4 + 4);
const yminI = Math.min(iCoord[0], iCoord[2]);
const xminI = Math.min(iCoord[1], iCoord[3]);
const ymaxI = Math.max(iCoord[0], iCoord[2]);
const xmaxI = Math.max(iCoord[1], iCoord[3]);
const yminJ = Math.min(jCoord[0], jCoord[2]);
const xminJ = Math.min(jCoord[1], jCoord[3]);
const ymaxJ = Math.max(jCoord[0], jCoord[2]);
const xmaxJ = Math.max(jCoord[1], jCoord[3]);
const areaI = (ymaxI - yminI) * (xmaxI - xminI);
const areaJ = (ymaxJ - yminJ) * (xmaxJ - xminJ);
if (areaI <= 0 || areaJ <= 0) {
return 0.0;
}
const intersectionYmin = Math.max(yminI, yminJ);
const intersectionXmin = Math.max(xminI, xminJ);
const intersectionYmax = Math.min(ymaxI, ymaxJ);
const intersectionXmax = Math.min(xmaxI, xmaxJ);
const intersectionArea = Math.max(intersectionYmax - intersectionYmin, 0.0) *
Math.max(intersectionXmax - intersectionXmin, 0.0);
return intersectionArea / (areaI + areaJ - intersectionArea);
}
// A Gaussian penalty function, this method always returns values in [0, 1].
// The weight is a function of similarity, the more overlap two boxes are, the
// smaller the weight is,meaning highly overlapping boxes will be significantly
// penalized. On the other hand, a non-overlapping box will not be penalized.
function suppressWeight(iouThreshold, scale, iou) {
const weight = Math.exp(scale * iou * iou);
return iou <= iouThreshold ? weight : 0.0;
}
function ascendingComparator(c1, c2) {
// For objects with same scores, we make the object with the larger index go
// first. In an array that pops from the end, this means that the object with
// the smaller index will be popped first. This ensures the same output as
// the TensorFlow python version.
return (c1.score - c2.score) ||
((c1.score === c2.score) && (c2.boxIndex - c1.boxIndex));
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Copy a tensor setting everything outside a central band in each innermost
* matrix to zero.
*
* The band part is computed as follows: Assume input has `k` dimensions
* `[I, J, K, ..., M, N]`, then the output is a tensor with the same shape where
* `band[i, j, k, ..., m, n] = in_band(m, n) * input[i, j, k, ..., m, n]`.
* The indicator function
* `in_band(m, n) = (num_lower < 0 || (m-n) <= num_lower)`
* `&& (num_upper < 0 || (n-m) <= num_upper)`
*
* ```js
* const x = tf.tensor2d([[ 0, 1, 2, 3],
* [-1, 0, 1, 2],
* [-2, -1, 0, 1],
* [-3, -2, -1, 0]]);
* let y = tf.linalg.bandPart(x, 1, -1);
* y.print(); // [[ 0, 1, 2, 3],
* // [-1, 0, 1, 2],
* // [ 0, -1, 0, 1],
* // [ 0, 0 , -1, 0]]
* let z = tf.linalg.bandPart(x, 2, 1);
* z.print(); // [[ 0, 1, 0, 0],
* // [-1, 0, 1, 0],
* // [-2, -1, 0, 1],
* // [ 0, -2, -1, 0]]
* ```
*
* @param x Rank `k` tensor
* @param numLower Number of subdiagonals to keep.
* If negative, keep entire lower triangle.
* @param numUpper Number of subdiagonals to keep.
* If negative, keep entire upper triangle.
* @returns Rank `k` tensor of the same shape as input.
* The extracted banded tensor.
*
* @doc {heading:'Operations', subheading:'Linear Algebra', namespace:'linalg'}
*/
function bandPart_(a, numLower, numUpper) {
const $a = convertToTensor(a, 'a', 'bandPart');
assert$1($a.rank >= 2, () => `bandPart(): Rank must be at least 2, got ${$a.rank}.`);
const shape = $a.shape;
const [M, N] = $a.shape.slice(-2);
let $numLower;
let $numUpper;
if (typeof numLower === 'number') {
assert$1(numLower % 1 === 0, () => `bandPart(): numLower must be an integer, got ${numLower}.`);
assert$1(numLower <= M, () => `bandPart(): numLower (${numLower})` +
` must not be greater than the number of rows (${M}).`);
$numLower =
convertToTensor(numLower < 0 ? M : numLower, 'numLower', 'bandPart');
}
else {
assert$1(numLower.dtype === 'int32', () => `bandPart(): numLower's dtype must be an int32.`);
// If numLower is a Scalar, checking `numLower <= M` could hurt performance,
// but minimum(numLower, M) could avoid unexpected results.
$numLower = where(less$2(numLower, 0), M, minimum$2(numLower, M));
}
if (typeof numUpper === 'number') {
assert$1(numUpper % 1 === 0, () => `bandPart(): numUpper must be an integer, got ${numUpper}.`);
assert$1(numUpper <= N, () => `bandPart(): numUpper (${numUpper})` +
` must not be greater than the number of columns (${N}).`);
$numUpper =
convertToTensor(numUpper < 0 ? N : numUpper, 'numUpper', 'bandPart');
}
else {
assert$1(numUpper.dtype === 'int32', () => `bandPart(): numUpper's dtype must be an int32.`);
$numUpper = where(less$2(numUpper, 0), N, minimum$2(numUpper, N));
}
const i = reshape$2(range$3(0, M, 1, 'int32'), [-1, 1]);
const j = range$3(0, N, 1, 'int32');
const ij = sub$2(i, j);
const inBand = logicalAnd$2(lessEqual$2(ij, $numLower), greaterEqual$2(ij, neg$2($numUpper)));
const zero = zeros$1([M, N], $a.dtype);
return reshape$2(stack(unstack(reshape$2($a, [-1, M, N]))
.map(mat => where(inBand, mat, zero))), shape);
}
const bandPart = /* @__PURE__ */ op({ bandPart_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Gram-Schmidt orthogonalization.
*
* ```js
* const x = tf.tensor2d([[1, 2], [3, 4]]);
* let y = tf.linalg.gramSchmidt(x);
* y.print();
* console.log('Orthogonalized:');
* y.dot(y.transpose()).print(); // should be nearly the identity matrix.
* console.log('First row direction maintained:');
* const data = await y.array();
* console.log(data[0][1] / data[0][0]); // should be nearly 2.
* ```
*
* @param xs The vectors to be orthogonalized, in one of the two following
* formats:
* - An Array of `tf.Tensor1D`.
* - A `tf.Tensor2D`, i.e., a matrix, in which case the vectors are the rows
* of `xs`.
* In each case, all the vectors must have the same length and the length
* must be greater than or equal to the number of vectors.
* @returns The orthogonalized and normalized vectors or matrix.
* Orthogonalization means that the vectors or the rows of the matrix
* are orthogonal (zero inner products). Normalization means that each
* vector or each row of the matrix has an L2 norm that equals `1`.
*
* @doc {heading:'Operations', subheading:'Linear Algebra', namespace:'linalg'}
*/
function gramSchmidt_(xs) {
let inputIsTensor2D;
if (Array.isArray(xs)) {
inputIsTensor2D = false;
assert$1(xs != null && xs.length > 0, () => 'Gram-Schmidt process: input must not be null, undefined, or ' +
'empty');
const dim = xs[0].shape[0];
for (let i = 1; i < xs.length; ++i) {
assert$1(xs[i].shape[0] === dim, () => 'Gram-Schmidt: Non-unique lengths found in the input vectors: ' +
`(${xs[i].shape[0]} vs. ${dim})`);
}
}
else {
inputIsTensor2D = true;
xs = split$1(xs, xs.shape[0], 0).map(x => squeeze(x, [0]));
}
assert$1(xs.length <= xs[0].shape[0], () => `Gram-Schmidt: Number of vectors (${xs.length}) exceeds ` +
`number of dimensions (${xs[0].shape[0]}).`);
const ys = [];
const xs1d = xs;
for (let i = 0; i < xs.length; ++i) {
ys.push(ENGINE.tidy(() => {
let x = xs1d[i];
if (i > 0) {
for (let j = 0; j < i; ++j) {
const proj = mul(sum$2(mul(ys[j], x)), ys[j]);
x = sub$2(x, proj);
}
}
return div$1(x, norm(x, 'euclidean'));
}));
}
if (inputIsTensor2D) {
return stack(ys, 0);
}
else {
return ys;
}
}
const gramSchmidt = /* @__PURE__ */ op({ gramSchmidt_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Compute QR decomposition of m-by-n matrix using Householder transformation.
*
* Implementation based on
* [http://www.cs.cornell.edu/~bindel/class/cs6210-f09/lec18.pdf]
* (http://www.cs.cornell.edu/~bindel/class/cs6210-f09/lec18.pdf)
*
* ```js
* const a = tf.tensor2d([[1, 2], [3, 4]]);
* let [q, r] = tf.linalg.qr(a);
* console.log('Q');
* q.print();
* console.log('R');
* r.print();
* console.log('Orthogonalized');
* q.dot(q.transpose()).print() // should be nearly the identity matrix.
* console.log('Reconstructed');
* q.dot(r).print(); // should be nearly [[1, 2], [3, 4]];
* ```
*
* @param x The `tf.Tensor` to be QR-decomposed. Must have rank >= 2. Suppose
* it has the shape `[..., M, N]`.
* @param fullMatrices An optional boolean parameter. Defaults to `false`.
* If `true`, compute full-sized `Q`. If `false` (the default),
* compute only the leading N columns of `Q` and `R`.
* @returns An `Array` of two `tf.Tensor`s: `[Q, R]`. `Q` is a unitary matrix,
* i.e., its columns all have unit norm and are mutually orthogonal.
* If `M >= N`,
* If `fullMatrices` is `false` (default),
* - `Q` has a shape of `[..., M, N]`,
* - `R` has a shape of `[..., N, N]`.
* If `fullMatrices` is `true` (default),
* - `Q` has a shape of `[..., M, M]`,
* - `R` has a shape of `[..., M, N]`.
* If `M < N`,
* - `Q` has a shape of `[..., M, M]`,
* - `R` has a shape of `[..., M, N]`.
* @throws If the rank of `x` is less than 2.
*
* @doc {heading:'Operations',
* subheading:'Linear Algebra',
* namespace:'linalg'}
*/
function qr_(x, fullMatrices = false) {
assert$1(x.rank >= 2, () => `qr() requires input tensor to have a rank >= 2, but got rank ${x.rank}`);
if (x.rank === 2) {
return qr2d(x, fullMatrices);
}
else {
// Rank > 2.
// TODO(cais): Below we split the input into individual 2D tensors,
// perform QR decomposition on them and then stack the results back
// together. We should explore whether this can be parallelized.
const outerDimsProd = x.shape.slice(0, x.shape.length - 2)
.reduce((value, prev) => value * prev);
const x2ds = unstack(reshape$2(x, [
outerDimsProd, x.shape[x.shape.length - 2],
x.shape[x.shape.length - 1]
]), 0);
const q2ds = [];
const r2ds = [];
x2ds.forEach(x2d => {
const [q2d, r2d] = qr2d(x2d, fullMatrices);
q2ds.push(q2d);
r2ds.push(r2d);
});
const q = reshape$2(stack(q2ds, 0), x.shape);
const r = reshape$2(stack(r2ds, 0), x.shape);
return [q, r];
}
}
function qr2d(x, fullMatrices = false) {
return ENGINE.tidy(() => {
assert$1(x.shape.length === 2, () => `qr2d() requires a 2D Tensor, but got a ${x.shape.length}D Tensor.`);
const m = x.shape[0];
const n = x.shape[1];
let q = eye(m); // Orthogonal transform so far.
let r = clone(x); // Transformed matrix so far.
const one2D = tensor2d([[1]], [1, 1]);
let w = clone(one2D);
const iters = m >= n ? n : m;
for (let j = 0; j < iters; ++j) {
// This tidy within the for-loop ensures we clean up temporary
// tensors as soon as they are no longer needed.
const rTemp = r;
const wTemp = w;
const qTemp = q;
[w, r, q] = ENGINE.tidy(() => {
// Find H = I - tau * w * w', to put zeros below R(j, j).
const rjEnd1 = slice$2(r, [j, j], [m - j, 1]);
const normX = norm(rjEnd1);
const rjj = slice$2(r, [j, j], [1, 1]);
// The sign() function returns 0 on 0, which causes division by zero.
const s = where(greater$2(rjj, 0), tensor2d([[-1]]), tensor2d([[1]]));
const u1 = sub$2(rjj, mul(s, normX));
const wPre = div$1(rjEnd1, u1);
if (wPre.shape[0] === 1) {
w = clone(one2D);
}
else {
w = concat$2([
one2D,
slice$2(wPre, [1, 0], [wPre.shape[0] - 1, wPre.shape[1]])
], 0);
}
const tau = neg$2(div$1(matMul$1(s, u1), normX));
// -- R := HR, Q := QH.
const rjEndAll = slice$2(r, [j, 0], [m - j, n]);
const tauTimesW = mul(tau, w);
const wT = transpose$2(w);
if (j === 0) {
r = sub$2(rjEndAll, matMul$1(tauTimesW, matMul$1(wT, rjEndAll)));
}
else {
const rTimesTau = sub$2(rjEndAll, matMul$1(tauTimesW, matMul$1(wT, rjEndAll)));
r = concat$2([slice$2(r, [0, 0], [j, n]), rTimesTau], 0);
}
const tawTimesWT = transpose$2(tauTimesW);
const qAllJEnd = slice$2(q, [0, j], [m, q.shape[1] - j]);
if (j === 0) {
q = sub$2(qAllJEnd, matMul$1(matMul$1(qAllJEnd, w), tawTimesWT));
}
else {
const qTimesTau = sub$2(qAllJEnd, matMul$1(matMul$1(qAllJEnd, w), tawTimesWT));
q = concat$2([slice$2(q, [0, 0], [m, j]), qTimesTau], 1);
}
return [w, r, q];
});
dispose([rTemp, wTemp, qTemp]);
}
if (!fullMatrices && m > n) {
q = slice$2(q, [0, 0], [m, n]);
r = slice$2(r, [0, 0], [n, n]);
}
return [q, r];
});
}
const qr = /* @__PURE__ */ op({ qr_ });
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Converts each string in the input Tensor to its hash mod by a number of
* buckets.
*
* The hash function is deterministic on the content of the string within the
* process and will never change. However, it is not suitable for cryptography.
* This function may be used when CPU time is scarce and inputs are trusted or
* unimportant. There is a risk of adversaries constructing inputs that all hash
* to the same bucket.
*
* ```js
* const result = tf.string.stringToHashBucketFast(
* ['Hello', 'TensorFlow', '2.x'], 3);
* result.print(); // [0, 2, 2]
* ```
* @param input: The strings to assign a hash bucket.
* @param numBuckets: The number of buckets.
* @return A Tensor of the same shape as the input tensor.
*
* @doc {heading: 'Operations', subheading: 'String'}
*/
function stringToHashBucketFast_(input, numBuckets) {
const $input = convertToTensor(input, 'input', 'stringToHashBucketFast', 'string');
const attrs = { numBuckets };
if (numBuckets <= 0) {
throw new Error(`Number of buckets must be at least 1`);
}
const inputs = { input: $input };
return ENGINE.runKernel(StringToHashBucketFast, inputs, attrs);
}
const stringToHashBucketFast$2 = /* @__PURE__ */ op({ stringToHashBucketFast_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Modularized ops.
const linalg = {
bandPart,
gramSchmidt,
qr
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Maps to mapping between the custom object and its name.
*
* After registering a custom class, these two maps will add key-value pairs
* for the class object and the registered name.
*
* Therefore we can get the relative registered name by calling
* getRegisteredName() function.
*
* For example:
* GLOBAL_CUSTOM_OBJECT: {key=registeredName: value=corresponding
* CustomObjectClass}
*
* GLOBAL_CUSTOM_NAMES: {key=CustomObjectClass: value=corresponding
* registeredName}
*
*/
const GLOBAL_CUSTOM_OBJECT = new Map();
const GLOBAL_CUSTOM_NAMES = new Map();
/**
* Serializable defines the serialization contract.
*
* TFJS requires serializable classes to return their className when asked
* to avoid issues with minification.
*/
class Serializable {
/**
* Return the class name for this class to use in serialization contexts.
*
* Generally speaking this will be the same thing that constructor.name
* would have returned. However, the class name needs to be robust
* against minification for serialization/deserialization to work properly.
*
* There's also places such as initializers.VarianceScaling, where
* implementation details between different languages led to different
* class hierarchies and a non-leaf node is used for serialization purposes.
*/
getClassName() {
return this.constructor
.className;
}
/**
* Creates an instance of T from a ConfigDict.
*
* This works for most descendants of serializable. A few need to
* provide special handling.
* @param cls A Constructor for the class to instantiate.
* @param config The Configuration for the object.
*/
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config);
}
}
/**
* Maps string keys to class constructors.
*
* Used during (de)serialization from the cross-language JSON format, which
* requires the class name in the serialization format matches the class
* names as used in Python, should it exist.
*/
class SerializationMap {
constructor() {
this.classNameMap = {};
}
/**
* Returns the singleton instance of the map.
*/
static getMap() {
if (SerializationMap.instance == null) {
SerializationMap.instance = new SerializationMap();
}
return SerializationMap.instance;
}
/**
* Registers the class as serializable.
*/
static register(cls) {
SerializationMap.getMap().classNameMap[cls.className] =
[cls, cls.fromConfig];
}
}
/**
* Register a class with the serialization map of TensorFlow.js.
*
* This is often used for registering custom Layers, so they can be
* serialized and deserialized.
*
* Example 1. Register the class without package name and specified name.
*
* ```js
* class MyCustomLayer extends tf.layers.Layer {
* static className = 'MyCustomLayer';
*
* constructor(config) {
* super(config);
* }
* }
* tf.serialization.registerClass(MyCustomLayer);
* console.log(tf.serialization.GLOBALCUSTOMOBJECT.get("Custom>MyCustomLayer"));
* console.log(tf.serialization.GLOBALCUSTOMNAMES.get(MyCustomLayer));
* ```
*
* Example 2. Register the class with package name: "Package" and specified
* name: "MyLayer".
* ```js
* class MyCustomLayer extends tf.layers.Layer {
* static className = 'MyCustomLayer';
*
* constructor(config) {
* super(config);
* }
* }
* tf.serialization.registerClass(MyCustomLayer, "Package", "MyLayer");
* console.log(tf.serialization.GLOBALCUSTOMOBJECT.get("Package>MyLayer"));
* console.log(tf.serialization.GLOBALCUSTOMNAMES.get(MyCustomLayer));
* ```
*
* Example 3. Register the class with specified name: "MyLayer".
* ```js
* class MyCustomLayer extends tf.layers.Layer {
* static className = 'MyCustomLayer';
*
* constructor(config) {
* super(config);
* }
* }
* tf.serialization.registerClass(MyCustomLayer, undefined, "MyLayer");
* console.log(tf.serialization.GLOBALCUSTOMOBJECT.get("Custom>MyLayer"));
* console.log(tf.serialization.GLOBALCUSTOMNAMES.get(MyCustomLayer));
* ```
*
* Example 4. Register the class with specified package name: "Package".
* ```js
* class MyCustomLayer extends tf.layers.Layer {
* static className = 'MyCustomLayer';
*
* constructor(config) {
* super(config);
* }
* }
* tf.serialization.registerClass(MyCustomLayer, "Package");
* console.log(tf.serialization.GLOBALCUSTOMOBJECT
* .get("Package>MyCustomLayer"));
* console.log(tf.serialization.GLOBALCUSTOMNAMES
* .get(MyCustomLayer));
* ```
*
* @param cls The class to be registered. It must have a public static member
* called `className` defined and the value must be a non-empty string.
* @param pkg The package name that this class belongs to. This used to define
* the key in GlobalCustomObject. If not defined, it defaults to `Custom`.
* @param name The name that user specified. It defaults to the actual name of
* the class as specified by its static `className` property.
* @doc {heading: 'Models', subheading: 'Serialization', ignoreCI: true}
*/
function registerClass(cls, pkg, name) {
assert$1(cls.className != null, () => `Class being registered does not have the static className ` +
`property defined.`);
assert$1(typeof cls.className === 'string', () => `className is required to be a string, but got type ` +
typeof cls.className);
assert$1(cls.className.length > 0, () => `Class being registered has an empty-string as its className, ` +
`which is disallowed.`);
if (typeof pkg === 'undefined') {
pkg = 'Custom';
}
if (typeof name === 'undefined') {
name = cls.className;
}
const className = name;
const registerName = pkg + '>' + className;
SerializationMap.register(cls);
GLOBAL_CUSTOM_OBJECT.set(registerName, cls);
GLOBAL_CUSTOM_NAMES.set(cls, registerName);
return cls;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** @doc {heading: 'Training', subheading: 'Classes', namespace: 'train'} */
class Optimizer extends Serializable {
/**
* Executes `f()` and minimizes the scalar output of `f()` by computing
* gradients of y with respect to the list of trainable variables provided by
* `varList`. If no list is provided, it defaults to all trainable variables.
*
* @param f The function to execute and whose output to minimize.
* @param returnCost Whether to return the scalar cost value produced by
* executing `f()`.
* @param varList An optional list of variables to update. If specified, only
* the trainable variables in varList will be updated by minimize. Defaults to
* all trainable variables.
*
* @doc {heading: 'Training', subheading: 'Optimizers'}
*/
minimize(f, returnCost = false, varList) {
const { value, grads } = this.computeGradients(f, varList);
if (varList != null) {
const gradArray = varList.map(v => ({ name: v.name, tensor: grads[v.name] }));
this.applyGradients(gradArray);
}
else {
this.applyGradients(grads);
}
// Dispose gradients.
dispose(grads);
if (returnCost) {
return value;
}
else {
value.dispose();
return null;
}
}
/**
* The number of iterations that this optimizer instance has been invoked for.
*/
get iterations() {
if (this.iterations_ == null) {
this.iterations_ = 0;
}
return this.iterations_;
}
incrementIterations() {
this.iterations_ = this.iterations + 1;
}
/**
* Executes f() and computes the gradient of the scalar output of f() with
* respect to the list of trainable variables provided by `varList`. If no
* list is provided, it defaults to all trainable variables.
*
* @param f The function to execute and whose output to use for computing
* gradients with respect to variables.
* @param varList An optional list of variables to compute gradients with
* respect to. If specified, only the trainable variables in varList will have
* gradients computed with respect to. Defaults to all trainable variables.
*
* @doc {heading: 'Training', subheading: 'Optimizers'}
*/
computeGradients(f, varList) {
return variableGrads(f, varList);
}
/**
* Dispose the variables (if any) owned by this optimizer instance.
*/
dispose() {
if (this.iterations_ != null) {
dispose(this.iterations_);
}
}
async saveIterations() {
if (this.iterations_ == null) {
this.iterations_ = 0;
}
return {
name: 'iter',
// TODO(cais): Use 'int64' type when available.
tensor: scalar(this.iterations_, 'int32')
};
}
async getWeights() {
throw new Error('getWeights() is not implemented for this optimizer yet.');
}
async setWeights(weightValues) {
throw new Error(`setWeights() is not implemented for this optimizer class ` +
`${this.getClassName()}`);
}
/**
* Extract the first element of the weight values and set it
* as the iterations counter variable of this instance of optimizer.
*
* @param weightValues
* @returns Weight values with the first element consumed and excluded.
*/
async extractIterations(weightValues) {
this.iterations_ = (await weightValues[0].tensor.data())[0];
return weightValues.slice(1);
}
}
Object.defineProperty(Optimizer, Symbol.hasInstance, {
value: (instance) => {
return instance.minimize != null && instance.computeGradients != null &&
instance.applyGradients != null;
}
});
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** @doclink Optimizer */
class AdadeltaOptimizer extends Optimizer {
/** @nocollapse */
static get className() {
// Name matters for Python compatibility.
// This is a getter instead of a property because when it's a property, it
// prevents the entire class from being tree-shaken.
return 'Adadelta';
}
constructor(learningRate, rho, epsilon = null) {
super();
this.learningRate = learningRate;
this.rho = rho;
this.epsilon = epsilon;
this.accumulatedGrads = [];
this.accumulatedUpdates = [];
if (epsilon == null) {
this.epsilon = ENGINE.backend.epsilon();
}
}
applyGradients(variableGradients) {
const variableNames = Array.isArray(variableGradients) ?
variableGradients.map(item => item.name) :
Object.keys(variableGradients);
variableNames.forEach((name, i) => {
const value = ENGINE.registeredVariables[name];
const trainable = false;
if (this.accumulatedGrads[i] == null) {
this.accumulatedGrads[i] = {
originalName: `${name}/accum_grad`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
if (this.accumulatedUpdates[i] == null) {
this.accumulatedUpdates[i] = {
originalName: `${name}/accum_var`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
const gradient = Array.isArray(variableGradients) ?
variableGradients[i].tensor :
variableGradients[name];
if (gradient == null) {
return;
}
const accumulatedGrad = this.accumulatedGrads[i].variable;
const accumulatedUpdate = this.accumulatedUpdates[i].variable;
tidy(() => {
const newAccumulatedGrad = add$1(mul(accumulatedGrad, this.rho), mul(square$2(gradient), 1 - this.rho));
const updates = mul(div$1(sqrt$2(add$1(accumulatedUpdate, this.epsilon)), sqrt$2(add$1(accumulatedGrad, this.epsilon))), gradient);
const newAccumulatedUpdate = add$1(mul(accumulatedUpdate, this.rho), mul(square$2(updates), 1 - this.rho));
accumulatedGrad.assign(newAccumulatedGrad);
accumulatedUpdate.assign(newAccumulatedUpdate);
const newValue = add$1(mul(updates, -this.learningRate), value);
value.assign(newValue);
});
});
this.incrementIterations();
}
dispose() {
if (this.accumulatedUpdates != null) {
dispose(this.accumulatedGrads.map(v => v.variable));
dispose(this.accumulatedUpdates.map(v => v.variable));
}
}
async getWeights() {
// Order matters for Python compatibility.
const variables = [...this.accumulatedGrads, ...this.accumulatedUpdates];
return [await this.saveIterations()].concat(variables.map(v => ({ name: v.originalName, tensor: v.variable })));
}
async setWeights(weightValues) {
weightValues = await this.extractIterations(weightValues);
const variableCount = weightValues.length / 2;
const trainable = false;
this.accumulatedGrads =
weightValues.slice(0, variableCount).map(v => ({
originalName: v.name,
variable: v.tensor.variable(trainable)
}));
this.accumulatedUpdates =
weightValues.slice(variableCount, variableCount * 2)
.map(v => ({
originalName: v.name,
variable: v.tensor.variable(trainable)
}));
}
getConfig() {
return {
'learningRate': this.learningRate,
'rho': this.rho,
'epsilon': this.epsilon
};
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config['learningRate'], config['rho'], config['epsilon']);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** @doclink Optimizer */
class AdagradOptimizer extends Optimizer {
/** @nocollapse */
static get className() {
// Name matters for Python compatibility.
// This is a getter instead of a property because when it's a property, it
// prevents the entire class from being tree-shaken.
return 'Adagrad';
}
constructor(learningRate, initialAccumulatorValue = 0.1) {
super();
this.learningRate = learningRate;
this.initialAccumulatorValue = initialAccumulatorValue;
this.accumulatedGrads = [];
}
applyGradients(variableGradients) {
const variableNames = Array.isArray(variableGradients) ?
variableGradients.map(item => item.name) :
Object.keys(variableGradients);
variableNames.forEach((name, i) => {
const value = ENGINE.registeredVariables[name];
if (this.accumulatedGrads[i] == null) {
const trainable = false;
this.accumulatedGrads[i] = {
originalName: `${name}/accumulator`,
variable: tidy(() => fill$2(value.shape, this.initialAccumulatorValue)
.variable(trainable))
};
}
const gradient = Array.isArray(variableGradients) ?
variableGradients[i].tensor :
variableGradients[name];
if (gradient == null) {
return;
}
const accumulatedGrad = this.accumulatedGrads[i].variable;
tidy(() => {
const newAccumulatedGrad = add$1(accumulatedGrad, square$2(gradient));
accumulatedGrad.assign(newAccumulatedGrad);
const newValue = add$1(mul(div$1(gradient, sqrt$2(add$1(newAccumulatedGrad, ENGINE.backend.epsilon()))), -this.learningRate), value);
value.assign(newValue);
});
});
this.incrementIterations();
}
dispose() {
if (this.accumulatedGrads != null) {
dispose(this.accumulatedGrads.map(v => v.variable));
}
}
async getWeights() {
// Order matters for Python compatibility.
return [await this.saveIterations()].concat(this.accumulatedGrads.map(v => ({ name: v.originalName, tensor: v.variable })));
}
async setWeights(weightValues) {
weightValues = await this.extractIterations(weightValues);
const trainable = false;
this.accumulatedGrads = weightValues.map(v => ({ originalName: v.name, variable: v.tensor.variable(trainable) }));
}
getConfig() {
return {
'learningRate': this.learningRate,
'initialAccumulatorValue': this.initialAccumulatorValue,
};
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config['learningRate'], config['initialAccumulatorValue']);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class AdamOptimizer extends Optimizer {
/** @nocollapse */
static get className() {
// Name matters for Python compatibility.
// This is a getter instead of a property because when it's a property, it
// prevents the entire class from being tree-shaken.
return 'Adam';
}
constructor(learningRate, beta1, beta2, epsilon = null) {
super();
this.learningRate = learningRate;
this.beta1 = beta1;
this.beta2 = beta2;
this.epsilon = epsilon;
this.accumulatedFirstMoment = [];
this.accumulatedSecondMoment = [];
tidy(() => {
// accB* will be updated by batch.
this.accBeta1 = scalar(beta1).variable();
this.accBeta2 = scalar(beta2).variable();
});
if (epsilon == null) {
this.epsilon = ENGINE.backend.epsilon();
}
}
applyGradients(variableGradients) {
const varNames = Array.isArray(variableGradients) ?
variableGradients.map(v => v.name) :
Object.keys(variableGradients);
tidy(() => {
const oneMinusAccBeta1 = sub$2(1, this.accBeta1);
const oneMinusAccBeta2 = sub$2(1, this.accBeta2);
varNames.forEach((name, i) => {
const value = ENGINE.registeredVariables[name];
const trainable = false;
if (this.accumulatedFirstMoment[i] == null) {
this.accumulatedFirstMoment[i] = {
originalName: `${name}/m`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
if (this.accumulatedSecondMoment[i] == null) {
this.accumulatedSecondMoment[i] = {
originalName: `${name}/v`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
const gradient = Array.isArray(variableGradients) ?
variableGradients[i].tensor :
variableGradients[name];
if (gradient == null) {
return;
}
const firstMoment = this.accumulatedFirstMoment[i].variable;
const secondMoment = this.accumulatedSecondMoment[i].variable;
const newFirstMoment = add$1(mul(firstMoment, this.beta1), mul(gradient, 1 - this.beta1));
const newSecondMoment = add$1(mul(secondMoment, this.beta2), mul(square$2(gradient), 1 - this.beta2));
const biasCorrectedFirstMoment = div$1(newFirstMoment, oneMinusAccBeta1);
const biasCorrectedSecondMoment = div$1(newSecondMoment, oneMinusAccBeta2);
firstMoment.assign(newFirstMoment);
secondMoment.assign(newSecondMoment);
const newValue = add$1(mul(div$1(biasCorrectedFirstMoment, add$1(sqrt$2(biasCorrectedSecondMoment), this.epsilon)), -this.learningRate), value);
value.assign(newValue);
});
this.accBeta1.assign(mul(this.accBeta1, this.beta1));
this.accBeta2.assign(mul(this.accBeta2, this.beta2));
});
this.incrementIterations();
}
dispose() {
this.accBeta1.dispose();
this.accBeta2.dispose();
if (this.accumulatedFirstMoment != null) {
dispose(this.accumulatedFirstMoment.map(v => v.variable));
}
if (this.accumulatedSecondMoment != null) {
dispose(this.accumulatedSecondMoment.map(v => v.variable));
}
}
async getWeights() {
// Order matters for Python compatibility.
const variables = [...this.accumulatedFirstMoment, ...this.accumulatedSecondMoment];
return [await this.saveIterations()].concat(variables.map(v => ({ name: v.originalName, tensor: v.variable })));
}
async setWeights(weightValues) {
weightValues = await this.extractIterations(weightValues);
tidy(() => {
this.accBeta1.assign(pow$2(this.beta1, this.iterations_ + 1));
this.accBeta2.assign(pow$2(this.beta2, this.iterations_ + 1));
});
const variableCount = weightValues.length / 2;
const trainable = false;
this.accumulatedFirstMoment =
weightValues.slice(0, variableCount).map(v => ({
originalName: v.name,
variable: v.tensor.variable(trainable)
}));
this.accumulatedSecondMoment =
weightValues.slice(variableCount, variableCount * 2)
.map(v => ({
originalName: v.name,
variable: v.tensor.variable(trainable)
}));
}
getConfig() {
return {
'learningRate': this.learningRate,
'beta1': this.beta1,
'beta2': this.beta2,
'epsilon': this.epsilon,
};
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config['learningRate'], config['beta1'], config['beta2'], config['epsilon']);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class AdamaxOptimizer extends Optimizer {
/** @nocollapse */
static get className() {
// Name matters for Python compatibility.
// This is a getter instead of a property because when it's a property, it
// prevents the entire class from being tree-shaken.
return 'Adamax';
}
constructor(learningRate, beta1, beta2, epsilon = null, decay = 0.0) {
super();
this.learningRate = learningRate;
this.beta1 = beta1;
this.beta2 = beta2;
this.epsilon = epsilon;
this.decay = decay;
this.accumulatedFirstMoment = [];
this.accumulatedWeightedInfNorm = [];
tidy(() => {
this.iteration = scalar(0).variable();
this.accBeta1 = scalar(beta1).variable();
});
if (epsilon == null) {
this.epsilon = ENGINE.backend.epsilon();
}
}
applyGradients(variableGradients) {
const variableNames = Array.isArray(variableGradients) ?
variableGradients.map(item => item.name) :
Object.keys(variableGradients);
tidy(() => {
const oneMinusAccBeta1 = sub$2(1, this.accBeta1);
const lr = div$1(-this.learningRate, add$1(mul(this.iteration, this.decay), 1));
variableNames.forEach((name, i) => {
const value = ENGINE.registeredVariables[name];
const trainable = false;
if (this.accumulatedFirstMoment[i] == null) {
this.accumulatedFirstMoment[i] = {
originalName: `${name}/m`,
variable: zerosLike$2(value).variable(trainable)
};
}
if (this.accumulatedWeightedInfNorm[i] == null) {
this.accumulatedWeightedInfNorm[i] = {
originalName: `${name}/v`,
variable: zerosLike$2(value).variable(trainable)
};
}
const gradient = Array.isArray(variableGradients) ?
variableGradients[i].tensor :
variableGradients[name];
if (gradient == null) {
return;
}
const firstMoment = this.accumulatedFirstMoment[i].variable;
const weightedInfNorm = this.accumulatedWeightedInfNorm[i].variable;
const newFirstMoment = add$1(mul(firstMoment, this.beta1), mul(gradient, 1 - this.beta1));
const ut0 = mul(weightedInfNorm, this.beta2);
const ut1 = abs$2(gradient);
const newWeightedInfNorm = maximum$2(ut0, ut1);
firstMoment.assign(newFirstMoment);
weightedInfNorm.assign(newWeightedInfNorm);
const newValue = add$1(mul(div$1(lr, oneMinusAccBeta1), div$1(newFirstMoment, add$1(newWeightedInfNorm, this.epsilon))), value);
value.assign(newValue);
});
this.iteration.assign(add$1(this.iteration, 1));
this.accBeta1.assign(mul(this.accBeta1, this.beta1));
});
this.incrementIterations();
}
dispose() {
this.accBeta1.dispose();
this.iteration.dispose();
if (this.accumulatedFirstMoment != null) {
dispose(this.accumulatedFirstMoment.map(v => v.variable));
}
if (this.accumulatedWeightedInfNorm != null) {
dispose(this.accumulatedWeightedInfNorm.map(v => v.variable));
}
}
async getWeights() {
throw new Error('getWeights() is not implemented for Adamax yet.');
}
async setWeights(weightValues) {
throw new Error('setWeights() is not implemented for Adamax yet.');
}
getConfig() {
return {
'learningRate': this.learningRate,
'beta1': this.beta1,
'beta2': this.beta2,
'epsilon': this.epsilon,
'decay': this.decay
};
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config['learningRate'], config['beta1'], config['beta2'], config['epsilon'], config['decay']);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** @doclink Optimizer */
class SGDOptimizer extends Optimizer {
/** @nocollapse */
static get className() {
// Name matters for Python compatibility.
// This is a getter instead of a property because when it's a property, it
// prevents the entire class from being tree-shaken.
return 'SGD';
}
constructor(learningRate) {
super();
this.learningRate = learningRate;
this.setLearningRate(learningRate);
}
applyGradients(variableGradients) {
const varNames = Array.isArray(variableGradients) ?
variableGradients.map(v => v.name) :
Object.keys(variableGradients);
varNames.forEach((name, i) => {
const gradient = Array.isArray(variableGradients) ?
variableGradients[i].tensor :
variableGradients[name];
if (gradient == null) {
return;
}
const value = ENGINE.registeredVariables[name];
tidy(() => {
const newValue = add$1(mul(this.c, gradient), value);
value.assign(newValue);
});
});
this.incrementIterations();
}
/**
* Sets the learning rate of the optimizer.
*/
setLearningRate(learningRate) {
this.learningRate = learningRate;
if (this.c != null) {
this.c.dispose();
}
this.c = keep(scalar(-learningRate));
}
dispose() {
this.c.dispose();
}
async getWeights() {
return [await this.saveIterations()];
}
async setWeights(weightValues) {
weightValues = await this.extractIterations(weightValues);
if (weightValues.length !== 0) {
throw new Error('SGD optimizer does not have settable weights.');
}
}
getConfig() {
return { 'learningRate': this.learningRate };
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config['learningRate']);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** @doclink Optimizer */
class MomentumOptimizer extends SGDOptimizer {
/** @nocollapse */
// Name matters for Python compatibility.
static get className() {
// Name matters for Python compatibility.
// This is a getter instead of a property because when it's a property, it
// prevents the entire class from being tree-shaken.
return 'Momentum';
}
constructor(learningRate, momentum, useNesterov = false) {
super(learningRate);
this.learningRate = learningRate;
this.momentum = momentum;
this.useNesterov = useNesterov;
this.accumulations = [];
this.m = scalar(this.momentum);
}
applyGradients(variableGradients) {
const variableNames = Array.isArray(variableGradients) ?
variableGradients.map(item => item.name) :
Object.keys(variableGradients);
variableNames.forEach((name, i) => {
const value = ENGINE.registeredVariables[name];
if (this.accumulations[i] == null) {
const trainable = false;
this.accumulations[i] = {
originalName: `${name}/momentum`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
const accumulation = this.accumulations[i].variable;
const gradient = Array.isArray(variableGradients) ?
variableGradients[i].tensor :
variableGradients[name];
if (gradient == null) {
return;
}
tidy(() => {
let newValue;
const newAccumulation = add$1(mul(this.m, accumulation), gradient);
if (this.useNesterov) {
newValue = add$1(mul(this.c, add$1(gradient, mul(newAccumulation, this.m))), value);
}
else {
newValue = add$1(mul(this.c, newAccumulation), value);
}
accumulation.assign(newAccumulation);
value.assign(newValue);
});
});
this.incrementIterations();
}
dispose() {
this.m.dispose();
if (this.accumulations != null) {
dispose(this.accumulations.map(v => v.variable));
}
}
/**
* Sets the momentum of the optimizer.
*
* @param momentum
*/
setMomentum(momentum) {
this.momentum = momentum;
}
async getWeights() {
// Order matters for Python compatibility.
return [await this.saveIterations()].concat(this.accumulations.map(v => ({ name: v.originalName, tensor: v.variable })));
}
async setWeights(weightValues) {
weightValues = await this.extractIterations(weightValues);
const trainable = false;
this.accumulations = weightValues.map(v => ({ originalName: v.name, variable: v.tensor.variable(trainable) }));
}
getConfig() {
return {
'learningRate': this.learningRate,
'momentum': this.momentum,
'useNesterov': this.useNesterov
};
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config['learningRate'], config['momentum'], config['useNesterov']);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** @doclink Optimizer */
class RMSPropOptimizer extends Optimizer {
/** @nocollapse */
static get className() {
// Name matters for Python compatibility.
// This is a getter instead of a property because when it's a property, it
// prevents the entire class from being tree-shaken.
return 'RMSProp';
}
constructor(learningRate, decay = 0.9, momentum = 0.0, epsilon = null, centered = false) {
super();
this.learningRate = learningRate;
this.decay = decay;
this.momentum = momentum;
this.epsilon = epsilon;
this.accumulatedMeanSquares = [];
this.accumulatedMoments = [];
this.accumulatedMeanGrads = [];
this.centered = centered;
if (epsilon == null) {
this.epsilon = ENGINE.backend.epsilon();
}
if (learningRate == null) {
throw new Error(`learningRate for RMSPropOptimizer must be defined.`);
}
}
applyGradients(variableGradients) {
const variableNames = Array.isArray(variableGradients) ?
variableGradients.map(item => item.name) :
Object.keys(variableGradients);
variableNames.forEach((name, i) => {
const value = ENGINE.registeredVariables[name];
const trainable = false;
if (this.accumulatedMeanSquares[i] == null) {
this.accumulatedMeanSquares[i] = {
originalName: `${name}/rms`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
if (this.accumulatedMoments[i] == null) {
this.accumulatedMoments[i] = {
originalName: `${name}/momentum`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
if (this.accumulatedMeanGrads[i] == null && this.centered) {
this.accumulatedMeanGrads[i] = {
originalName: `${name}/mg`,
variable: tidy(() => zerosLike$2(value).variable(trainable))
};
}
const gradient = Array.isArray(variableGradients) ?
variableGradients[i].tensor :
variableGradients[name];
if (gradient == null) {
return;
}
const accumulatedMeanSquare = this.accumulatedMeanSquares[i].variable;
const accumulatedMoments = this.accumulatedMoments[i].variable;
tidy(() => {
const newAccumulatedMeanSquare = add$1(mul(accumulatedMeanSquare, this.decay), mul(square$2(gradient), 1 - this.decay));
if (this.centered) {
const accumulatedMeanGrad = this.accumulatedMeanGrads[i].variable;
// Centered gradient
const newAccumulatedMeanGrad = add$1(mul(accumulatedMeanGrad, this.decay), mul(gradient, 1 - this.decay));
const gradContribution = div$1(mul(gradient, this.learningRate), sqrt$2(sub$2(newAccumulatedMeanSquare, add$1(square$2(newAccumulatedMeanGrad), this.epsilon))));
const newAccumulatedMoments = add$1(mul(accumulatedMoments, this.momentum), gradContribution);
accumulatedMeanSquare.assign(newAccumulatedMeanSquare);
accumulatedMeanGrad.assign(newAccumulatedMeanGrad);
accumulatedMoments.assign(newAccumulatedMoments);
const newValue = sub$2(value, newAccumulatedMoments);
value.assign(newValue);
}
else {
// Plain gradient
const newAccumulatedMeanSquare = add$1(mul(accumulatedMeanSquare, this.decay), mul(square$2(gradient), 1 - this.decay));
const newAccumulatedMoments = add$1(mul(accumulatedMoments, this.momentum), div$1(mul(gradient, this.learningRate), sqrt$2(add$1(newAccumulatedMeanSquare, this.epsilon))));
accumulatedMeanSquare.assign(newAccumulatedMeanSquare);
accumulatedMoments.assign(newAccumulatedMoments);
const newValue = sub$2(value, newAccumulatedMoments);
value.assign(newValue);
}
});
});
this.incrementIterations();
}
dispose() {
if (this.accumulatedMeanSquares != null) {
dispose(this.accumulatedMeanSquares.map(v => v.variable));
}
if (this.accumulatedMeanGrads != null && this.centered) {
dispose(this.accumulatedMeanGrads.map(v => v.variable));
}
if (this.accumulatedMoments != null) {
dispose(this.accumulatedMoments.map(v => v.variable));
}
}
async getWeights() {
// Order matters for Python compatibility.
const variables = [...this.accumulatedMeanSquares, ...this.accumulatedMoments];
if (this.centered) {
variables.push(...this.accumulatedMeanGrads);
}
return [await this.saveIterations()].concat(variables.map(v => ({ name: v.originalName, tensor: v.variable })));
}
async setWeights(weightValues) {
weightValues = await this.extractIterations(weightValues);
const variableCount = this.centered ? weightValues.length / 3 : weightValues.length / 2;
const trainable = false;
this.accumulatedMeanSquares =
weightValues.slice(0, variableCount).map(v => ({
originalName: v.name,
variable: v.tensor.variable(trainable)
}));
this.accumulatedMoments =
weightValues.slice(variableCount, variableCount * 2)
.map(v => ({
originalName: v.name,
variable: v.tensor.variable(trainable)
}));
if (this.centered) {
this.accumulatedMeanGrads =
weightValues.slice(variableCount * 2, variableCount * 3)
.map(v => ({
originalName: v.name,
variable: v.tensor.variable(trainable)
}));
}
}
getConfig() {
return {
'learningRate': this.learningRate,
'decay': this.decay,
'momentum': this.momentum,
'epsilon': this.epsilon,
'centered': this.centered
};
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls(config['learningRate'], config['decay'], config['momentum'], config['epsilon'], config['centered']);
}
}
/**
* @license
* Copyright 2022 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const OPTIMIZERS = [
AdadeltaOptimizer,
AdagradOptimizer,
AdamOptimizer,
AdamaxOptimizer,
MomentumOptimizer,
RMSPropOptimizer,
SGDOptimizer,
];
function registerOptimizers() {
for (const optimizer of OPTIMIZERS) {
registerClass(optimizer);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/* Type definitions for exporting and importing of models. */
/**
* A map from Tensor dtype to number of bytes per element of the Tensor.
*/
const DTYPE_VALUE_SIZE_MAP = {
'float32': 4,
'float16': 2,
'int32': 4,
'uint16': 2,
'uint8': 1,
'bool': 1,
'complex64': 8
};
/**
* Wraps a list of ArrayBuffers into a `slice()`-able object without allocating
* a large ArrayBuffer.
*
* Allocating large ArrayBuffers (~2GB) can be unstable on Chrome. TFJS loads
* its weights as a list of (usually) 4MB ArrayBuffers and then slices the
* weight tensors out of them. For small models, it's safe to concatenate all
* the weight buffers into a single ArrayBuffer and then slice the weight
* tensors out of it, but for large models, a different approach is needed.
*/
class CompositeArrayBuffer {
/**
* Concatenate a number of ArrayBuffers into one.
*
* @param buffers An array of ArrayBuffers to concatenate, or a single
* ArrayBuffer.
* @returns Result of concatenating `buffers` in order.
*/
static join(buffers) {
return new CompositeArrayBuffer(buffers).slice();
}
constructor(buffers) {
this.shards = [];
this.previousShardIndex = 0;
if (buffers == null) {
return;
}
// Normalize the `buffers` input to be `ArrayBuffer[]`.
if (!(buffers instanceof Array)) {
buffers = [buffers];
}
buffers = buffers.map((bufferOrTypedArray) => {
if (isTypedArray(bufferOrTypedArray)) {
return bufferOrTypedArray.buffer;
}
return bufferOrTypedArray;
});
// Skip setting up shards if there are no buffers.
if (buffers.length === 0) {
return;
}
this.bufferUniformSize = buffers[0].byteLength;
let start = 0;
for (let i = 0; i < buffers.length; i++) {
const buffer = buffers[i];
// Check that all buffers except the last one have the same length.
if (i !== buffers.length - 1 &&
buffer.byteLength !== this.bufferUniformSize) {
// Unset the buffer uniform size, since the buffer sizes are not
// uniform.
this.bufferUniformSize = undefined;
}
// Create the shards, including their start and end points.
const end = start + buffer.byteLength;
this.shards.push({ buffer, start, end });
start = end;
}
// Set the byteLength
if (this.shards.length === 0) {
this.byteLength = 0;
}
this.byteLength = this.shards[this.shards.length - 1].end;
}
slice(start = 0, end = this.byteLength) {
// If there are no shards, then the CompositeArrayBuffer was initialized
// with no data.
if (this.shards.length === 0) {
return new ArrayBuffer(0);
}
// NaN is treated as zero for slicing. This matches ArrayBuffer's behavior.
start = isNaN(Number(start)) ? 0 : start;
end = isNaN(Number(end)) ? 0 : end;
// Fix the bounds to within the array.
start = Math.max(0, start);
end = Math.min(this.byteLength, end);
if (end <= start) {
return new ArrayBuffer(0);
}
const startShardIndex = this.findShardForByte(start);
if (startShardIndex === -1) {
// This should not happen since the start and end indices are always
// within 0 and the composite array's length.
throw new Error(`Could not find start shard for byte ${start}`);
}
const size = end - start;
const outputBuffer = new ArrayBuffer(size);
const outputArray = new Uint8Array(outputBuffer);
let sliced = 0;
for (let i = startShardIndex; i < this.shards.length; i++) {
const shard = this.shards[i];
const globalStart = start + sliced;
const localStart = globalStart - shard.start;
const outputStart = sliced;
const globalEnd = Math.min(end, shard.end);
const localEnd = globalEnd - shard.start;
const outputSlice = new Uint8Array(shard.buffer, localStart, localEnd - localStart);
outputArray.set(outputSlice, outputStart);
sliced += outputSlice.length;
if (end < shard.end) {
break;
}
}
return outputBuffer;
}
/**
* Get the index of the shard that contains the byte at `byteIndex`.
*/
findShardForByte(byteIndex) {
if (this.shards.length === 0 || byteIndex < 0 ||
byteIndex >= this.byteLength) {
return -1;
}
// If the buffers have a uniform size, compute the shard directly.
if (this.bufferUniformSize != null) {
this.previousShardIndex = Math.floor(byteIndex / this.bufferUniformSize);
return this.previousShardIndex;
}
// If the buffers don't have a uniform size, we need to search for the
// shard. That means we need a function to check where the byteIndex lies
// relative to a given shard.
function check(shard) {
if (byteIndex < shard.start) {
return -1;
}
if (byteIndex >= shard.end) {
return 1;
}
return 0;
}
// For efficiency, try the previous shard first.
if (check(this.shards[this.previousShardIndex]) === 0) {
return this.previousShardIndex;
}
// Otherwise, use a generic search function.
// This should almost never end up being used in practice since the weight
// entries should always be in order.
const index = search(this.shards, check);
if (index === -1) {
return -1;
}
this.previousShardIndex = index;
return this.previousShardIndex;
}
}
/**
* Search for an element of a sorted array.
*
* @param sortedArray The sorted array to search
* @param compare A function to compare the current value against the searched
* value. Return 0 on a match, negative if the searched value is less than
* the value passed to the function, and positive if the searched value is
* greater than the value passed to the function.
* @returns The index of the element, or -1 if it's not in the array.
*/
function search(sortedArray, compare) {
// Binary search
let min = 0;
let max = sortedArray.length;
while (min <= max) {
const middle = Math.floor((max - min) / 2) + min;
const side = compare(sortedArray[middle]);
if (side === 0) {
return middle;
}
else if (side < 0) {
max = middle;
}
else {
min = middle + 1;
}
}
return -1;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** Number of bytes reserved for the length of the string. (32bit integer). */
const NUM_BYTES_STRING_LENGTH = 4;
/**
* Encode a map from names to weight values as an ArrayBuffer, along with an
* `Array` of `WeightsManifestEntry` as specification of the encoded weights.
*
* This function does not perform sharding.
*
* This function is the reverse of `decodeWeights`.
*
* @param tensors A map ("dict") from names to tensors.
* @param group Group to which the weights belong (optional).
* @returns A `Promise` of
* - A flat `ArrayBuffer` with all the binary values of the `Tensor`s
* concatenated.
* - An `Array` of `WeightManifestEntry`s, carrying information including
* tensor names, `dtype`s and shapes.
* @throws Error: on unsupported tensor `dtype`.
*/
async function encodeWeights(tensors, group) {
// TODO(adarob, cais): Support quantization.
const specs = [];
const dataPromises = [];
const names = Array.isArray(tensors) ?
tensors.map(tensor => tensor.name) :
Object.keys(tensors);
for (let i = 0; i < names.length; ++i) {
const name = names[i];
const t = Array.isArray(tensors) ? tensors[i].tensor : tensors[name];
if (t.dtype !== 'float32' && t.dtype !== 'int32' && t.dtype !== 'bool' &&
t.dtype !== 'string' && t.dtype !== 'complex64') {
throw new Error(`Unsupported dtype in weight '${name}': ${t.dtype}`);
}
const spec = { name, shape: t.shape, dtype: t.dtype };
if (t.dtype === 'string') {
const utf8bytes = new Promise(async (resolve) => {
const vals = await t.bytes();
const totalNumBytes = vals.reduce((p, c) => p + c.length, 0) +
NUM_BYTES_STRING_LENGTH * vals.length;
const bytes = new Uint8Array(totalNumBytes);
let offset = 0;
for (let i = 0; i < vals.length; i++) {
const val = vals[i];
const bytesOfLength = new Uint8Array(new Uint32Array([val.length]).buffer);
bytes.set(bytesOfLength, offset);
offset += NUM_BYTES_STRING_LENGTH;
bytes.set(val, offset);
offset += val.length;
}
resolve(bytes);
});
dataPromises.push(utf8bytes);
}
else {
dataPromises.push(t.data());
}
if (group != null) {
spec.group = group;
}
specs.push(spec);
}
const tensorValues = await Promise.all(dataPromises);
return { data: concatenateTypedArrays(tensorValues), specs };
}
/**
* Decode flat ArrayBuffer as weights.
*
* This function does not handle sharding.
*
* This function is the reverse of `encodeWeights`.
*
* @param weightData A flat ArrayBuffer or an array of ArrayBuffers carrying the
* binary values of the tensors concatenated in the order specified in
* `specs`.
* @param specs Specifications of the names, dtypes and shapes of the tensors
* whose value are encoded by `buffer`.
* @return A map from tensor name to tensor value, with the names corresponding
* to names in `specs`.
* @throws Error, if any of the tensors has unsupported dtype.
*/
function decodeWeights(weightData, specs) {
// TODO(adarob, cais): Support quantization.
const compositeBuffer = new CompositeArrayBuffer(weightData);
const out = {};
let offset = 0;
for (const spec of specs) {
const byteLength = getWeightBytelength(spec, (start, end) => {
return compositeBuffer.slice(offset + start, offset + end);
});
out[spec.name] = decodeWeight(spec, compositeBuffer
.slice(offset, offset + byteLength));
offset += byteLength;
}
return out;
}
function getWeightBytelength(spec, slice) {
const size = sizeFromShape(spec.shape);
let bytesPerValue;
if ('quantization' in spec) {
const quantization = spec.quantization;
bytesPerValue = DTYPE_VALUE_SIZE_MAP[quantization.dtype];
}
else if (spec.dtype === 'string') {
// Can not statically determine string length.
let byteLength = 0;
for (let i = 0; i < size; i++) {
byteLength += NUM_BYTES_STRING_LENGTH + new Uint32Array(slice(byteLength, byteLength + NUM_BYTES_STRING_LENGTH))[0];
}
return byteLength;
}
else {
bytesPerValue = DTYPE_VALUE_SIZE_MAP[spec.dtype];
}
return size * bytesPerValue;
}
function decodeWeight(spec, byteBuffer) {
const name = spec.name;
const dtype = spec.dtype;
const shape = spec.shape;
const size = sizeFromShape(shape);
let values;
let offset = 0;
if ('quantization' in spec) {
const quantization = spec.quantization;
if (quantization.dtype === 'uint8' || quantization.dtype === 'uint16') {
if (!('min' in quantization && 'scale' in quantization)) {
throw new Error(`Weight ${spec.name} with quantization ${quantization.dtype} ` +
`doesn't have corresponding metadata min and scale.`);
}
}
else if (quantization.dtype === 'float16') {
if (dtype !== 'float32') {
throw new Error(`Weight ${spec.name} is quantized with ${quantization.dtype} ` +
`which only supports weights of type float32 not ${dtype}.`);
}
}
else {
throw new Error(`Weight ${spec.name} has unknown ` +
`quantization dtype ${quantization.dtype}. ` +
`Supported quantization dtypes are: ` +
`'uint8', 'uint16', and 'float16'.`);
}
const quantizationSizeFactor = DTYPE_VALUE_SIZE_MAP[quantization.dtype];
const quantizedArray = (quantization.dtype === 'uint8') ?
new Uint8Array(byteBuffer) :
new Uint16Array(byteBuffer);
if (dtype === 'float32') {
if (quantization.dtype === 'uint8' || quantization.dtype === 'uint16') {
values = new Float32Array(quantizedArray.length);
for (let i = 0; i < quantizedArray.length; i++) {
const v = quantizedArray[i];
values[i] = v * quantization.scale + quantization.min;
}
}
else if (quantization.dtype === 'float16') {
// TODO: This is inefficient. Make getFloat16Decoder efficient.
const float16Decode = getFloat16Decoder();
values = float16Decode(quantizedArray);
}
else {
throw new Error(`Unsupported quantization type ${quantization.dtype} ` +
`for weight type float32.`);
}
}
else if (dtype === 'int32') {
if (quantization.dtype !== 'uint8' && quantization.dtype !== 'uint16') {
throw new Error(`Unsupported quantization type ${quantization.dtype} ` +
`for weight type int32.`);
}
values = new Int32Array(quantizedArray.length);
for (let i = 0; i < quantizedArray.length; i++) {
const v = quantizedArray[i];
values[i] = Math.round(v * quantization.scale + quantization.min);
}
}
else {
throw new Error(`Unsupported dtype in weight '${name}': ${dtype}`);
}
offset += size * quantizationSizeFactor;
}
else if (dtype === 'string') {
const size = sizeFromShape(spec.shape);
values = [];
for (let i = 0; i < size; i++) {
const byteLength = new Uint32Array(byteBuffer.slice(offset, offset + NUM_BYTES_STRING_LENGTH))[0];
offset += NUM_BYTES_STRING_LENGTH;
const bytes = new Uint8Array(byteBuffer.slice(offset, offset + byteLength));
values.push(bytes);
offset += byteLength;
}
}
else {
const dtypeFactor = DTYPE_VALUE_SIZE_MAP[dtype];
if (dtype === 'float32') {
values = new Float32Array(byteBuffer);
}
else if (dtype === 'int32') {
values = new Int32Array(byteBuffer);
}
else if (dtype === 'bool') {
values = new Uint8Array(byteBuffer);
}
else if (dtype === 'complex64') {
values = new Float32Array(byteBuffer);
const real = new Float32Array(values.length / 2);
const image = new Float32Array(values.length / 2);
for (let i = 0; i < real.length; i++) {
real[i] = values[i * 2];
image[i] = values[i * 2 + 1];
}
const realTensor = tensor(real, shape, 'float32');
const imageTensor = tensor(image, shape, 'float32');
const complexTensor = complex$2(realTensor, imageTensor);
realTensor.dispose();
imageTensor.dispose();
return complexTensor;
}
else {
throw new Error(`Unsupported dtype in weight '${name}': ${dtype}`);
}
offset += size * dtypeFactor;
}
return tensor(values, shape, dtype);
}
/**
* Concatenate TypedArrays into an ArrayBuffer.
*/
function concatenateTypedArrays(xs) {
// TODO(adarob, cais): Support quantization.
if (xs === null) {
throw new Error(`Invalid input value: ${JSON.stringify(xs)}`);
}
let totalByteLength = 0;
// `normalizedXs` is here for this reason: a `TypedArray`'s `buffer'
// can have a different byte length from that of the `TypedArray` itself,
// for example, when the `TypedArray` is created from an offset in an
// `ArrayBuffer`. `normliazedXs` holds `TypedArray`s whose `buffer`s match
// the `TypedArray` in byte length. If an element of `xs` does not show
// this property, a new `TypedArray` that satisfy this property will be
// constructed and pushed into `normalizedXs`.
const normalizedXs = [];
xs.forEach((x) => {
totalByteLength += x.byteLength;
// tslint:disable:no-any
normalizedXs.push(x.byteLength === x.buffer.byteLength ? x :
new x.constructor(x));
if (!(x instanceof Float32Array || x instanceof Int32Array ||
x instanceof Uint8Array)) {
throw new Error(`Unsupported TypedArray subtype: ${x.constructor.name}`);
}
// tslint:enable:no-any
});
const y = new Uint8Array(totalByteLength);
let offset = 0;
normalizedXs.forEach((x) => {
y.set(new Uint8Array(x.buffer), offset);
offset += x.byteLength;
});
return y.buffer;
}
// Use Buffer on Node.js instead of Blob/atob/btoa
const useNodeBuffer = typeof Buffer !== 'undefined' &&
(typeof Blob === 'undefined' || typeof atob === 'undefined' ||
typeof btoa === 'undefined');
/**
* Calculate the byte length of a JavaScript string.
*
* Note that a JavaScript string can contain wide characters, therefore the
* length of the string is not necessarily equal to the byte length.
*
* @param str Input string.
* @returns Byte length.
*/
function stringByteLength(str) {
if (useNodeBuffer) {
return Buffer.byteLength(str, 'utf8');
}
return new Blob([str]).size;
}
/**
* Encode an ArrayBuffer as a base64 encoded string.
*
* @param buffer `ArrayBuffer` to be converted.
* @returns A string that base64-encodes `buffer`.
*/
function arrayBufferToBase64String(buffer) {
if (useNodeBuffer) {
return Buffer.from(buffer).toString('base64');
}
const buf = new Uint8Array(buffer);
let s = '';
for (let i = 0, l = buf.length; i < l; i++) {
s += String.fromCharCode(buf[i]);
}
return btoa(s);
}
/**
* Decode a base64 string as an ArrayBuffer.
*
* @param str Base64 string.
* @returns Decoded `ArrayBuffer`.
*/
function base64StringToArrayBuffer(str) {
if (useNodeBuffer) {
const buf = Buffer.from(str, 'base64');
return buf.buffer.slice(buf.byteOffset, buf.byteOffset + buf.byteLength);
}
const s = atob(str);
const buffer = new Uint8Array(s.length);
for (let i = 0; i < s.length; ++i) {
buffer.set([s.charCodeAt(i)], i);
}
return buffer.buffer;
}
/**
* Concatenate a number of ArrayBuffers into one.
*
* @param buffers An array of ArrayBuffers to concatenate, or a single
* ArrayBuffer.
* @returns Result of concatenating `buffers` in order.
*
* @deprecated Use tf.io.CompositeArrayBuffer.join() instead.
*/
function concatenateArrayBuffers(buffers) {
return CompositeArrayBuffer.join(buffers);
}
/**
* Create `ModelJSON` from `ModelArtifacts`.
*
* @param artifacts Model artifacts, describing the model and its weights.
* @param manifest Weight manifest, describing where the weights of the
* `ModelArtifacts` are stored, and some metadata about them.
* @returns Object representing the `model.json` file describing the model
* artifacts and weights
*/
function getModelJSONForModelArtifacts(artifacts, manifest) {
const result = {
modelTopology: artifacts.modelTopology,
format: artifacts.format,
generatedBy: artifacts.generatedBy,
convertedBy: artifacts.convertedBy,
weightsManifest: manifest
};
if (artifacts.signature != null) {
result.signature = artifacts.signature;
}
if (artifacts.userDefinedMetadata != null) {
result.userDefinedMetadata = artifacts.userDefinedMetadata;
}
if (artifacts.modelInitializer != null) {
result.modelInitializer = artifacts.modelInitializer;
}
if (artifacts.initializerSignature != null) {
result.initializerSignature = artifacts.initializerSignature;
}
if (artifacts.trainingConfig != null) {
result.trainingConfig = artifacts.trainingConfig;
}
return result;
}
/**
* Populate ModelArtifactsInfo fields for a model with JSON topology.
* @param modelArtifacts
* @returns A ModelArtifactsInfo object.
*/
function getModelArtifactsInfoForJSON(modelArtifacts) {
if (modelArtifacts.modelTopology instanceof ArrayBuffer) {
throw new Error('Expected JSON model topology, received ArrayBuffer.');
}
return {
dateSaved: new Date(),
modelTopologyType: 'JSON',
modelTopologyBytes: modelArtifacts.modelTopology == null ?
0 :
stringByteLength(JSON.stringify(modelArtifacts.modelTopology)),
weightSpecsBytes: modelArtifacts.weightSpecs == null ?
0 :
stringByteLength(JSON.stringify(modelArtifacts.weightSpecs)),
weightDataBytes: modelArtifacts.weightData == null ?
0 :
new CompositeArrayBuffer(modelArtifacts.weightData).byteLength,
};
}
/**
* Computes mantisa table for casting Float16 to Float32
* See http://www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf
*
* @returns Uint32Array, 2048 mantissa lookup values.
*/
function computeFloat16MantisaTable() {
const convertMantissa = (i) => {
let m = i << 13;
let e = 0;
while ((m & 0x00800000) === 0) {
e -= 0x00800000;
m <<= 1;
}
m &= -8388609;
e += 0x38800000;
return m | e;
};
const mantisaTable = new Uint32Array(2048);
mantisaTable[0] = 0;
for (let i = 1; i < 1024; i++) {
mantisaTable[i] = convertMantissa(i);
}
for (let i = 1024; i < 2048; i++) {
mantisaTable[i] = 0x38000000 + ((i - 1024) << 13);
}
return mantisaTable;
}
/**
* Computes exponent table for casting Float16 to Float32
* See http://www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf
*
* @returns Uint32Array, 64 exponent lookup values.
*/
function computeFloat16ExponentTable() {
const exponentTable = new Uint32Array(64);
exponentTable[0] = 0;
exponentTable[31] = 0x47800000;
exponentTable[32] = 0x80000000;
exponentTable[63] = 0xc7800000;
for (let i = 1; i < 31; i++) {
exponentTable[i] = i << 23;
}
for (let i = 33; i < 63; i++) {
exponentTable[i] = 0x80000000 + ((i - 32) << 23);
}
return exponentTable;
}
/**
* Computes offset table for casting Float16 to Float32
* See http://www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf
*
* @returns Uint32Array, 6d offset values.
*/
function computeFloat16OffsetTable() {
const offsetTable = new Uint32Array(64);
for (let i = 0; i < 64; i++) {
offsetTable[i] = 1024;
}
offsetTable[0] = offsetTable[32] = 0;
return offsetTable;
}
/**
* Retrieve a Float16 decoder which will decode a ByteArray of Float16 values
* to a Float32Array.
*
* @returns Function (buffer: Uint16Array) => Float32Array which decodes
* the Uint16Array of Float16 bytes to a Float32Array.
*/
function getFloat16Decoder() {
// Algorithm is based off of
// http://www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf
// Cache lookup tables
const mantisaTable = computeFloat16MantisaTable();
const exponentTable = computeFloat16ExponentTable();
const offsetTable = computeFloat16OffsetTable();
return (quantizedArray) => {
const buffer = new ArrayBuffer(4 * quantizedArray.length);
const bufferUint32View = new Uint32Array(buffer);
for (let index = 0; index < quantizedArray.length; index++) {
const float16Bits = quantizedArray[index];
const float32Bits = mantisaTable[offsetTable[float16Bits >> 10] + (float16Bits & 0x3ff)] +
exponentTable[float16Bits >> 10];
bufferUint32View[index] = float32Bits;
}
return new Float32Array(buffer);
};
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class IORouterRegistry {
constructor() {
this.saveRouters = [];
this.loadRouters = [];
}
static getInstance() {
if (IORouterRegistry.instance == null) {
IORouterRegistry.instance = new IORouterRegistry();
}
return IORouterRegistry.instance;
}
/**
* Register a save-handler router.
*
* @param saveRouter A function that maps a URL-like string onto an instance
* of `IOHandler` with the `save` method defined or `null`.
*/
static registerSaveRouter(saveRouter) {
IORouterRegistry.getInstance().saveRouters.push(saveRouter);
}
/**
* Register a load-handler router.
*
* @param loadRouter A function that maps a URL-like string onto an instance
* of `IOHandler` with the `load` method defined or `null`.
*/
static registerLoadRouter(loadRouter) {
IORouterRegistry.getInstance().loadRouters.push(loadRouter);
}
/**
* Look up IOHandler for saving, given a URL-like string.
*
* @param url
* @returns If only one match is found, an instance of IOHandler with the
* `save` method defined. If no match is found, `null`.
* @throws Error, if more than one match is found.
*/
static getSaveHandlers(url) {
return IORouterRegistry.getHandlers(url, 'save');
}
/**
* Look up IOHandler for loading, given a URL-like string.
*
* @param url
* @param loadOptions Optional, custom load options.
* @returns All valid handlers for `url`, given the currently registered
* handler routers.
*/
static getLoadHandlers(url, loadOptions) {
return IORouterRegistry.getHandlers(url, 'load', loadOptions);
}
static getHandlers(url, handlerType, loadOptions) {
const validHandlers = [];
const routers = handlerType === 'load' ?
IORouterRegistry.getInstance().loadRouters :
IORouterRegistry.getInstance().saveRouters;
routers.forEach(router => {
const handler = router(url, loadOptions);
if (handler !== null) {
validHandlers.push(handler);
}
});
return validHandlers;
}
}
const getSaveHandlers = (url) => IORouterRegistry.getSaveHandlers(url);
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const DATABASE_NAME = 'tensorflowjs';
const DATABASE_VERSION = 1;
// Model data and ModelArtifactsInfo (metadata) are stored in two separate
// stores for efficient access of the list of stored models and their metadata.
// 1. The object store for model data: topology, weights and weight manifests.
const MODEL_STORE_NAME = 'models_store';
// 2. The object store for ModelArtifactsInfo, including meta-information such
// as the type of topology (JSON vs binary), byte size of the topology, byte
// size of the weights, etc.
const INFO_STORE_NAME = 'model_info_store';
function getIndexedDBFactory() {
if (!env().getBool('IS_BROWSER')) {
// TODO(cais): Add more info about what IOHandler subtypes are available.
// Maybe point to a doc page on the web and/or automatically determine
// the available IOHandlers and print them in the error message.
throw new Error('Failed to obtain IndexedDB factory because the current environment' +
'is not a web browser.');
}
// tslint:disable-next-line:no-any
const theWindow = typeof window === 'undefined' ? self : window;
const factory = theWindow.indexedDB || theWindow.mozIndexedDB ||
theWindow.webkitIndexedDB || theWindow.msIndexedDB ||
theWindow.shimIndexedDB;
if (factory == null) {
throw new Error('The current browser does not appear to support IndexedDB.');
}
return factory;
}
function setUpDatabase(openRequest) {
const db = openRequest.result;
db.createObjectStore(MODEL_STORE_NAME, { keyPath: 'modelPath' });
db.createObjectStore(INFO_STORE_NAME, { keyPath: 'modelPath' });
}
/**
* IOHandler subclass: Browser IndexedDB.
*
* See the doc string of `browserIndexedDB` for more details.
*/
class BrowserIndexedDB {
constructor(modelPath) {
this.indexedDB = getIndexedDBFactory();
if (modelPath == null || !modelPath) {
throw new Error('For IndexedDB, modelPath must not be null, undefined or empty.');
}
this.modelPath = modelPath;
}
async save(modelArtifacts) {
// TODO(cais): Support saving GraphDef models.
if (modelArtifacts.modelTopology instanceof ArrayBuffer) {
throw new Error('BrowserLocalStorage.save() does not support saving model topology ' +
'in binary formats yet.');
}
return this.databaseAction(this.modelPath, modelArtifacts);
}
async load() {
return this.databaseAction(this.modelPath);
}
/**
* Perform database action to put model artifacts into or read model artifacts
* from IndexedDB object store.
*
* Whether the action is put or get depends on whether `modelArtifacts` is
* specified. If it is specified, the action will be put; otherwise the action
* will be get.
*
* @param modelPath A unique string path for the model.
* @param modelArtifacts If specified, it will be the model artifacts to be
* stored in IndexedDB.
* @returns A `Promise` of `SaveResult`, if the action is put, or a `Promise`
* of `ModelArtifacts`, if the action is get.
*/
databaseAction(modelPath, modelArtifacts) {
return new Promise((resolve, reject) => {
const openRequest = this.indexedDB.open(DATABASE_NAME, DATABASE_VERSION);
openRequest.onupgradeneeded = () => setUpDatabase(openRequest);
openRequest.onsuccess = () => {
const db = openRequest.result;
if (modelArtifacts == null) {
// Read model out from object store.
const modelTx = db.transaction(MODEL_STORE_NAME, 'readonly');
const modelStore = modelTx.objectStore(MODEL_STORE_NAME);
const getRequest = modelStore.get(this.modelPath);
getRequest.onsuccess = () => {
if (getRequest.result == null) {
db.close();
return reject(new Error(`Cannot find model with path '${this.modelPath}' ` +
`in IndexedDB.`));
}
else {
resolve(getRequest.result.modelArtifacts);
}
};
getRequest.onerror = error => {
db.close();
return reject(getRequest.error);
};
modelTx.oncomplete = () => db.close();
}
else {
// Put model into object store.
// Concatenate all the model weights into a single ArrayBuffer. Large
// models (~1GB) have problems saving if they are not concatenated.
// TODO(mattSoulanille): Save large models to multiple indexeddb
// records.
modelArtifacts.weightData = CompositeArrayBuffer.join(modelArtifacts.weightData);
const modelArtifactsInfo = getModelArtifactsInfoForJSON(modelArtifacts);
// First, put ModelArtifactsInfo into info store.
const infoTx = db.transaction(INFO_STORE_NAME, 'readwrite');
let infoStore = infoTx.objectStore(INFO_STORE_NAME);
let putInfoRequest;
try {
putInfoRequest =
infoStore.put({ modelPath: this.modelPath, modelArtifactsInfo });
}
catch (error) {
return reject(error);
}
let modelTx;
putInfoRequest.onsuccess = () => {
// Second, put model data into model store.
modelTx = db.transaction(MODEL_STORE_NAME, 'readwrite');
const modelStore = modelTx.objectStore(MODEL_STORE_NAME);
let putModelRequest;
try {
putModelRequest = modelStore.put({
modelPath: this.modelPath,
modelArtifacts,
modelArtifactsInfo
});
}
catch (error) {
// Sometimes, the serialized value is too large to store.
return reject(error);
}
putModelRequest.onsuccess = () => resolve({ modelArtifactsInfo });
putModelRequest.onerror = error => {
// If the put-model request fails, roll back the info entry as
// well.
infoStore = infoTx.objectStore(INFO_STORE_NAME);
const deleteInfoRequest = infoStore.delete(this.modelPath);
deleteInfoRequest.onsuccess = () => {
db.close();
return reject(putModelRequest.error);
};
deleteInfoRequest.onerror = error => {
db.close();
return reject(putModelRequest.error);
};
};
};
putInfoRequest.onerror = error => {
db.close();
return reject(putInfoRequest.error);
};
infoTx.oncomplete = () => {
if (modelTx == null) {
db.close();
}
else {
modelTx.oncomplete = () => db.close();
}
};
}
};
openRequest.onerror = error => reject(openRequest.error);
});
}
}
BrowserIndexedDB.URL_SCHEME = 'indexeddb://';
const indexedDBRouter = (url) => {
if (!env().getBool('IS_BROWSER')) {
return null;
}
else {
if (!Array.isArray(url) && url.startsWith(BrowserIndexedDB.URL_SCHEME)) {
return browserIndexedDB(url.slice(BrowserIndexedDB.URL_SCHEME.length));
}
else {
return null;
}
}
};
IORouterRegistry.registerSaveRouter(indexedDBRouter);
IORouterRegistry.registerLoadRouter(indexedDBRouter);
/**
* Creates a browser IndexedDB IOHandler for saving and loading models.
*
* ```js
* const model = tf.sequential();
* model.add(
* tf.layers.dense({units: 1, inputShape: [100], activation: 'sigmoid'}));
*
* const saveResult = await model.save('indexeddb://MyModel'));
* console.log(saveResult);
* ```
*
* @param modelPath A unique identifier for the model to be saved. Must be a
* non-empty string.
* @returns An instance of `BrowserIndexedDB` (subclass of `IOHandler`),
* which can be used with, e.g., `tf.Model.save`.
*/
function browserIndexedDB(modelPath) {
return new BrowserIndexedDB(modelPath);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const PATH_SEPARATOR = '/';
const PATH_PREFIX = 'tensorflowjs_models';
const INFO_SUFFIX = 'info';
const MODEL_TOPOLOGY_SUFFIX = 'model_topology';
const WEIGHT_SPECS_SUFFIX = 'weight_specs';
const WEIGHT_DATA_SUFFIX = 'weight_data';
const MODEL_METADATA_SUFFIX = 'model_metadata';
function getModelKeys(path) {
return {
info: [PATH_PREFIX, path, INFO_SUFFIX].join(PATH_SEPARATOR),
topology: [PATH_PREFIX, path, MODEL_TOPOLOGY_SUFFIX].join(PATH_SEPARATOR),
weightSpecs: [PATH_PREFIX, path, WEIGHT_SPECS_SUFFIX].join(PATH_SEPARATOR),
weightData: [PATH_PREFIX, path, WEIGHT_DATA_SUFFIX].join(PATH_SEPARATOR),
modelMetadata: [PATH_PREFIX, path, MODEL_METADATA_SUFFIX].join(PATH_SEPARATOR)
};
}
function removeItems(keys) {
for (const key of Object.values(keys)) {
window.localStorage.removeItem(key);
}
}
/**
* IOHandler subclass: Browser Local Storage.
*
* See the doc string to `browserLocalStorage` for more details.
*/
class BrowserLocalStorage {
constructor(modelPath) {
if (!env().getBool('IS_BROWSER') || typeof window === 'undefined' ||
typeof window.localStorage === 'undefined') {
// TODO(cais): Add more info about what IOHandler subtypes are
// available.
// Maybe point to a doc page on the web and/or automatically determine
// the available IOHandlers and print them in the error message.
throw new Error('The current environment does not support local storage.');
}
this.LS = window.localStorage;
if (modelPath == null || !modelPath) {
throw new Error('For local storage, modelPath must not be null, undefined or empty.');
}
this.modelPath = modelPath;
this.keys = getModelKeys(this.modelPath);
}
/**
* Save model artifacts to browser local storage.
*
* See the documentation to `browserLocalStorage` for details on the saved
* artifacts.
*
* @param modelArtifacts The model artifacts to be stored.
* @returns An instance of SaveResult.
*/
async save(modelArtifacts) {
if (modelArtifacts.modelTopology instanceof ArrayBuffer) {
throw new Error('BrowserLocalStorage.save() does not support saving model topology ' +
'in binary formats yet.');
}
else {
const topology = JSON.stringify(modelArtifacts.modelTopology);
const weightSpecs = JSON.stringify(modelArtifacts.weightSpecs);
const modelArtifactsInfo = getModelArtifactsInfoForJSON(modelArtifacts);
// TODO(mattsoulanille): Support saving models over 2GB that exceed
// Chrome's ArrayBuffer size limit.
const weightBuffer = CompositeArrayBuffer.join(modelArtifacts.weightData);
try {
this.LS.setItem(this.keys.info, JSON.stringify(modelArtifactsInfo));
this.LS.setItem(this.keys.topology, topology);
this.LS.setItem(this.keys.weightSpecs, weightSpecs);
this.LS.setItem(this.keys.weightData, arrayBufferToBase64String(weightBuffer));
// Note that JSON.stringify doesn't write out keys that have undefined
// values, so for some keys, we set undefined instead of a null-ish
// value.
const metadata = {
format: modelArtifacts.format,
generatedBy: modelArtifacts.generatedBy,
convertedBy: modelArtifacts.convertedBy,
signature: modelArtifacts.signature != null ?
modelArtifacts.signature :
undefined,
userDefinedMetadata: modelArtifacts.userDefinedMetadata != null ?
modelArtifacts.userDefinedMetadata :
undefined,
modelInitializer: modelArtifacts.modelInitializer != null ?
modelArtifacts.modelInitializer :
undefined,
initializerSignature: modelArtifacts.initializerSignature != null ?
modelArtifacts.initializerSignature :
undefined,
trainingConfig: modelArtifacts.trainingConfig != null ?
modelArtifacts.trainingConfig :
undefined
};
this.LS.setItem(this.keys.modelMetadata, JSON.stringify(metadata));
return { modelArtifactsInfo };
}
catch (err) {
// If saving failed, clean up all items saved so far.
removeItems(this.keys);
throw new Error(`Failed to save model '${this.modelPath}' to local storage: ` +
`size quota being exceeded is a possible cause of this failure: ` +
`modelTopologyBytes=${modelArtifactsInfo.modelTopologyBytes}, ` +
`weightSpecsBytes=${modelArtifactsInfo.weightSpecsBytes}, ` +
`weightDataBytes=${modelArtifactsInfo.weightDataBytes}.`);
}
}
}
/**
* Load a model from local storage.
*
* See the documentation to `browserLocalStorage` for details on the saved
* artifacts.
*
* @returns The loaded model (if loading succeeds).
*/
async load() {
const info = JSON.parse(this.LS.getItem(this.keys.info));
if (info == null) {
throw new Error(`In local storage, there is no model with name '${this.modelPath}'`);
}
if (info.modelTopologyType !== 'JSON') {
throw new Error('BrowserLocalStorage does not support loading non-JSON model ' +
'topology yet.');
}
const out = {};
// Load topology.
const topology = JSON.parse(this.LS.getItem(this.keys.topology));
if (topology == null) {
throw new Error(`In local storage, the topology of model '${this.modelPath}' ` +
`is missing.`);
}
out.modelTopology = topology;
// Load weight specs.
const weightSpecs = JSON.parse(this.LS.getItem(this.keys.weightSpecs));
if (weightSpecs == null) {
throw new Error(`In local storage, the weight specs of model '${this.modelPath}' ` +
`are missing.`);
}
out.weightSpecs = weightSpecs;
// Load meta-data fields.
const metadataString = this.LS.getItem(this.keys.modelMetadata);
if (metadataString != null) {
const metadata = JSON.parse(metadataString);
out.format = metadata.format;
out.generatedBy = metadata.generatedBy;
out.convertedBy = metadata.convertedBy;
if (metadata.signature != null) {
out.signature = metadata.signature;
}
if (metadata.userDefinedMetadata != null) {
out.userDefinedMetadata = metadata.userDefinedMetadata;
}
if (metadata.modelInitializer != null) {
out.modelInitializer = metadata.modelInitializer;
}
if (metadata.initializerSignature != null) {
out.initializerSignature = metadata.initializerSignature;
}
if (metadata.trainingConfig != null) {
out.trainingConfig = metadata.trainingConfig;
}
}
// Load weight data.
const weightDataBase64 = this.LS.getItem(this.keys.weightData);
if (weightDataBase64 == null) {
throw new Error(`In local storage, the binary weight values of model ` +
`'${this.modelPath}' are missing.`);
}
out.weightData = base64StringToArrayBuffer(weightDataBase64);
return out;
}
}
BrowserLocalStorage.URL_SCHEME = 'localstorage://';
const localStorageRouter = (url) => {
if (!env().getBool('IS_BROWSER')) {
return null;
}
else {
if (!Array.isArray(url) && url.startsWith(BrowserLocalStorage.URL_SCHEME)) {
return browserLocalStorage(url.slice(BrowserLocalStorage.URL_SCHEME.length));
}
else {
return null;
}
}
};
IORouterRegistry.registerSaveRouter(localStorageRouter);
IORouterRegistry.registerLoadRouter(localStorageRouter);
/**
* Factory function for local storage IOHandler.
*
* This `IOHandler` supports both `save` and `load`.
*
* For each model's saved artifacts, four items are saved to local storage.
* - `${PATH_SEPARATOR}/${modelPath}/info`: Contains meta-info about the
* model, such as date saved, type of the topology, size in bytes, etc.
* - `${PATH_SEPARATOR}/${modelPath}/topology`: Model topology. For Keras-
* style models, this is a stringized JSON.
* - `${PATH_SEPARATOR}/${modelPath}/weight_specs`: Weight specs of the
* model, can be used to decode the saved binary weight values (see
* item below).
* - `${PATH_SEPARATOR}/${modelPath}/weight_data`: Concatenated binary
* weight values, stored as a base64-encoded string.
*
* Saving may throw an `Error` if the total size of the artifacts exceed the
* browser-specific quota.
*
* @param modelPath A unique identifier for the model to be saved. Must be a
* non-empty string.
* @returns An instance of `IOHandler`, which can be used with, e.g.,
* `tf.Model.save`.
*/
function browserLocalStorage(modelPath) {
return new BrowserLocalStorage(modelPath);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* IOHandlers related to files, such as browser-triggered file downloads,
* user-selected files in browser.
*/
const DEFAULT_FILE_NAME_PREFIX = 'model';
const DEFAULT_JSON_EXTENSION_NAME = '.json';
const DEFAULT_WEIGHT_DATA_EXTENSION_NAME = '.weights.bin';
function defer(f) {
return new Promise(resolve => setTimeout(resolve)).then(f);
}
class BrowserDownloads {
constructor(fileNamePrefix) {
if (!env().getBool('IS_BROWSER')) {
// TODO(cais): Provide info on what IOHandlers are available under the
// current environment.
throw new Error('browserDownloads() cannot proceed because the current environment ' +
'is not a browser.');
}
if (fileNamePrefix.startsWith(BrowserDownloads.URL_SCHEME)) {
fileNamePrefix = fileNamePrefix.slice(BrowserDownloads.URL_SCHEME.length);
}
if (fileNamePrefix == null || fileNamePrefix.length === 0) {
fileNamePrefix = DEFAULT_FILE_NAME_PREFIX;
}
this.modelJsonFileName = fileNamePrefix + DEFAULT_JSON_EXTENSION_NAME;
this.weightDataFileName =
fileNamePrefix + DEFAULT_WEIGHT_DATA_EXTENSION_NAME;
}
async save(modelArtifacts) {
if (typeof (document) === 'undefined') {
throw new Error('Browser downloads are not supported in ' +
'this environment since `document` is not present');
}
// TODO(mattsoulanille): Support saving models over 2GB that exceed
// Chrome's ArrayBuffer size limit.
const weightBuffer = CompositeArrayBuffer.join(modelArtifacts.weightData);
const weightsURL = window.URL.createObjectURL(new Blob([weightBuffer], { type: 'application/octet-stream' }));
if (modelArtifacts.modelTopology instanceof ArrayBuffer) {
throw new Error('BrowserDownloads.save() does not support saving model topology ' +
'in binary formats yet.');
}
else {
const weightsManifest = [{
paths: ['./' + this.weightDataFileName],
weights: modelArtifacts.weightSpecs
}];
const modelJSON = getModelJSONForModelArtifacts(modelArtifacts, weightsManifest);
const modelJsonURL = window.URL.createObjectURL(new Blob([JSON.stringify(modelJSON)], { type: 'application/json' }));
// If anchor elements are not provided, create them without attaching them
// to parents, so that the downloaded file names can be controlled.
const jsonAnchor = this.modelJsonAnchor == null ?
document.createElement('a') :
this.modelJsonAnchor;
jsonAnchor.download = this.modelJsonFileName;
jsonAnchor.href = modelJsonURL;
// Trigger downloads by evoking a click event on the download anchors.
// When multiple downloads are started synchronously, Firefox will only
// save the last one.
await defer(() => jsonAnchor.dispatchEvent(new MouseEvent('click')));
if (modelArtifacts.weightData != null) {
const weightDataAnchor = this.weightDataAnchor == null ?
document.createElement('a') :
this.weightDataAnchor;
weightDataAnchor.download = this.weightDataFileName;
weightDataAnchor.href = weightsURL;
await defer(() => weightDataAnchor.dispatchEvent(new MouseEvent('click')));
}
return { modelArtifactsInfo: getModelArtifactsInfoForJSON(modelArtifacts) };
}
}
}
BrowserDownloads.URL_SCHEME = 'downloads://';
const browserDownloadsRouter = (url) => {
if (!env().getBool('IS_BROWSER')) {
return null;
}
else {
if (!Array.isArray(url) && url.startsWith(BrowserDownloads.URL_SCHEME)) {
return browserDownloads(url.slice(BrowserDownloads.URL_SCHEME.length));
}
else {
return null;
}
}
};
IORouterRegistry.registerSaveRouter(browserDownloadsRouter);
/**
* Creates an IOHandler that triggers file downloads from the browser.
*
* The returned `IOHandler` instance can be used as model exporting methods such
* as `tf.Model.save` and supports only saving.
*
* ```js
* const model = tf.sequential();
* model.add(tf.layers.dense(
* {units: 1, inputShape: [10], activation: 'sigmoid'}));
* const saveResult = await model.save('downloads://mymodel');
* // This will trigger downloading of two files:
* // 'mymodel.json' and 'mymodel.weights.bin'.
* console.log(saveResult);
* ```
*
* @param fileNamePrefix Prefix name of the files to be downloaded. For use with
* `tf.Model`, `fileNamePrefix` should follow either of the following two
* formats:
* 1. `null` or `undefined`, in which case the default file
* names will be used:
* - 'model.json' for the JSON file containing the model topology and
* weights manifest.
* - 'model.weights.bin' for the binary file containing the binary weight
* values.
* 2. A single string or an Array of a single string, as the file name prefix.
* For example, if `'foo'` is provided, the downloaded JSON
* file and binary weights file will be named 'foo.json' and
* 'foo.weights.bin', respectively.
* @param config Additional configuration for triggering downloads.
* @returns An instance of `BrowserDownloads` `IOHandler`.
*
* @doc {
* heading: 'Models',
* subheading: 'Loading',
* namespace: 'io',
* ignoreCI: true
* }
*/
function browserDownloads(fileNamePrefix = 'model') {
return new BrowserDownloads(fileNamePrefix);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class PassthroughLoader {
constructor(modelArtifacts) {
this.modelArtifacts = modelArtifacts;
}
load() {
return this.modelArtifacts;
}
}
class PassthroughSaver {
constructor(saveHandler) {
this.saveHandler = saveHandler;
}
save(modelArtifacts) {
return this.saveHandler(modelArtifacts);
}
}
class PassthroughAsync {
constructor(handler) {
if (handler.load) {
this.load = () => Promise.resolve(handler.load());
}
if (handler.save) {
this.save = (modelArtifacts) => Promise.resolve(handler.save(modelArtifacts));
}
}
}
/**
* Creates an IOHandler that loads model artifacts from memory.
*
* When used in conjunction with `tf.loadLayersModel`, an instance of
* `tf.LayersModel` (Keras-style) can be constructed from the loaded artifacts.
*
* ```js
* const model = await tf.loadLayersModel(tf.io.fromMemory(
* modelTopology, weightSpecs, weightData));
* ```
*
* @param modelArtifacts a object containing model topology (i.e., parsed from
* the JSON format).
* @param weightSpecs An array of `WeightsManifestEntry` objects describing the
* names, shapes, types, and quantization of the weight data. Optional.
* @param weightData A single `ArrayBuffer` containing the weight data,
* concatenated in the order described by the weightSpecs. Optional.
* @param trainingConfig Model training configuration. Optional.
*
* @returns A passthrough `IOHandler` that simply loads the provided data.
*/
function fromMemory(modelArtifacts, weightSpecs, weightData, trainingConfig) {
const args = arguments;
return new PassthroughAsync(fromMemorySync(...args));
}
/**
* Creates an IOHandler that loads model artifacts from memory.
*
* When used in conjunction with `tf.loadLayersModel`, an instance of
* `tf.LayersModel` (Keras-style) can be constructed from the loaded artifacts.
*
* ```js
* const model = await tf.loadLayersModel(tf.io.fromMemory(
* modelTopology, weightSpecs, weightData));
* ```
*
* @param modelArtifacts a object containing model topology (i.e., parsed from
* the JSON format).
* @param weightSpecs An array of `WeightsManifestEntry` objects describing the
* names, shapes, types, and quantization of the weight data. Optional.
* @param weightData A single `ArrayBuffer` containing the weight data,
* concatenated in the order described by the weightSpecs. Optional.
* @param trainingConfig Model training configuration. Optional.
*
* @returns A passthrough `IOHandlerSync` that simply loads the provided data.
*/
function fromMemorySync(modelArtifacts, weightSpecs, weightData, trainingConfig) {
if (arguments.length === 1) {
const isModelArtifacts = modelArtifacts.modelTopology != null ||
modelArtifacts.weightSpecs != null;
if (isModelArtifacts) {
return new PassthroughLoader(modelArtifacts);
}
else {
// Legacy support: with only modelTopology.
// TODO(cais): Remove this deprecated API.
console.warn('Please call tf.io.fromMemory() with only one argument. ' +
'The argument should be of type ModelArtifacts. ' +
'The multi-argument signature of tf.io.fromMemory() has been ' +
'deprecated and will be removed in a future release.');
return new PassthroughLoader({ modelTopology: modelArtifacts });
}
}
else {
// Legacy support.
// TODO(cais): Remove this deprecated API.
console.warn('Please call tf.io.fromMemory() with only one argument. ' +
'The argument should be of type ModelArtifacts. ' +
'The multi-argument signature of tf.io.fromMemory() has been ' +
'deprecated and will be removed in a future release.');
return new PassthroughLoader({
modelTopology: modelArtifacts,
weightSpecs,
weightData,
trainingConfig
});
}
}
/**
* Creates an IOHandler that passes saved model artifacts to a callback.
*
* ```js
* function handleSave(artifacts) {
* // ... do something with the artifacts ...
* return {modelArtifactsInfo: {...}, ...};
* }
*
* const saveResult = model.save(tf.io.withSaveHandler(handleSave));
* ```
*
* @param saveHandler A function that accepts a `ModelArtifacts` and returns a
* promise that resolves to a `SaveResult`.
*/
function withSaveHandler(saveHandler) {
return new PassthroughSaver(saveHandler);
}
/**
* Validate gather nd inputs.
*
* @param tensor The tensor contains the source values.
* @param indices The tensor contains the indices to slice the source.
*
* @returns [resultShape, numUpdates, sliceSize, strides]
*/
function prepareAndValidate(tensor, indices) {
const tensorRank = tensor.shape.length;
const indicesRank = indices.shape.length;
if (tensorRank < 1) {
throw new Error('tf.gatherND() expects the input to be rank 1 or higher,' +
` but the rank was ${tensorRank}.`);
}
if (indicesRank < 1) {
throw new Error('tf.gatherND() expects the indices to be rank 1 or higher,' +
` but the rank was ${indicesRank}.`);
}
if (indices.dtype !== 'int32') {
throw new Error('tf.gatherND() expects the indices to be int32 type,' +
` but the dtype was ${indices.dtype}.`);
}
if (indices.shape[indicesRank - 1] > tensorRank) {
throw new Error('index innermost dimension length must be <= tensor rank; saw: ' +
`${indices.shape[indicesRank - 1]} vs. ${tensorRank}`);
}
if (sizeFromShape(tensor.shape) === 0) {
throw new Error('Requested more than 0 entries, but input is empty.' +
` Input shape: ${tensor.shape}.`);
}
const indicesShape = indices.shape;
const sliceRank = indicesShape[indicesShape.length - 1];
// The result shape is
// indices.shape[:-1] + params.shape[indices.shape[-1]:]
let nResult = 1;
for (let i = 0; i < indicesShape.length - 1; ++i) {
nResult *= indicesShape[i];
}
const inputShape = tensor.shape;
const resultShape = indicesShape.slice();
resultShape.pop();
let sliceSize = 1;
for (let i = sliceRank; i < tensorRank; ++i) {
sliceSize *= inputShape[i];
resultShape.push(inputShape[i]);
}
const strides = [...computeStrides(tensor.shape).map(stride => stride / sliceSize),
1].slice(0, sliceRank);
return [resultShape, nResult, sliceSize, strides];
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const NEW_AXIS = -2;
const SHRINK_AXIS = -1;
function assertParamsValid(input, begin, size) {
const inputRank = input.shape.length;
assert$1(inputRank === begin.length, () => `Error in slice${inputRank}D: Length of begin ${begin} must ` +
`match the rank of the array (${inputRank}).`);
assert$1(inputRank === size.length, () => `Error in slice${inputRank}D: Length of size ${size} must ` +
`match the rank of the array (${inputRank}).`);
for (let i = 0; i < inputRank; ++i) {
assert$1(begin[i] + size[i] <= input.shape[i], () => `Error in slice${inputRank}D: begin[${i}] + size[${i}] ` +
`(${begin[i] + size[i]}) would overflow input.shape[${i}] (${input.shape[i]})`);
}
}
/** Converts a binary mask to an array of axes. Used in stridedSlice(). */
function maskToAxes(mask) {
const axes = [];
let axis = 0;
while (mask > 0) {
if (mask & 1) {
axes.push(axis);
}
mask /= 2;
axis++;
}
return axes;
}
/** Computes the output shape given the strided slice params. */
function computeOutShape$2(begin, end, strides) {
const size = [];
for (let axis = 0; axis < begin.length; axis++) {
size[axis] = Math.ceil((end[axis] - begin[axis]) / strides[axis]);
}
return size;
}
// Creates full selection at the elided dimensions. If the dimension matches
// the ellipsis mask, override the current stride value. Otherwise, insert.
function stridesWithElidedDims(strides, ellipsisInsertionIndex, numElidedAxes, inputShape) {
const newStrides = [...strides];
for (let i = newStrides.length; i < inputShape.length; i++) {
newStrides.push(1);
}
for (let i = 0; i < numElidedAxes; i++) {
if (i === 0) {
newStrides[ellipsisInsertionIndex] = 1;
}
else {
newStrides.splice(ellipsisInsertionIndex, 0 /* num elements to delete */, 1 /* element to add */);
newStrides.pop();
}
}
return newStrides;
}
function unnormalizeAxis(ellipsisInsertionIndex, numElidedAxes, normalizedAxis) {
if (normalizedAxis <= ellipsisInsertionIndex) {
return normalizedAxis;
}
return normalizedAxis - (numElidedAxes - 1);
}
function getElidedAxes(numElidedAxes, ellipsisInsertionIndex) {
const elidedAxes = [];
for (let i = 0; i < numElidedAxes; i++) {
elidedAxes.push(ellipsisInsertionIndex + i);
}
return elidedAxes;
}
// Normalize the start, end and strides.
function getNormalizedAxes(inputShape, ellipsisAxes, numInterpolatedAxes, begin, end, strides, beginMask, endMask, ellipsisMask) {
const inputRank = inputShape.length;
let normalizedBegin = new Array(inputRank), normalizedEnd = new Array(inputRank), normalizedStrides = new Array(inputRank);
if (ellipsisAxes.length && numInterpolatedAxes > 0) {
const fullIndex = ellipsisAxes[0];
// The ellipsis applies to the masked index as well as any dimensions
// that are interpolated.
const numElidedAxes = numInterpolatedAxes + 1;
normalizedBegin = startIndicesWithElidedDims(beginMask, fullIndex, numElidedAxes, begin, inputShape);
normalizedEnd = stopIndicesWithElidedDims(endMask, fullIndex, numElidedAxes, end, inputShape);
normalizedStrides =
stridesWithElidedDims(strides, fullIndex, numElidedAxes, inputShape);
}
else {
for (let axis = 0; axis < inputRank; axis++) {
normalizedBegin[axis] = startForAxis(beginMask, begin, strides, inputShape, axis, ellipsisMask);
normalizedEnd[axis] =
stopForAxis(endMask, end, strides, inputShape, axis, ellipsisMask);
normalizedStrides[axis] = stridesForAxis(strides, axis, ellipsisMask);
}
}
return {
begin: normalizedBegin,
end: normalizedEnd,
strides: normalizedStrides
};
}
// Creates full selection at the elided dimensions. If the dimension matches
// the ellipsis mask, override the current start value. Otherwise, insert.
function startIndicesWithElidedDims(beginMask, ellipsisInsertionIndex, numElidedAxes, originalBegin, inputShape) {
const newIndices = [...inputShape];
const elidedAxes = getElidedAxes(numElidedAxes, ellipsisInsertionIndex);
for (let axis = 0; axis < newIndices.length; axis++) {
if (elidedAxes.indexOf(axis) > -1) {
newIndices[axis] = 0;
}
else {
const originalAxis = unnormalizeAxis(ellipsisInsertionIndex, numElidedAxes, axis);
let originalValue = originalBegin[originalAxis];
if (beginMask & 1 << originalAxis) {
originalValue = 0;
}
newIndices[axis] = originalValue;
}
}
return newIndices;
}
// Creates full selection at the elided dimensions. If the dimension matches
// the ellipsis mask, override the current stop value. Otherwise, insert.
function stopIndicesWithElidedDims(endMask, ellipsisInsertionIndex, numElidedAxes, originalEnd, inputShape) {
const newIndices = [...inputShape];
const elidedAxes = getElidedAxes(numElidedAxes, ellipsisInsertionIndex);
for (let axis = 0; axis < newIndices.length; axis++) {
if (elidedAxes.indexOf(axis) > -1) {
newIndices[axis] = Number.MAX_SAFE_INTEGER;
}
else {
const originalAxis = unnormalizeAxis(ellipsisInsertionIndex, numElidedAxes, axis);
let originalValue = originalEnd[originalAxis];
if (endMask & 1 << originalAxis) {
originalValue = Number.MAX_SAFE_INTEGER;
}
newIndices[axis] = originalValue;
}
}
for (let i = 0; i < newIndices.length; i++) {
// Handle negative indices
const axisSize = inputShape[i];
if (newIndices[i] < 0) {
newIndices[i] += axisSize;
}
newIndices[i] = clamp(0, newIndices[i], inputShape[i]);
}
return newIndices;
}
function stridesForAxis(strides, axis, ellipsisMask) {
let stride = strides[axis];
if (ellipsisMask & (1 << axis) || stride == null) {
stride = 1;
}
return stride;
}
function startForAxis(beginMask, startIndices, strides, inputShape, axis, ellipsisMask) {
// Begin with the specified index
let start = startIndices[axis];
const stride = strides[axis] || 1;
// Check the axis bit from right of masked axes, or the begin index is not set
// for the axis.
if (beginMask & 1 << axis || ellipsisMask & 1 << axis || start == null) {
if (stride > 0) {
// Forward iteration - use the first element. These values will get
// clamped below (Note: We could have set them to 0 and axis_size-1, but
// use lowest() and max() to maintain symmetry with StopForAxis())
start = Number.MIN_SAFE_INTEGER;
}
else {
// Backward iteration - use the last element.
start = Number.MAX_SAFE_INTEGER;
}
}
// Handle negative indices
const axisSize = inputShape[axis];
if (start < 0) {
start += axisSize;
}
// Clamping
start = clamp(0, start, axisSize - 1);
return start;
}
function stopForAxis(endMask, stopIndices, strides, inputShape, axis, ellipsisMask) {
// Begin with the specified index
let stop = stopIndices[axis];
const stride = strides[axis] || 1;
// Check the axis bit from right of masked axes, or if the stop index is not
// set for this axis.
if (endMask & (1 << axis) || ellipsisMask & (1 << axis) || stop == null) {
if (stride > 0) {
// Forward iteration - use the last element. These values will get
// clamped below
stop = Number.MAX_SAFE_INTEGER;
}
else {
// Backward iteration - use the first element.
stop = Number.MIN_SAFE_INTEGER;
}
}
// Handle negative indices
const axisSize = inputShape[axis];
if (stop < 0) {
stop += axisSize;
}
// Clamping
// Because the end index points one past the last element, we need slightly
// different clamping ranges depending on the direction.
if (stride > 0) {
// Forward iteration
stop = clamp(0, stop, axisSize);
}
else {
// Backward iteration
stop = clamp(-1, stop, axisSize - 1);
}
return stop;
}
/**
* Returns true if the slice occupies a continous set of elements in the
* 'flat' space.
*/
function isSliceContinous(shape, begin, size) {
// Index of the first axis that has size > 1.
let firstNonOneAxis = size.length;
for (let i = 0; i < size.length; i++) {
if (size[i] > 1) {
firstNonOneAxis = i;
break;
}
}
for (let i = firstNonOneAxis + 1; i < size.length; i++) {
if (begin[i] > 0 || size[i] !== shape[i]) {
return false;
}
}
return true;
}
function computeFlatOffset(begin, strides) {
let flatOffset = begin.length > 0 ? begin[begin.length - 1] : 1;
for (let i = 0; i < begin.length - 1; i++) {
flatOffset += begin[i] * strides[i];
}
return flatOffset;
}
function parseSliceParams(x, begin, size) {
// The following logic allows for more ergonomic calls.
let begin_;
const xRank = x.shape.length;
if (typeof begin === 'number') {
begin_ = [begin, ...new Array(xRank - 1).fill(0)];
}
else if (begin.length < xRank) {
begin_ = begin.concat(new Array(xRank - begin.length).fill(0));
}
else {
begin_ = begin.slice();
}
begin_.forEach(d => {
assert$1(d !== -1, () => 'slice() does not support negative begin indexing.');
});
let size_;
if (size == null) {
size_ = new Array(xRank).fill(-1);
}
else if (typeof size === 'number') {
size_ = [size, ...new Array(xRank - 1).fill(-1)];
}
else if (size.length < xRank) {
size_ = size.concat(new Array(xRank - size.length).fill(-1));
}
else {
size_ = size;
}
size_ = size_.map((d, i) => {
if (d >= 0) {
return d;
}
else {
assert$1(d === -1, () => `Negative size values should be exactly -1 but got ` +
`${d} for the slice() size at index ${i}.`);
return x.shape[i] - begin_[i];
}
});
return [begin_, size_];
}
// Convert the slicing specification from a sparse representation to a dense
// representation. This means that all ellipses and newaxis are expanded out.
function sliceInfo(xShape, begin, end, strides, beginMask, endMask, ellipsisMask, newAxisMask, shrinkAxisMask) {
let stridesNonNull;
if (strides == null) {
stridesNonNull = new Array(begin.length);
stridesNonNull.fill(1);
}
else {
stridesNonNull = strides;
}
// Only one non-zero bit is allowed in ellipsisMask, which means ellipsisMask
// is a power of 2. Use bit compares to ensure ellipsisMask is 0 or a power
// of 2. When i is a power of 2, i & (i - 1) is always 0.
// Also ref:
// https://stackoverflow.com/questions/600293/how-to-check-if-a-number-is-a-power-of-2
if (ellipsisMask != null && (ellipsisMask & (ellipsisMask - 1)) !== 0) {
throw new Error('Multiple ellipses in slice is not allowed.');
}
// Step 1: Account for ellipsis and new axis.
// Check for ellipsis and count how many non-newaxis there are after.
let ellipsisSeen = false;
const sparseSpec = {
dims: stridesNonNull.length,
numAddAxisAfterEllipsis: 0,
begin: begin.slice(),
end: end.slice(),
strides: stridesNonNull.slice(),
beginMask,
endMask,
ellipsisMask,
newAxisMask,
shrinkAxisMask
};
for (let i = 0; i < sparseSpec.dims; i++) {
if (ellipsisSeen && ((1 << i) & newAxisMask) !== 0) {
sparseSpec.numAddAxisAfterEllipsis++;
}
if ((1 << i) & ellipsisMask) {
ellipsisSeen = true;
}
}
// If no ellipsis insert one at the end.
if (!ellipsisSeen) {
sparseSpec.ellipsisMask |= (1 << sparseSpec.dims);
sparseSpec.dims++; // this effects loop iteration below
}
// Step 2: Make a sparse spec into a full index spec.
//
// The sparse spec deos not correspond to the number of dimensions.
// Make a dense spec that cooresponds to the number of dimensions.
//
// For example suppose foo[...,3:] on foo.shape = [2, 2, 3] then we need to
// produce the missing beginMask for the first two dimensions i.e. from
// beginMaskSpec = 0, endMaskSpec = 2, we achieve beginMask = 6 (110),
// endMask = 7 (111).
const denseSpec = {
dims: xShape.length,
beginMask: 0,
endMask: 0,
beginValid: false,
endValid: false
};
buildDenseSpec(sparseSpec, denseSpec);
// Step 3: Make implicit ranges (non-zero beginMasks and endMasks) explicit
// and bounds check.
let isIdentity = true;
let sliceDim0 = true;
let isSimpleSlice = true;
const processingShape = [];
const finalShape = [];
for (let i = 0; i < xShape.length; ++i) {
if (denseSpec.strides[i] === 0) {
throw Error(`strides[${i}] must be non-zero`);
}
const shrinkI = !!(denseSpec.shrinkAxisMask & (1 << i));
const dimI = xShape[i];
if (dimI === -1) {
processingShape.push(shrinkI ? 1 : -1);
continue;
}
const masks = [denseSpec.beginMask & (1 << i), denseSpec.endMask & (1 << i)];
const validRange = [
denseSpec.strides[i] > 0 ? 0 : -1,
denseSpec.strides[i] > 0 ? dimI : dimI - 1
];
if (shrinkI && denseSpec.strides[i] <= 0) {
throw Error('only stride 1 allowed on non-range indexing.');
}
isSimpleSlice = isSimpleSlice && (denseSpec.strides[i] === 1);
const beginAndEndMasked = !!((denseSpec.beginMask & (1 << i)) && (denseSpec.endMask & (1 << i)));
if (denseSpec.beginValid && denseSpec.endValid) {
if (shrinkI) {
// If we are shrinking, the end index is now possibly incorrect. In
// particular foo[-1] produces sparseBegin = -1, sparseEnd = 0.
// and canonical puts these to n-1 and 0, which implies a degenerate
// interval. Fortunately, it is now safe to re-create end as begin + 1.
const xFwd = denseSpec.begin[i] < 0 ? dimI + denseSpec.begin[i] :
denseSpec.begin[i];
denseSpec.begin[i] = xFwd;
denseSpec.end[i] = denseSpec.begin[i] + 1;
if (xFwd < 0 || xFwd >= dimI) {
throw Error(`slice index ${denseSpec.begin[i]} of dimension ${i} out of bounds.`);
}
}
else {
denseSpec.begin[i] = canonical(denseSpec.begin[i], 0, denseSpec.strides[i], dimI, masks, validRange);
denseSpec.end[i] = canonical(denseSpec.end[i], 1, denseSpec.strides[i], dimI, masks, validRange);
}
// Update optimization values
const takeAllInDimension = denseSpec.strides[i] === 1 &&
denseSpec.begin[i] === 0 && denseSpec.end[i] === dimI;
isIdentity = isIdentity && takeAllInDimension;
sliceDim0 = sliceDim0 &&
((i === 0 && denseSpec.strides[i] === 1) || takeAllInDimension);
}
else {
isIdentity =
isIdentity && ((denseSpec.strides[i] === 1) && beginAndEndMasked);
sliceDim0 = sliceDim0 &&
((i === 0 && denseSpec.strides[i] === 1) || beginAndEndMasked);
}
// Compute the processing shape (the intermediate Eigen will produce)
let intervalLength;
let knownInterval = false;
if (denseSpec.beginValid && denseSpec.endValid) {
intervalLength = denseSpec.end[i] - denseSpec.begin[i];
knownInterval = true;
}
else if (shrinkI) {
// The dimension is still known as 1 for the processingShape, but will be
// discarded for the final shape.
intervalLength = 1;
knownInterval = true;
}
else if (beginAndEndMasked) {
// Even if we don't have values for begin or end, we do know that this
// dimension covers the whole interval. If we have shape information for
// this dimension, that tells us the interval length.
if (dimI >= 0) {
if (denseSpec.strides[i] < 0) {
intervalLength = -dimI;
}
else {
intervalLength = dimI;
}
knownInterval = true;
}
}
if (knownInterval) {
let sizeI;
// Hold zero if the interval is degenerate, otherwise account for
// remainder
if (intervalLength === 0 ||
((intervalLength < 0) !== (denseSpec.strides[i] < 0))) {
sizeI = 0;
}
else {
sizeI = Math.trunc(intervalLength / denseSpec.strides[i]) +
(intervalLength % denseSpec.strides[i] !== 0 ? 1 : 0);
}
processingShape.push(sizeI);
}
else {
processingShape.push(-1);
}
}
// Step 4: Compute the final shape
//
// newAxis will increase dimension by 1 (with a one-size dimension)
// slices like foo[3, ...] will reduce dimension by 1.
// This cannot be done earlier, because it depends on Step 3.
for (let denseDim = 0; denseDim < denseSpec.finalShapeGatherIndices.length; ++denseDim) {
const gatherIndex = denseSpec.finalShapeGatherIndices[denseDim];
if (gatherIndex >= 0) {
finalShape.push(processingShape[gatherIndex]);
}
else if (gatherIndex === NEW_AXIS) {
finalShape.push(1);
}
}
const finalShapeSparse = finalShape.filter((dim, i) => denseSpec.finalShapeGatherIndices[i] !== NEW_AXIS);
return {
finalShapeSparse,
finalShape,
isIdentity,
sliceDim0,
isSimpleSlice,
begin: denseSpec.begin,
end: denseSpec.end,
strides: denseSpec.strides
};
}
function buildDenseSpec(sparse, dense) {
dense.beginMask = 0;
dense.endMask = 0;
dense.shrinkAxisMask = 0;
let fullIndex = 0;
dense.beginValid = sparse.begin != null;
dense.endValid = sparse.end != null;
dense.begin = new Array(dense.dims);
dense.end = new Array(dense.dims);
dense.strides = new Array(dense.dims);
dense.finalShapeGatherIndices = [];
dense.finalShapeGatherIndicesSparse = [];
dense.inputShapeGatherIndicesSparse = new Array(dense.dims);
for (let i = 0; i < sparse.dims; i++) {
if ((1 << i) & sparse.ellipsisMask) {
// Only the bit that has ellipsis will fall in this condition.
// Expand the ellipsis into the appropriate indices
// Note: this only works because we guaranteed one ellipsis.
const nextIndex = Math.min(dense.dims - (sparse.dims - i) + 1 + sparse.numAddAxisAfterEllipsis, dense.dims);
for (; fullIndex < nextIndex; fullIndex++) {
// newAxis aren't real axis so you have to skip.
dense.begin[fullIndex] = 0;
dense.end[fullIndex] = 0;
dense.strides[fullIndex] = 1;
dense.beginMask |= (1 << fullIndex);
dense.endMask |= (1 << fullIndex);
dense.finalShapeGatherIndices.push(fullIndex);
dense.finalShapeGatherIndicesSparse.push(-1);
dense.inputShapeGatherIndicesSparse[fullIndex] = i;
}
}
else if ((1 << i) & sparse.newAxisMask) {
// Only the bit that has newAxis will fall in this condition.
dense.finalShapeGatherIndices.push(NEW_AXIS);
dense.finalShapeGatherIndicesSparse.push(-1);
}
else {
if (fullIndex === dense.begin.length) {
throw Error(`Index out of range using input dim ${fullIndex}; input ` +
`has only ${dense.dims} dims, ${dense.begin.length}.`);
}
// Gather slicing spec into appropriate index.
if (sparse.begin != null) {
dense.begin[fullIndex] = sparse.begin[i];
}
if (sparse.end != null) {
dense.end[fullIndex] = sparse.end[i];
}
dense.strides[fullIndex] = sparse.strides[i];
if (sparse.beginMask & (1 << i)) {
dense.beginMask |= (1 << fullIndex);
}
if (sparse.endMask & (1 << i)) {
dense.endMask |= (1 << fullIndex);
}
// If shrink, record where to get the dimensionality from (i.e. newAxis)
// creates a fake 1 size dimension. Also remember shrink axis (now in
// dense form) so we can ignore dense.end below.
if (sparse.shrinkAxisMask & (1 << i)) {
dense.finalShapeGatherIndices.push(SHRINK_AXIS);
dense.finalShapeGatherIndicesSparse.push(-1);
dense.shrinkAxisMask |= (1 << fullIndex);
}
else {
dense.finalShapeGatherIndices.push(fullIndex);
// Remember that where in the sparse shape the dense dim comes from.
dense.finalShapeGatherIndicesSparse.push(i);
}
dense.inputShapeGatherIndicesSparse[fullIndex] = i;
fullIndex++;
}
}
}
function canonical(x, c, strideI, dimI, masks, validRange) {
if (masks[c]) {
return strideI > 0 ? validRange[c] : validRange[(c + 1) & 1];
}
else {
const xFwd = x < 0 ? dimI + x : x; // make negative indices positive
return xFwd < validRange[0] ? validRange[0] :
xFwd > validRange[1] ? validRange[1] : xFwd;
}
}
var slice_util = /*#__PURE__*/Object.freeze({
__proto__: null,
assertParamsValid: assertParamsValid,
computeFlatOffset: computeFlatOffset,
computeOutShape: computeOutShape$2,
getNormalizedAxes: getNormalizedAxes,
isSliceContinous: isSliceContinous,
maskToAxes: maskToAxes,
parseSliceParams: parseSliceParams,
sliceInfo: sliceInfo,
startForAxis: startForAxis,
startIndicesWithElidedDims: startIndicesWithElidedDims,
stopForAxis: stopForAxis,
stopIndicesWithElidedDims: stopIndicesWithElidedDims,
stridesForAxis: stridesForAxis,
stridesWithElidedDims: stridesWithElidedDims
});
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class OptimizerConstructors {
/**
* Constructs a `tf.SGDOptimizer` that uses stochastic gradient descent.
*
* ```js
* // Fit a quadratic function by learning the coefficients a, b, c.
* const xs = tf.tensor1d([0, 1, 2, 3]);
* const ys = tf.tensor1d([1.1, 5.9, 16.8, 33.9]);
*
* const a = tf.scalar(Math.random()).variable();
* const b = tf.scalar(Math.random()).variable();
* const c = tf.scalar(Math.random()).variable();
*
* // y = a * x^2 + b * x + c.
* const f = x => a.mul(x.square()).add(b.mul(x)).add(c);
* const loss = (pred, label) => pred.sub(label).square().mean();
*
* const learningRate = 0.01;
* const optimizer = tf.train.sgd(learningRate);
*
* // Train the model.
* for (let i = 0; i < 10; i++) {
* optimizer.minimize(() => loss(f(xs), ys));
* }
*
* // Make predictions.
* console.log(
* `a: ${a.dataSync()}, b: ${b.dataSync()}, c: ${c.dataSync()}`);
* const preds = f(xs).dataSync();
* preds.forEach((pred, i) => {
* console.log(`x: ${i}, pred: ${pred}`);
* });
* ```
*
* @param learningRate The learning rate to use for the SGD algorithm.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static sgd(learningRate) {
return new SGDOptimizer(learningRate);
}
/**
* Constructs a `tf.MomentumOptimizer` that uses momentum gradient
* descent.
*
* See
* [http://proceedings.mlr.press/v28/sutskever13.pdf](
* http://proceedings.mlr.press/v28/sutskever13.pdf)
*
* @param learningRate The learning rate to use for the Momentum gradient
* descent algorithm.
* @param momentum The momentum to use for the momentum gradient descent
* algorithm.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static momentum(learningRate, momentum, useNesterov = false) {
return new MomentumOptimizer(learningRate, momentum, useNesterov);
}
/**
* Constructs a `tf.RMSPropOptimizer` that uses RMSProp gradient
* descent. This implementation uses plain momentum and is not centered
* version of RMSProp.
*
* See
* [http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf](
* http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
*
* @param learningRate The learning rate to use for the RMSProp gradient
* descent algorithm.
* @param decay The discounting factor for the history/coming gradient.
* @param momentum The momentum to use for the RMSProp gradient descent
* algorithm.
* @param epsilon Small value to avoid zero denominator.
* @param centered If true, gradients are normalized by the estimated
* variance of the gradient.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static rmsprop(learningRate, decay = .9, momentum = 0.0, epsilon = null, centered = false) {
return new RMSPropOptimizer(learningRate, decay, momentum, epsilon, centered);
}
/**
* Constructs a `tf.AdamOptimizer` that uses the Adam algorithm.
* See [https://arxiv.org/abs/1412.6980](https://arxiv.org/abs/1412.6980)
*
* @param learningRate The learning rate to use for the Adam gradient
* descent algorithm.
* @param beta1 The exponential decay rate for the 1st moment estimates.
* @param beta2 The exponential decay rate for the 2nd moment estimates.
* @param epsilon A small constant for numerical stability.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adam(learningRate = 0.001, beta1 = 0.9, beta2 = 0.999, epsilon = null) {
return new AdamOptimizer(learningRate, beta1, beta2, epsilon);
}
/**
* Constructs a `tf.AdadeltaOptimizer` that uses the Adadelta algorithm.
* See [https://arxiv.org/abs/1212.5701](https://arxiv.org/abs/1212.5701)
*
* @param learningRate The learning rate to use for the Adadelta gradient
* descent algorithm.
* @param rho The learning rate decay over each update.
* @param epsilon A constant epsilon used to better condition the grad
* update.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adadelta(learningRate = .001, rho = .95, epsilon = null) {
return new AdadeltaOptimizer(learningRate, rho, epsilon);
}
/**
* Constructs a `tf.AdamaxOptimizer` that uses the Adamax algorithm.
* See [https://arxiv.org/abs/1412.6980](https://arxiv.org/abs/1412.6980)
*
* @param learningRate The learning rate to use for the Adamax gradient
* descent algorithm.
* @param beta1 The exponential decay rate for the 1st moment estimates.
* @param beta2 The exponential decay rate for the 2nd moment estimates.
* @param epsilon A small constant for numerical stability.
* @param decay The learning rate decay over each update.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adamax(learningRate = 0.002, beta1 = 0.9, beta2 = 0.999, epsilon = null, decay = 0.0) {
return new AdamaxOptimizer(learningRate, beta1, beta2, epsilon, decay);
}
/**
* Constructs a `tf.AdagradOptimizer` that uses the Adagrad algorithm.
* See
* [http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf](
* http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
* or
* [http://ruder.io/optimizing-gradient-descent/index.html#adagrad](
* http://ruder.io/optimizing-gradient-descent/index.html#adagrad)
*
* @param learningRate The learning rate to use for the Adagrad gradient
* descent algorithm.
* @param initialAccumulatorValue Starting value for the accumulators, must be
* positive.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adagrad(learningRate, initialAccumulatorValue = 0.1) {
return new AdagradOptimizer(learningRate, initialAccumulatorValue);
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const train = OptimizerConstructors;
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const delayCallback = (() => {
if (typeof requestAnimationFrame !== 'undefined') {
return requestAnimationFrame;
}
else if (typeof setImmediate !== 'undefined') {
return setImmediate;
}
return (f) => f(); // no delays
})();
/**
* Returns a promise that resolves when a requestAnimationFrame has completed.
*
* On Node.js this uses setImmediate instead of requestAnimationFrame.
*
* This is simply a sugar method so that users can do the following:
* `await tf.nextFrame();`
*
* @doc {heading: 'Performance', subheading: 'Timing'}
*/
function nextFrame() {
return new Promise(resolve => delayCallback(() => resolve()));
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function assertParamsConsistent(shapes, axis) {
const rank = shapes[0].length;
shapes.forEach((shape, i) => {
assert$1(shape.length === rank, () => `Error in concat${rank}D: rank of tensors[${i}] must be the same ` +
`as the rank of the rest (${rank})`);
});
assert$1(axis >= 0 && axis < rank, () => `Error in concat${rank}D: axis must be between 0 and ${rank - 1}.`);
const firstShape = shapes[0];
shapes.forEach((shape, i) => {
for (let r = 0; r < rank; r++) {
assert$1((r === axis) || (shape[r] === firstShape[r]), () => `Error in concat${rank}D: Shape of tensors[${i}] (${shape}) ` +
`does not match the shape of the rest (${firstShape}) ` +
`along the non-concatenated axis ${i}.`);
}
});
}
function computeOutShape$1(shapes, axis) {
const outputShape = shapes[0].slice();
for (let i = 1; i < shapes.length; i++) {
outputShape[axis] += shapes[i][axis];
}
return outputShape;
}
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
var RowPartitionType$1;
(function (RowPartitionType) {
RowPartitionType[RowPartitionType["FIRST_DIM_SIZE"] = 0] = "FIRST_DIM_SIZE";
RowPartitionType[RowPartitionType["VALUE_ROWIDS"] = 1] = "VALUE_ROWIDS";
RowPartitionType[RowPartitionType["ROW_LENGTHS"] = 2] = "ROW_LENGTHS";
RowPartitionType[RowPartitionType["ROW_SPLITS"] = 3] = "ROW_SPLITS";
RowPartitionType[RowPartitionType["ROW_LIMITS"] = 4] = "ROW_LIMITS";
RowPartitionType[RowPartitionType["ROW_STARTS"] = 5] = "ROW_STARTS";
})(RowPartitionType$1 || (RowPartitionType$1 = {}));
function combineRaggedTensorToTensorShapes(raggedRank, shape, valueShape) {
// Test for consistency of valueShape and shape specified.
// If shape is unspecified and valueShape is specified, then copy
// over the size from the valueShape dimension.
let outputShape = new Array();
if (valueShape == null && shape == null) {
return outputShape;
}
if (shape == null) {
// Here, value_shape must be of known size.
while (outputShape.length < raggedRank + valueShape.length) {
outputShape.push(-1);
}
}
else {
outputShape = shape.slice();
}
if (valueShape == null) {
return outputShape;
}
// At this point, valueShape and output_shape have known ranks.
if (raggedRank + valueShape.length !== outputShape.length) {
throw new Error(`rt input.shape and shape=${shape} are incompatible: rt input.rank = ${raggedRank +
valueShape.length}, but shape.rank = ${outputShape.length}`);
}
for (let i = 1; i < valueShape.length; ++i) {
const valueDim = valueShape[i];
const outputShapeDimIndex = outputShape[outputShape.length - valueShape.length + i];
const outputShapeDim = outputShape[outputShapeDimIndex];
if (valueDim >= 0) {
if (outputShapeDim >= 0) {
if (outputShapeDim !== valueDim) {
throw new Error(`rt input.shape and shape=${shape} are incompatible: rt input.shape[${i + raggedRank}] = ${valueDim} but shape[${i + raggedRank}] = ${outputShapeDim}`);
}
}
else {
outputShape[outputShapeDimIndex] = valueDim;
}
}
}
return outputShape;
}
function getRowPartitionTypesHelper(rowPartitionTypeStrings) {
const stringToType = {
'FIRST_DIM_SIZE': RowPartitionType$1.FIRST_DIM_SIZE,
'VALUE_ROWIDS': RowPartitionType$1.VALUE_ROWIDS,
'ROW_LENGTHS': RowPartitionType$1.ROW_LENGTHS,
'ROW_SPLITS': RowPartitionType$1.ROW_SPLITS,
'ROW_LIMITS': RowPartitionType$1.ROW_LIMITS,
'ROW_STARTS': RowPartitionType$1.ROW_STARTS
};
const result = [];
for (const typeStr of rowPartitionTypeStrings) {
if (typeStr in stringToType) {
result.push(stringToType[typeStr]);
}
else {
break;
}
}
return result;
}
function getRaggedRank(rowPartitionTypes) {
if (rowPartitionTypes.length === 0) {
return 0;
}
if (rowPartitionTypes[0] === RowPartitionType$1.FIRST_DIM_SIZE) {
return rowPartitionTypes.length - 1;
}
return rowPartitionTypes.length;
}
function validateDefaultValueShape(defaultValueShape, valueShape) {
if (defaultValueShape == null || valueShape == null) {
return;
}
const defaultNDims = defaultValueShape.length;
const valuesNDims = valueShape.length;
if (defaultNDims >= valuesNDims) {
throw new Error(`defaultValue.shape=${defaultValueShape} and ragged tensor flatValues.shape=${valueShape}, are incompatible: defaultValue.rank = ${defaultNDims} must be less than ragged tensor input flatValues.rank = ${valuesNDims})`);
}
for (let i = 0; i < Math.min(defaultNDims, valuesNDims - 1); ++i) {
const defaultDim = defaultValueShape[i];
const valueDim = valueShape[i + 1];
if (defaultDim >= 0 && valueDim >= 0 && defaultDim !== 1 &&
defaultDim !== valueDim) {
throw new Error(`defaultValue.shape=${defaultValueShape}, and ragged tensor input flatValues.shape=${valueShape} are incompatible: defaultValue.shape[${i - defaultValueShape.length}] = ${defaultDim} but ragged tensor input.flatValues.shape[${i - defaultValueShape.length}] = ${valueDim}`);
}
}
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Inputs of size above this threshold will be parallelized by calling multiple
* shader programs.
*/
const PARALLELIZE_THRESHOLD = 30;
function computeOptimalWindowSize(inSize) {
if (inSize <= PARALLELIZE_THRESHOLD) {
return inSize;
}
return nearestDivisor(inSize, Math.floor(Math.sqrt(inSize)));
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Returns the image center in pixels.
function getImageCenter(center, imageHeight, imageWidth) {
const centerX = imageWidth * (typeof center === 'number' ? center : center[0]);
const centerY = imageHeight * (typeof center === 'number' ? center : center[1]);
return [centerX, centerY];
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Gets the new shape of the input Tensor after it's been reshaped
* to:
* [blockShape[0], ..., blockShape[M-1], batch / prod(blockShape),
* inputShape[1], ..., inputShape[N-1]]
*
* See step 1: https://www.tensorflow.org/api_docs/python/tf/batch_to_space_nd
*/
function getReshaped(inputShape, blockShape, prod, batchToSpace = true) {
let reshaped = [];
if (batchToSpace) {
reshaped = reshaped.concat(blockShape.slice(0));
reshaped.push(inputShape[0] / prod);
reshaped = reshaped.concat(inputShape.slice(1));
}
else {
reshaped = reshaped.concat(inputShape[0]);
const spatialLength = blockShape.length;
for (let i = 0; i < spatialLength; ++i) {
reshaped =
reshaped.concat([inputShape[i + 1] / blockShape[i], blockShape[i]]);
}
reshaped = reshaped.concat(inputShape.slice(spatialLength + 1));
}
return reshaped;
}
/**
* Gets the permutation that will transpose the dimensions of the
* reshaped tensor to shape:
*
* [batch / prod(block_shape),inputShape[1], blockShape[0], ...,
* inputShape[M], blockShape[M-1],inputShape[M+1], ..., inputShape[N-1]]
*
* see step 2: https://www.tensorflow.org/api_docs/python/tf/batch_to_space_nd
*/
function getPermuted(reshapedRank, blockShapeRank, batchToSpace = true) {
const permuted = [];
if (batchToSpace) {
permuted.push(blockShapeRank);
for (let i = blockShapeRank + 1; i < reshapedRank; ++i) {
if (i <= 2 * blockShapeRank) {
permuted.push(i);
permuted.push(i - (blockShapeRank + 1));
}
else {
permuted.push(i);
}
}
}
else {
const permutedBeforeBatch = [];
const permutedAfterBatch = [];
for (let i = 1; i < reshapedRank; ++i) {
if (i >= blockShapeRank * 2 + 1 || i % 2 === 1) {
permutedAfterBatch.push(i);
}
else {
permutedBeforeBatch.push(i);
}
}
permuted.push(...permutedBeforeBatch);
permuted.push(0);
permuted.push(...permutedAfterBatch);
}
return permuted;
}
/**
* Gets the shape of the reshaped and permuted input Tensor before any cropping
* is applied. The new shape will be:
*
* [batch / prod(blockShape),inputShape[1] * blockShape[0], ...,
* inputShape[M] * blockShape[M-1],inputShape[M+1], ..., inputShape[N-1]]
*
* See step 3: https://www.tensorflow.org/api_docs/python/tf/batch_to_space_nd
*/
function getReshapedPermuted(inputShape, blockShape, prod, batchToSpace = true) {
const reshapedPermuted = [];
if (batchToSpace) {
reshapedPermuted.push(inputShape[0] / prod);
}
else {
reshapedPermuted.push(inputShape[0] * prod);
}
for (let i = 1; i < inputShape.length; ++i) {
if (i <= blockShape.length) {
if (batchToSpace) {
reshapedPermuted.push(blockShape[i - 1] * inputShape[i]);
}
else {
reshapedPermuted.push(inputShape[i] / blockShape[i - 1]);
}
}
else {
reshapedPermuted.push(inputShape[i]);
}
}
return reshapedPermuted;
}
/**
* Converts the crops argument into the beginning coordinates of a slice
* operation.
*/
function getSliceBeginCoords(crops, blockShape) {
const sliceBeginCoords = [0];
for (let i = 0; i < blockShape; ++i) {
sliceBeginCoords.push(crops[i][0]);
}
return sliceBeginCoords;
}
/**
* Converts the crops argument into the size of a slice operation. When
* combined with getSliceBeginCoords this function allows the reshaped and
* permuted Tensor to be cropped to its final output shape of:
*
* inputShape[1] * blockShape[0] - crops[0,0] - crops[0,1], ...,
* inputShape[M] * blockShape[M-1] -crops[M-1,0] -
* crops[M-1,1],inputShape[M+1], ..., inputShape[N-1]]
*
* See step 4: https://www.tensorflow.org/api_docs/python/tf/batch_to_space_nd
*/
function getSliceSize(uncroppedShape, crops, blockShape) {
const sliceSize = uncroppedShape.slice(0, 1);
for (let i = 0; i < blockShape; ++i) {
sliceSize.push(uncroppedShape[i + 1] - crops[i][0] - crops[i][1]);
}
return sliceSize;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SELU_SCALEALPHA = 1.7580993408473768599402175208123;
const SELU_SCALE = 1.0507009873554804934193349852946;
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ERF_P = 0.3275911;
const ERF_A1 = 0.254829592;
const ERF_A2 = -0.284496736;
const ERF_A3 = 1.421413741;
const ERF_A4 = -1.453152027;
const ERF_A5 = 1.061405429;
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Merges real and imaginary Float32Arrays into a single complex Float32Array.
*
* The memory layout is interleaved as follows:
* real: [r0, r1, r2]
* imag: [i0, i1, i2]
* complex: [r0, i0, r1, i1, r2, i2]
*
* This is the inverse of splitRealAndImagArrays.
*
* @param real The real values of the complex tensor values.
* @param imag The imag values of the complex tensor values.
* @returns A complex tensor as a Float32Array with merged values.
*/
function mergeRealAndImagArrays(real, imag) {
if (real.length !== imag.length) {
throw new Error(`Cannot merge real and imag arrays of different lengths. real:` +
`${real.length}, imag: ${imag.length}.`);
}
const result = new Float32Array(real.length * 2);
for (let i = 0; i < result.length; i += 2) {
result[i] = real[i / 2];
result[i + 1] = imag[i / 2];
}
return result;
}
/**
* Splits a complex Float32Array into real and imag parts.
*
* The memory layout is interleaved as follows:
* complex: [r0, i0, r1, i1, r2, i2]
* real: [r0, r1, r2]
* imag: [i0, i1, i2]
*
* This is the inverse of mergeRealAndImagArrays.
*
* @param complex The complex tensor values.
* @returns An object with real and imag Float32Array components of the complex
* tensor.
*/
function splitRealAndImagArrays(complex) {
const real = new Float32Array(complex.length / 2);
const imag = new Float32Array(complex.length / 2);
for (let i = 0; i < complex.length; i += 2) {
real[i / 2] = complex[i];
imag[i / 2] = complex[i + 1];
}
return { real, imag };
}
/**
* Extracts even indexed complex values in the given array.
* @param complex The complex tensor values
*/
function complexWithEvenIndex(complex) {
const len = Math.ceil(complex.length / 4);
const real = new Float32Array(len);
const imag = new Float32Array(len);
for (let i = 0; i < complex.length; i += 4) {
real[Math.floor(i / 4)] = complex[i];
imag[Math.floor(i / 4)] = complex[i + 1];
}
return { real, imag };
}
/**
* Extracts odd indexed complete values in the given array.
* @param complex The complex tensor values
*/
function complexWithOddIndex(complex) {
const len = Math.floor(complex.length / 4);
const real = new Float32Array(len);
const imag = new Float32Array(len);
for (let i = 2; i < complex.length; i += 4) {
real[Math.floor(i / 4)] = complex[i];
imag[Math.floor(i / 4)] = complex[i + 1];
}
return { real, imag };
}
/**
* Get the map representing a complex value in the given array.
* @param complex The complex tensor values.
* @param index An index of the target complex value.
*/
function getComplexWithIndex(complex, index) {
const real = complex[index * 2];
const imag = complex[index * 2 + 1];
return { real, imag };
}
/**
* Insert a given complex value into the TypedArray.
* @param data The array in which the complex value is inserted.
* @param c The complex value to be inserted.
* @param index An index of the target complex value.
*/
function assignToTypedArray(data, real, imag, index) {
data[index * 2] = real;
data[index * 2 + 1] = imag;
}
/**
* Make the list of exponent terms used by FFT.
*/
function exponents(n, inverse) {
const real = new Float32Array(n / 2);
const imag = new Float32Array(n / 2);
for (let i = 0; i < Math.ceil(n / 2); i++) {
const x = (inverse ? 2 : -2) * Math.PI * (i / n);
real[i] = Math.cos(x);
imag[i] = Math.sin(x);
}
return { real, imag };
}
/**
* Make the exponent term used by FFT.
*/
function exponent(k, n, inverse) {
const x = (inverse ? 2 : -2) * Math.PI * (k / n);
const real = Math.cos(x);
const imag = Math.sin(x);
return { real, imag };
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ARROW = '->';
const ARROW_REGEX = /->/g;
const COMMA = ',';
const ELLIPSIS = '...';
/**
* Parse an equation for einsum.
*
* @param equation The einsum equation (e.g., "ij,jk->ik").
* @param numTensors Number of tensors provided along with `equation`. Used to
* check matching number of input tensors.
* @returns An object consisting of the following fields:
* - allDims: all dimension names as strings.
* - summedDims: a list of all dimensions being summed over, as indices to
* the elements of `allDims`.
* - idDims: indices of the dimensions in each input tensor, as indices to
* the elements of `allDims.
*/
function decodeEinsumEquation(equation, numTensors) {
equation = equation.replace(/\s/g, ''); // Remove witespace in equation.
const numArrows = (equation.length - equation.replace(ARROW_REGEX, '').length) /
ARROW.length;
if (numArrows < 1) {
throw new Error('Equations without an arrow are not supported.');
}
else if (numArrows > 1) {
throw new Error(`Equation must contain exactly one arrow ("${ARROW}").`);
}
const [inputString, outputString] = equation.split(ARROW);
assert$1(inputString.indexOf(ELLIPSIS) === -1, () => `The ellipsis notation ("${ELLIPSIS}") is not supported yet.`);
const inputTerms = inputString.split(COMMA);
const numInputs = inputTerms.length;
if (numTensors !== numInputs) {
throw new Error(`Expected ${numInputs} input tensors, received ${numTensors}`);
}
if (numInputs > 2) {
throw new Error('Support for more than 2 input tensors is not implemented yet.');
}
const allDims = [];
for (let i = 0; i < outputString.length; ++i) {
const dimName = outputString[i];
if (!inputTerms.some(inputTerm => inputTerm.indexOf(dimName) !== -1)) {
throw new Error(`Output subscripts contain the label ${dimName} ` +
`not present in the input subscripts.`);
}
if (allDims.indexOf(dimName) === -1) {
allDims.push(dimName);
}
}
for (let i = 0; i < inputString.length; ++i) {
const dimName = inputString[i];
if (allDims.indexOf(dimName) === -1 && dimName !== COMMA) {
allDims.push(dimName);
}
}
const idDims = new Array(inputTerms.length);
for (let i = 0; i < numInputs; ++i) {
if (new Set(inputTerms[i].split('')).size !== inputTerms[i].length) {
throw new Error(`Found duplicate axes in input component ${inputTerms[i]}. ` +
`Support for duplicate axes in input is not implemented yet.`);
}
idDims[i] = [];
for (let j = 0; j < inputTerms[i].length; ++j) {
idDims[i].push(allDims.indexOf(inputTerms[i][j]));
}
}
const numDims = allDims.length; // Number of unique dimensions.
const numOutDims = outputString.length; // Number of output dimensions.
const summedDims = []; // Dimensions being summed over.
for (let i = numOutDims; i < numDims; ++i) {
summedDims.push(i);
}
return { allDims, summedDims, idDims };
}
/**
* Get the permutation for a given input tensor.
*
* @param nDims Total number of dimension of all tensors involved in the einsum
* operation.
* @param idDims Dimension indices involve in the tensor in question.
* @returns An object consisting of the following fields:
* - permutationIndices: Indices to permute the axes of the tensor with.
* - expandDims: Indices to the dimension that need to be expanded from the
* tensor after permutation.
*/
function getEinsumPermutation(nDims, idDims) {
let permutationIndices = new Array(nDims);
permutationIndices.fill(-1);
for (let i = 0; i < idDims.length; ++i) {
permutationIndices[idDims[i]] = i;
}
const expandDims = [];
for (let i = 0; i < nDims; ++i) {
if (permutationIndices[i] === -1) {
expandDims.push(i);
}
}
permutationIndices = permutationIndices.filter(d => d !== -1);
return { permutationIndices, expandDims };
}
/**
* Checks that the dimension sizes from different input tensors match the
* equation.
*/
function checkEinsumDimSizes(nDims, idDims, tensors) {
const dimSizes = new Array(nDims);
for (let i = 0; i < tensors.length; ++i) {
const shape = tensors[i].shape;
for (let j = 0; j < idDims[i].length; ++j) {
if (dimSizes[idDims[i][j]] === undefined) {
dimSizes[idDims[i][j]] = shape[j];
}
else {
assert$1(dimSizes[idDims[i][j]] === shape[j], () => `Expected dimension ${dimSizes[idDims[i][j]]} at axis ${j} ` +
`of input shaped ${JSON.stringify(shape)}, ` +
`but got dimension ${shape[j]}`);
}
}
}
}
/**
* Gets path of computation for einsum.
*
* @param summedDims indices to the dimensions being summed over.
* @param idDims A look up table for the dimensions present in each input
* tensor.Each constituent array contains indices for the dimensions in the
* corresponding input tensor.
*
* @return A map with two fields:
* - path: The path of computation, with each element indicating the dimension
* being summed over after the element-wise multiplication in that step.
* - steps: With the same length as `path`. Each element contains the indices
* to the input tensors being used for element-wise multiplication in the
* corresponding step.
*/
function getEinsumComputePath(summedDims, idDims) {
const path = summedDims;
const steps = [];
let nSteps = 0;
if (summedDims.length === 0) {
// Einsum that involes no summing: e.g., transpose and outer product.
path.push(-1);
}
nSteps = summedDims.length + 1;
for (let i = 0; i < nSteps; ++i) {
steps.push([]);
}
const computedTermIndices = [];
for (let i = 0; i < path.length; ++i) {
const summedDim = path[i];
const termIndices = findTermsWithDim(idDims, summedDim);
for (const termIndex of termIndices) {
if (computedTermIndices.indexOf(termIndex) === -1) {
steps[i].push(termIndex);
computedTermIndices.push(termIndex);
}
}
}
return { path, steps };
}
/** Determines if an axes permutation is the identity permutation. */
function isIdentityPermutation(perm) {
return perm.every((dim, index) => dim === index);
}
function findTermsWithDim(idDims, dim) {
const termIndices = [];
for (let i = 0; i < idDims.length; ++i) {
if (idDims[i].length === 0 || idDims[i].indexOf(dim) !== -1 || dim === -1) {
termIndices.push(i);
}
}
return termIndices;
}
/**
* Prepare the split size array. When the input is a number, the axis is evenly
* divided among the split size. When the input contains the negative value, the
* rest of the axis is allocated toward that.
*/
function prepareSplitSize(x, numOrSizeSplits, axis = 0) {
let splitSizes = [];
if (typeof (numOrSizeSplits) === 'number') {
assert$1(x.shape[axis] % numOrSizeSplits === 0, () => 'Number of splits must evenly divide the axis.');
splitSizes =
new Array(numOrSizeSplits).fill(x.shape[axis] / numOrSizeSplits);
}
else {
const numOfNegs = numOrSizeSplits.reduce((count, value) => {
if (value === -1) {
count += 1;
}
return count;
}, 0);
assert$1(numOfNegs <= 1, () => 'There should be only one negative value in split array.');
const negIndex = numOrSizeSplits.indexOf(-1);
// Allow the number of split array to be -1, which indicates the rest
// of dimension is allocated to that split.
if (negIndex !== -1) {
const total = numOrSizeSplits.reduce((a, b) => b > 0 ? a + b : a);
numOrSizeSplits[negIndex] = x.shape[axis] - total;
}
assert$1(x.shape[axis] === numOrSizeSplits.reduce((a, b) => a + b), () => 'The sum of sizes must match the size of the axis dimension.');
splitSizes = numOrSizeSplits;
}
return splitSizes;
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Generates sparse fill empty rows indices, dense shape mismatch error message.
*
* @param indicesLength The first dimension of indices.
*/
function getSparseFillEmptyRowsIndicesDenseShapeMismatch(indicesLength) {
return `Received SparseTensor with denseShape[0] = 0 but
indices.shape[0] = ${indicesLength}`;
}
/**
* Generates sparse fill empty rows negative index error message.
*
* @param index The index with a negative value.
* @param value The negative value.
*/
function getSparseFillEmptyRowsNegativeIndexErrorMessage(index, value) {
return `indices(${index}, 0) is invalid: ${value} < 0`;
}
/**
* Generates sparse fill empty rows out of range index error message.
*
* @param index The index with an out of range value.
* @param value The out of range value.
* @param limit The upper limit for indices.
*/
function getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(index, value, limit) {
return `indices(${index}, 0) is invalid: ${value} >= ${limit}`;
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Generates sparse reshape multiple negative 1 output dimension error message.
*
* @param dim1 The first dimension with a negative 1 value.
* @param dim2 The second dimension with a negative 1 value.
*/
function getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(dim1, dim2) {
return `only one output dimension may be -1, not both ${dim1} and ${dim2}`;
}
/**
* Generates sparse reshape negative output dimension error message.
*
* @param dim The dimension with a negative value.
* @param value The negative value.
*/
function getSparseReshapeNegativeOutputDimErrorMessage(dim, value) {
return `size ${dim} must be non-negative, not ${value}`;
}
/**
* Generates sparse reshape empty tensor zero output dimension error message.
*
*/
function getSparseReshapeEmptyTensorZeroOutputDimErrorMessage() {
return 'reshape cannot infer the missing input size for an empty tensor ' +
'unless all specified input sizes are non-zero';
}
/**
* Generates sparse reshape input output multiple mismatch error message.
*
* @param inputShape the input shape.
* @param outputShape the requested output shape.
*/
function getSparseReshapeInputOutputMultipleErrorMessage(inputShape, outputShape) {
const inputSize = sizeFromShape(inputShape);
const outputSize = sizeFromShape(outputShape);
return `Input to reshape is a SparseTensor with ${inputSize}
dense values, but the requested shape requires a multiple of ${outputSize}. inputShape=${inputShape} outputShape= ${outputShape}`;
}
/**
* Generates sparse reshape input output inequality error message.
*
* @param inputShape the input shape.
* @param outputShape the requested output shape.
*/
function getSparseReshapeInputOutputMismatchErrorMessage(inputShape, outputShape) {
const inputSize = sizeFromShape(inputShape);
const outputSize = sizeFromShape(outputShape);
return `Input to reshape is a tensor with ${inputSize} dense values, but the requested shape has ${outputSize}. inputShape=${inputShape} outputShape=${outputShape}`;
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Generates sparse segment reduction negative segment ids error message.
*
*/
function getSparseSegmentReductionNegativeSegmentIdsErrorMessage() {
return `segment ids must be >= 0`;
}
/**
* Generates sparse segment reduction non increasing segment ids error message.
*
*/
function getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage() {
return `segment ids are not increasing`;
}
/**
* Generates sparse segment reduction segment id out of range error message.
*
* @param segmentId The segment id index that is out of range.
* @param outputRows Upper bound of valid segment id values.
*/
function getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(segmentId, outputRows) {
return `Segment id ${segmentId} out of range [0, ${outputRows}), possibly because segmentIds input is not sorted.`;
}
/**
* Generates sparse segment reduction input indice out of range error message.
*
* @param index The index that holds the out of range value.
* @param indexValue The value that is out of range.
* @param inputRows Upper bound of valid index values.
*/
function getSparseSegmentReductionIndicesOutOfRangeErrorMessage(index, indexValue, inputRows) {
return `Bad: indices[${index}] == ${indexValue} out of range [0, ${inputRows})`;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function segOpComputeOptimalWindowSize(inSize, numSegments) {
let done = false;
let res;
if (inSize <= PARALLELIZE_THRESHOLD) {
res = inSize;
done = true;
}
else {
res = nearestDivisor(inSize, Math.floor(Math.sqrt(inSize)));
}
while (!done) {
if (res > numSegments || res === inSize) {
done = true;
}
else {
res = nearestDivisor(inSize, res + 1);
}
}
return res;
}
function computeOutShape(aShape, axis, numSegments) {
const outShape = [];
const rank = aShape.length;
for (let dim = 0; dim < rank; dim++) {
if (dim !== axis) {
outShape.push(aShape[dim]);
}
else {
outShape.push(numSegments);
}
}
return outShape;
}
function collectGatherOpShapeInfo(x, indices, axis, batchDims) {
const indicesRank = indices.shape.length;
const xRank = x.shape.length;
if (batchDims !== 0) {
if (batchDims < -indicesRank || batchDims > indicesRank) {
throw new Error(`Expect batchDims in the range of [-${indicesRank}, ${indicesRank}], but got ${batchDims}`);
}
}
if (batchDims < 0) {
batchDims += indicesRank;
}
if (batchDims > xRank) {
throw new Error(`batchDims (${batchDims}) must be less than rank(x) (
${xRank}).`);
}
if (axis < batchDims) {
throw new Error(`batchDims (${batchDims}) must be less than or equal to axis (${axis}).`);
}
for (let i = 0; i < batchDims; ++i) {
if (x.shape[i] !== indices.shape[i]) {
throw new Error(`x.shape[${i}]: ${x.shape[i]} should be equal to indices.shape[${i}]: ${indices.shape[i]}.`);
}
}
const dimSize = x.shape[axis];
const outputShape = [];
let batchSize = 1;
let outerSize = 1;
let sliceSize = 1;
for (let i = 0; i < batchDims; ++i) {
outputShape.push(x.shape[i]);
batchSize *= x.shape[i];
}
for (let i = batchDims; i < axis; i++) {
outputShape.push(x.shape[i]);
outerSize *= x.shape[i];
}
for (let i = batchDims; i < indicesRank; i++) {
outputShape.push(indices.shape[i]);
}
for (let i = axis + 1; i < xRank; i++) {
outputShape.push(x.shape[i]);
sliceSize *= x.shape[i];
}
return { batchSize, sliceSize, outerSize, dimSize, outputShape };
}
var segment_util = /*#__PURE__*/Object.freeze({
__proto__: null,
collectGatherOpShapeInfo: collectGatherOpShapeInfo,
computeOutShape: computeOutShape,
segOpComputeOptimalWindowSize: segOpComputeOptimalWindowSize
});
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fromUint8ToStringArray(vals) {
try {
// Decode the bytes into string.
return vals.map(val => decodeString(val));
}
catch (err) {
throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${err}`);
}
}
function fromStringArrayToUint8(strings) {
return strings.map(s => encodeString(s));
}
var backend_util = /*#__PURE__*/Object.freeze({
__proto__: null,
ERF_A1: ERF_A1,
ERF_A2: ERF_A2,
ERF_A3: ERF_A3,
ERF_A4: ERF_A4,
ERF_A5: ERF_A5,
ERF_P: ERF_P,
PARALLELIZE_THRESHOLD: PARALLELIZE_THRESHOLD,
get RowPartitionType () { return RowPartitionType$1; },
SELU_SCALE: SELU_SCALE,
SELU_SCALEALPHA: SELU_SCALEALPHA,
applyActivation: applyActivation$1,
assertAndGetBroadcastShape: assertAndGetBroadcastShape,
assertAxesAreInnerMostDims: assertAxesAreInnerMostDims,
assertParamsConsistent: assertParamsConsistent,
assignToTypedArray: assignToTypedArray,
axesAreInnerMostDims: axesAreInnerMostDims,
calculateShapes: calculateShapes,
checkEinsumDimSizes: checkEinsumDimSizes,
checkPadOnDimRoundingMode: checkPadOnDimRoundingMode,
combineLocations: combineLocations,
combineRaggedTensorToTensorShapes: combineRaggedTensorToTensorShapes,
complexWithEvenIndex: complexWithEvenIndex,
complexWithOddIndex: complexWithOddIndex,
computeConv2DInfo: computeConv2DInfo,
computeConv3DInfo: computeConv3DInfo,
computeDefaultPad: computeDefaultPad,
computeDilation2DInfo: computeDilation2DInfo,
computeOptimalWindowSize: computeOptimalWindowSize,
computeOutAndReduceShapes: computeOutAndReduceShapes,
computeOutShape: computeOutShape$1,
computePool2DInfo: computePool2DInfo,
computePool3DInfo: computePool3DInfo,
convertConv2DDataFormat: convertConv2DDataFormat,
decodeEinsumEquation: decodeEinsumEquation,
eitherStridesOrDilationsAreOne: eitherStridesOrDilationsAreOne,
expandShapeToKeepDim: expandShapeToKeepDim,
exponent: exponent,
exponents: exponents,
fromStringArrayToUint8: fromStringArrayToUint8,
fromUint8ToStringArray: fromUint8ToStringArray,
getAxesPermutation: getAxesPermutation,
getBroadcastDims: getBroadcastDims$1,
getComplexWithIndex: getComplexWithIndex,
getEinsumComputePath: getEinsumComputePath,
getEinsumPermutation: getEinsumPermutation,
getFusedBiasGradient: getFusedBiasGradient,
getFusedDyActivation: getFusedDyActivation,
getImageCenter: getImageCenter,
getInnerMostAxes: getInnerMostAxes,
getPermuted: getPermuted,
getRaggedRank: getRaggedRank,
getReductionAxes: getReductionAxes,
getReshaped: getReshaped,
getReshapedPermuted: getReshapedPermuted,
getRowPartitionTypesHelper: getRowPartitionTypesHelper,
getSliceBeginCoords: getSliceBeginCoords,
getSliceSize: getSliceSize,
getSparseFillEmptyRowsIndicesDenseShapeMismatch: getSparseFillEmptyRowsIndicesDenseShapeMismatch,
getSparseFillEmptyRowsNegativeIndexErrorMessage: getSparseFillEmptyRowsNegativeIndexErrorMessage,
getSparseFillEmptyRowsOutOfRangeIndexErrorMessage: getSparseFillEmptyRowsOutOfRangeIndexErrorMessage,
getSparseReshapeEmptyTensorZeroOutputDimErrorMessage: getSparseReshapeEmptyTensorZeroOutputDimErrorMessage,
getSparseReshapeInputOutputMismatchErrorMessage: getSparseReshapeInputOutputMismatchErrorMessage,
getSparseReshapeInputOutputMultipleErrorMessage: getSparseReshapeInputOutputMultipleErrorMessage,
getSparseReshapeMultipleNegativeOneOutputDimErrorMessage: getSparseReshapeMultipleNegativeOneOutputDimErrorMessage,
getSparseReshapeNegativeOutputDimErrorMessage: getSparseReshapeNegativeOutputDimErrorMessage,
getSparseSegmentReductionIndicesOutOfRangeErrorMessage: getSparseSegmentReductionIndicesOutOfRangeErrorMessage,
getSparseSegmentReductionNegativeSegmentIdsErrorMessage: getSparseSegmentReductionNegativeSegmentIdsErrorMessage,
getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage: getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage,
getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage: getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage,
getUndoAxesPermutation: getUndoAxesPermutation,
isIdentityPermutation: isIdentityPermutation,
log: log$3,
mergeRealAndImagArrays: mergeRealAndImagArrays,
prepareAndValidate: prepareAndValidate,
prepareSplitSize: prepareSplitSize,
segment_util: segment_util,
shouldFuse: shouldFuse,
slice_util: slice_util,
splitRealAndImagArrays: splitRealAndImagArrays,
stridesOrDilationsArePositive: stridesOrDilationsArePositive,
tupleValuesAreOne: tupleValuesAreOne,
upcastType: upcastType,
validateDefaultValueShape: validateDefaultValueShape,
validateInput: validateInput,
validateUpdateShape: validateUpdateShape,
warn: warn
});
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Required side effectful code.
registerOptimizers();
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const contexts = {};
const WEBGL_ATTRIBUTES = {
alpha: false,
antialias: false,
premultipliedAlpha: false,
preserveDrawingBuffer: false,
depth: false,
stencil: false,
failIfMajorPerformanceCaveat: true
};
function setWebGLContext(webGLVersion, gl) {
contexts[webGLVersion] = gl;
}
function getWebGLContext(webGLVersion, customCanvas) {
if (!(webGLVersion in contexts) || customCanvas != null) {
const newCtx = getWebGLRenderingContext(webGLVersion, customCanvas);
if (newCtx !== null) {
contexts[webGLVersion] = newCtx;
}
else {
console.log('Could not get context for WebGL version', webGLVersion);
return null;
}
}
const gl = contexts[webGLVersion];
if (gl == null || gl.isContextLost()) {
delete contexts[webGLVersion];
return getWebGLContext(webGLVersion);
}
gl.disable(gl.DEPTH_TEST);
gl.disable(gl.STENCIL_TEST);
gl.disable(gl.BLEND);
gl.disable(gl.DITHER);
gl.disable(gl.POLYGON_OFFSET_FILL);
gl.disable(gl.SAMPLE_COVERAGE);
gl.enable(gl.SCISSOR_TEST);
gl.enable(gl.CULL_FACE);
gl.cullFace(gl.BACK);
return contexts[webGLVersion];
}
function createCanvas(webGLVersion) {
// Use canvas element for Safari, since its offscreen canvas does not support
// fencing.
if (!env().getBool('IS_SAFARI') && typeof OffscreenCanvas !== 'undefined' &&
webGLVersion === 2) {
return new OffscreenCanvas(300, 150);
}
else if (typeof document !== 'undefined') {
return document.createElement('canvas');
}
else {
throw new Error('Cannot create a canvas in this context');
}
}
function getWebGLRenderingContext(webGLVersion, customCanvas) {
if (webGLVersion !== 1 && webGLVersion !== 2) {
throw new Error('Cannot get WebGL rendering context, WebGL is disabled.');
}
const canvas = customCanvas == null ? createCanvas(webGLVersion) : customCanvas;
canvas.addEventListener('webglcontextlost', (ev) => {
ev.preventDefault();
delete contexts[webGLVersion];
}, false);
if (env().getBool('SOFTWARE_WEBGL_ENABLED')) {
WEBGL_ATTRIBUTES.failIfMajorPerformanceCaveat = false;
}
if (webGLVersion === 1) {
return (
// tslint:disable-next-line
canvas.getContext('webgl', WEBGL_ATTRIBUTES) ||
canvas
.getContext('experimental-webgl', WEBGL_ATTRIBUTES));
}
return canvas.getContext('webgl2', WEBGL_ATTRIBUTES);
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
var PackingScheme;
(function (PackingScheme) {
/**
* All values in a single texel are densely packed without any constraints.
*
* This is how the shader encodes a tensor with shape = [2, 3, 4]
* (indices are [batch, row, col]).
*
* 000|001 010|011 020|021
* ------- ------- -------
* 002|003 012|013 022|023
*
* 100|101 110|111 120|121
* ------- ------- -------
* 102|103 112|113 122|123
*
*/
PackingScheme[PackingScheme["DENSE"] = 0] = "DENSE";
/**
* Single texels contain only values from the same batch, and from adjacent
* rows and columns.
*
* This is how the shader encodes a tensor with shape = [2, 3, 5]
* (indices are [batch, row, col]).
*
* 000|001 002|003 004|xxx 020|021 022|023 024|xxx
* ------- ------- ------- ------- ------- -------
* 010|011 012|013 014|xxx xxx|xxx xxx|xxx xxx|xxx
*
* 100|101 102|103 104|xxx 120|121 122|123 124|xxx
* ------- ------- ------- ------- ------- -------
* 110|111 112|113 114|xxx xxx|xxx xxx|xxx xxx|xxx
*
*/
PackingScheme[PackingScheme["SHARED_BATCH"] = 1] = "SHARED_BATCH";
})(PackingScheme || (PackingScheme = {}));
var TextureUsage;
(function (TextureUsage) {
TextureUsage[TextureUsage["RENDER"] = 0] = "RENDER";
TextureUsage[TextureUsage["UPLOAD"] = 1] = "UPLOAD";
TextureUsage[TextureUsage["PIXELS"] = 2] = "PIXELS";
TextureUsage[TextureUsage["DOWNLOAD"] = 3] = "DOWNLOAD";
})(TextureUsage || (TextureUsage = {}));
var PhysicalTextureType;
(function (PhysicalTextureType) {
PhysicalTextureType[PhysicalTextureType["UNPACKED_FLOAT16"] = 0] = "UNPACKED_FLOAT16";
PhysicalTextureType[PhysicalTextureType["UNPACKED_FLOAT32"] = 1] = "UNPACKED_FLOAT32";
PhysicalTextureType[PhysicalTextureType["PACKED_4X1_UNSIGNED_BYTE"] = 2] = "PACKED_4X1_UNSIGNED_BYTE";
PhysicalTextureType[PhysicalTextureType["PACKED_2X2_FLOAT32"] = 3] = "PACKED_2X2_FLOAT32";
PhysicalTextureType[PhysicalTextureType["PACKED_2X2_FLOAT16"] = 4] = "PACKED_2X2_FLOAT16";
})(PhysicalTextureType || (PhysicalTextureType = {}));
function getUnpackedMatrixTextureShapeWidthHeight(rows, columns) {
return [columns, rows];
}
function getUnpackedArraySizeFromMatrixSize(matrixSize, channelsPerTexture) {
return matrixSize * channelsPerTexture;
}
/**
* Get shape for densely packed RGBA texture.
*/
function getDenseTexShape(shape) {
const size = sizeFromShape(shape);
const texelsNeeded = Math.ceil(size / 4);
return sizeToSquarishShape(texelsNeeded);
}
function getPackedMatrixTextureShapeWidthHeight(rows, columns) {
return [
Math.max(1, Math.ceil(columns / 2)), Math.max(1, Math.ceil(rows / 2))
];
}
function getPackedRGBAArraySizeFromMatrixShape(rows, columns) {
const [w, h] = getPackedMatrixTextureShapeWidthHeight(rows, columns);
return w * h * 4;
}
function getTextureConfig(
// tslint:disable-next-line:no-any
gl, textureHalfFloatExtension) {
// tslint:disable-next-line:no-any
const glany = gl;
let internalFormatFloat;
let internalFormatHalfFloat;
let internalFormatPackedHalfFloat;
let internalFormatPackedFloat;
let textureFormatFloat;
let downloadTextureFormat;
let downloadUnpackNumChannels;
let defaultNumChannels;
let textureTypeHalfFloat;
let textureTypeFloat;
if (env().getNumber('WEBGL_VERSION') === 2) {
internalFormatFloat = glany.R32F;
internalFormatHalfFloat = glany.R16F;
internalFormatPackedHalfFloat = glany.RGBA16F;
internalFormatPackedFloat = glany.RGBA32F;
textureFormatFloat = glany.RED;
downloadUnpackNumChannels = 4;
defaultNumChannels = 1;
textureTypeHalfFloat = glany.HALF_FLOAT;
textureTypeFloat = glany.FLOAT;
downloadTextureFormat = glany.RGBA8;
}
else {
internalFormatFloat = gl.RGBA;
internalFormatHalfFloat = gl.RGBA;
internalFormatPackedHalfFloat = gl.RGBA;
internalFormatPackedFloat = glany.RGBA;
textureFormatFloat = gl.RGBA;
downloadUnpackNumChannels = 4;
defaultNumChannels = 4;
textureTypeHalfFloat = textureHalfFloatExtension != null ?
textureHalfFloatExtension.HALF_FLOAT_OES :
null;
textureTypeFloat = gl.FLOAT;
downloadTextureFormat = gl.RGBA;
}
return {
internalFormatFloat,
internalFormatHalfFloat,
internalFormatPackedHalfFloat,
internalFormatPackedFloat,
textureFormatFloat,
downloadTextureFormat,
downloadUnpackNumChannels,
defaultNumChannels,
textureTypeHalfFloat,
textureTypeFloat
};
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function callAndCheck(gl, func) {
const returnValue = func();
if (env().getBool('DEBUG')) {
checkWebGLError(gl);
}
return returnValue;
}
function checkWebGLError(gl) {
const error = gl.getError();
if (error !== gl.NO_ERROR) {
throw new Error('WebGL Error: ' + getWebGLErrorMessage(gl, error));
}
}
// https://en.wikipedia.org/wiki/Half-precision_floating-point_format
const MIN_FLOAT16 = 5.96e-8;
const MAX_FLOAT16 = 65504;
function canBeRepresented(num) {
if (env().getBool('WEBGL_RENDER_FLOAT32_ENABLED') || num === 0 ||
(MIN_FLOAT16 < Math.abs(num) && Math.abs(num) < MAX_FLOAT16)) {
return true;
}
return false;
}
function getWebGLErrorMessage(gl, status) {
switch (status) {
case gl.NO_ERROR:
return 'NO_ERROR';
case gl.INVALID_ENUM:
return 'INVALID_ENUM';
case gl.INVALID_VALUE:
return 'INVALID_VALUE';
case gl.INVALID_OPERATION:
return 'INVALID_OPERATION';
case gl.INVALID_FRAMEBUFFER_OPERATION:
return 'INVALID_FRAMEBUFFER_OPERATION';
case gl.OUT_OF_MEMORY:
return 'OUT_OF_MEMORY';
case gl.CONTEXT_LOST_WEBGL:
return 'CONTEXT_LOST_WEBGL';
default:
return `Unknown error code ${status}`;
}
}
function getExtensionOrThrow(gl, extensionName) {
return throwIfNull(gl, () => gl.getExtension(extensionName), 'Extension "' + extensionName + '" not supported on this browser.');
}
function createVertexShader$1(gl, vertexShaderSource) {
const vertexShader = throwIfNull(gl, () => gl.createShader(gl.VERTEX_SHADER), 'Unable to create vertex WebGLShader.');
callAndCheck(gl, () => gl.shaderSource(vertexShader, vertexShaderSource));
callAndCheck(gl, () => gl.compileShader(vertexShader));
if (gl.getShaderParameter(vertexShader, gl.COMPILE_STATUS) === false) {
console.log(gl.getShaderInfoLog(vertexShader));
throw new Error('Failed to compile vertex shader.');
}
return vertexShader;
}
function createFragmentShader(gl, fragmentShaderSource) {
const fragmentShader = throwIfNull(gl, () => gl.createShader(gl.FRAGMENT_SHADER), 'Unable to create fragment WebGLShader.');
callAndCheck(gl, () => gl.shaderSource(fragmentShader, fragmentShaderSource));
callAndCheck(gl, () => gl.compileShader(fragmentShader));
if (env().get('ENGINE_COMPILE_ONLY')) {
return fragmentShader;
}
if (gl.getShaderParameter(fragmentShader, gl.COMPILE_STATUS) === false) {
logShaderSourceAndInfoLog(fragmentShaderSource, gl.getShaderInfoLog(fragmentShader));
throw new Error('Failed to compile fragment shader.');
}
return fragmentShader;
}
const lineNumberRegex = /ERROR: [0-9]+:([0-9]+):/g;
function logShaderSourceAndInfoLog(shaderSource, shaderInfoLog) {
const lineNumberRegexResult = lineNumberRegex.exec(shaderInfoLog);
if (lineNumberRegexResult == null) {
console.log(`Couldn't parse line number in error: ${shaderInfoLog}`);
console.log(shaderSource);
return;
}
const lineNumber = +lineNumberRegexResult[1];
const shaderLines = shaderSource.split('\n');
const pad = shaderLines.length.toString().length + 2;
const linesWithLineNumbers = shaderLines.map((line, lineNumber) => rightPad((lineNumber + 1).toString(), pad) + line);
let maxLineLength = 0;
for (let i = 0; i < linesWithLineNumbers.length; i++) {
maxLineLength = Math.max(linesWithLineNumbers[i].length, maxLineLength);
}
const beforeErrorLines = linesWithLineNumbers.slice(0, lineNumber - 1);
const errorLine = linesWithLineNumbers.slice(lineNumber - 1, lineNumber);
const afterErrorLines = linesWithLineNumbers.slice(lineNumber);
console.log(beforeErrorLines.join('\n'));
console.log(shaderInfoLog.split('\n')[0]);
console.log(`%c ${rightPad(errorLine[0], maxLineLength)}`, 'border:1px solid red; background-color:#e3d2d2; color:#a61717');
console.log(afterErrorLines.join('\n'));
}
function createProgram(gl) {
return throwIfNull(gl, () => gl.createProgram(), 'Unable to create WebGLProgram.');
}
function linkProgram(gl, program) {
callAndCheck(gl, () => gl.linkProgram(program));
if (env().get('ENGINE_COMPILE_ONLY')) {
return;
}
if (gl.getProgramParameter(program, gl.LINK_STATUS) === false) {
console.log(gl.getProgramInfoLog(program));
throw new Error('Failed to link vertex and fragment shaders.');
}
}
/// validateProgram is effectively "If we `useProgram(program); drawArrays();`,
/// give feedback in log about perf/correctness warnings or errors that would
/// occur."
/// So make sure we set up all vertex/texture/sampler/uniform data before
/// calling validateProgram!
function validateProgram(gl, program) {
callAndCheck(gl, () => gl.validateProgram(program));
if (gl.getProgramParameter(program, gl.VALIDATE_STATUS) === false) {
console.log(gl.getProgramInfoLog(program));
throw new Error('Shader program validation failed.');
}
}
function createStaticVertexBuffer(gl, data) {
const buffer = throwIfNull(gl, () => gl.createBuffer(), 'Unable to create WebGLBuffer');
callAndCheck(gl, () => gl.bindBuffer(gl.ARRAY_BUFFER, buffer));
callAndCheck(gl, () => gl.bufferData(gl.ARRAY_BUFFER, data, gl.STATIC_DRAW));
return buffer;
}
function createStaticIndexBuffer(gl, data) {
const buffer = throwIfNull(gl, () => gl.createBuffer(), 'Unable to create WebGLBuffer');
callAndCheck(gl, () => gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, buffer));
callAndCheck(gl, () => gl.bufferData(gl.ELEMENT_ARRAY_BUFFER, data, gl.STATIC_DRAW));
return buffer;
}
function createTexture(gl) {
return throwIfNull(gl, () => gl.createTexture(), 'Unable to create WebGLTexture.');
}
function validateTextureSize(width, height) {
const maxTextureSize = env().getNumber('WEBGL_MAX_TEXTURE_SIZE');
if ((width <= 0) || (height <= 0)) {
const requested = `[${width}x${height}]`;
throw new Error('Requested texture size ' + requested + ' is invalid.');
}
if ((width > maxTextureSize) || (height > maxTextureSize)) {
const requested = `[${width}x${height}]`;
const max = `[${maxTextureSize}x${maxTextureSize}]`;
throw new Error('Requested texture size ' + requested +
' greater than WebGL maximum on this browser / GPU ' + max + '.');
}
}
function createFramebuffer(gl) {
return throwIfNull(gl, () => gl.createFramebuffer(), 'Unable to create WebGLFramebuffer.');
}
function bindVertexBufferToProgramAttribute(gl, program, attribute, buffer, arrayEntriesPerItem, itemStrideInBytes, itemOffsetInBytes) {
const loc = gl.getAttribLocation(program, attribute);
if (loc === -1) {
// The GPU compiler decided to strip out this attribute because it's unused,
// thus no need to bind.
return false;
}
callAndCheck(gl, () => gl.bindBuffer(gl.ARRAY_BUFFER, buffer));
callAndCheck(gl, () => gl.vertexAttribPointer(loc, arrayEntriesPerItem, gl.FLOAT, false, itemStrideInBytes, itemOffsetInBytes));
callAndCheck(gl, () => gl.enableVertexAttribArray(loc));
return true;
}
function bindTextureUnit(gl, texture, textureUnit) {
validateTextureUnit(gl, textureUnit);
callAndCheck(gl, () => gl.activeTexture(gl.TEXTURE0 + textureUnit));
callAndCheck(gl, () => gl.bindTexture(gl.TEXTURE_2D, texture));
}
function getProgramUniformLocationOrThrow(gl, program, uniformName) {
return throwIfNull(gl, () => gl.getUniformLocation(program, uniformName), 'uniform "' + uniformName + '" not present in program.');
}
function getProgramUniformLocation(gl, program, uniformName) {
return gl.getUniformLocation(program, uniformName);
}
function bindTextureToProgramUniformSampler(gl, texture, uniformSamplerLocation, textureUnit) {
callAndCheck(gl, () => bindTextureUnit(gl, texture, textureUnit));
callAndCheck(gl, () => gl.uniform1i(uniformSamplerLocation, textureUnit));
}
function bindColorTextureToFramebuffer(gl, texture, framebuffer) {
callAndCheck(gl, () => gl.bindFramebuffer(gl.FRAMEBUFFER, framebuffer));
callAndCheck(gl, () => gl.framebufferTexture2D(gl.FRAMEBUFFER, gl.COLOR_ATTACHMENT0, gl.TEXTURE_2D, texture, 0));
}
function unbindColorTextureFromFramebuffer(gl, framebuffer) {
callAndCheck(gl, () => gl.bindFramebuffer(gl.FRAMEBUFFER, framebuffer));
callAndCheck(gl, () => gl.framebufferTexture2D(gl.FRAMEBUFFER, gl.COLOR_ATTACHMENT0, gl.TEXTURE_2D, null, 0));
}
function validateFramebuffer(gl) {
const status = gl.checkFramebufferStatus(gl.FRAMEBUFFER);
if (status !== gl.FRAMEBUFFER_COMPLETE) {
throw new Error('Error binding framebuffer: ' + getFramebufferErrorMessage(gl, status));
}
}
function getFramebufferErrorMessage(gl, status) {
switch (status) {
case gl.FRAMEBUFFER_INCOMPLETE_ATTACHMENT:
return 'FRAMEBUFFER_INCOMPLETE_ATTACHMENT';
case gl.FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT:
return 'FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT';
case gl.FRAMEBUFFER_INCOMPLETE_DIMENSIONS:
return 'FRAMEBUFFER_INCOMPLETE_DIMENSIONS';
case gl.FRAMEBUFFER_UNSUPPORTED:
return 'FRAMEBUFFER_UNSUPPORTED';
default:
return `unknown error ${status}`;
}
}
function throwIfNull(gl, returnTOrNull, failureMessage) {
const tOrNull = callAndCheck(gl, () => returnTOrNull());
if (tOrNull == null) {
throw new Error(failureMessage);
}
return tOrNull;
}
function validateTextureUnit(gl, textureUnit) {
const maxTextureUnit = gl.MAX_COMBINED_TEXTURE_IMAGE_UNITS - 1;
const glTextureUnit = textureUnit + gl.TEXTURE0;
if (glTextureUnit < gl.TEXTURE0 || glTextureUnit > maxTextureUnit) {
const textureUnitRange = `[gl.TEXTURE0, gl.TEXTURE${maxTextureUnit}]`;
throw new Error(`textureUnit must be in ${textureUnitRange}.`);
}
}
function getBatchDim(shape, dimsToSkip = 2) {
return sizeFromShape(shape.slice(0, shape.length - dimsToSkip));
}
function getRowsCols(shape) {
if (shape.length === 0) {
throw Error('Cannot get rows and columns of an empty shape array.');
}
return [
shape.length > 1 ? shape[shape.length - 2] : 1, shape[shape.length - 1]
];
}
function getShapeAs3D(shape) {
let shapeAs3D = [1, 1, 1];
const isScalar = shape.length === 0 || (shape.length === 1 && shape[0] === 1);
if (!isScalar) {
shapeAs3D =
[getBatchDim(shape), ...getRowsCols(shape)];
}
return shapeAs3D;
}
function getTextureShapeFromLogicalShape(logShape, isPacked = false) {
let maxTexSize = env().getNumber('WEBGL_MAX_TEXTURE_SIZE');
let maxSizeForNarrowTex = env().getNumber('WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE');
if (maxSizeForNarrowTex === Infinity &&
env().getBool('WEBGL_AUTO_SQUARIFY_NARROW_TEXTURE_SHAPE')) {
maxSizeForNarrowTex = maxTexSize / 2;
}
if (isPacked) {
maxTexSize = maxTexSize * 2;
maxSizeForNarrowTex = maxSizeForNarrowTex * 2;
// This logic ensures we accurately count the number of packed texels needed
// to accommodate the tensor. We can only pack values in the same texel if
// they are from adjacent pairs of rows/cols within the same batch. So if a
// tensor has 3 rows, we pretend it has 4 rows in order to account for the
// fact that the texels containing the third row are half empty.
logShape = logShape.map((d, i) => i >= logShape.length - 2 ?
nearestLargerEven(logShape[i]) :
logShape[i]);
// Packed texture height is at least 2 (the channel height of a single
// texel).
if (logShape.length === 1) {
logShape = [2, logShape[0]];
}
}
// If logical shape is 2, we don't squeeze, since we want to match physical.
if (logShape.length !== 2) {
const squeezeResult = squeezeShape(logShape);
logShape = squeezeResult.newShape;
}
let size = sizeFromShape(logShape);
let textureShape = null;
if (logShape.length <= 1 && size <= maxTexSize) {
textureShape = [1, size];
}
else if (logShape.length === 2 && logShape[0] <= maxTexSize &&
logShape[1] <= maxTexSize) {
textureShape = logShape;
}
else if (logShape.length === 3 && logShape[0] * logShape[1] <= maxTexSize &&
logShape[2] <= maxTexSize) {
textureShape = [logShape[0] * logShape[1], logShape[2]];
}
else if (logShape.length === 3 && logShape[0] <= maxTexSize &&
logShape[1] * logShape[2] <= maxTexSize) {
textureShape = [logShape[0], logShape[1] * logShape[2]];
}
else if (logShape.length === 4 &&
logShape[0] * logShape[1] * logShape[2] <= maxTexSize &&
logShape[3] <= maxTexSize) {
textureShape = [logShape[0] * logShape[1] * logShape[2], logShape[3]];
}
else if (logShape.length === 4 && logShape[0] <= maxTexSize &&
logShape[1] * logShape[2] * logShape[3] <= maxTexSize) {
textureShape = [logShape[0], logShape[1] * logShape[2] * logShape[3]];
}
// true if one edge length is 1 (1 or 2, if packed), while another edge
// length exceeds maxSizeForNarrowTex.
const isLongNarrowTex = textureShape != null &&
Math.max(...textureShape) > maxSizeForNarrowTex &&
Math.min(...textureShape) <= (isPacked ? 2 : 1) &&
Math.min(...textureShape) > 0;
if (textureShape == null || isLongNarrowTex) {
if (isPacked) {
// For packed textures size equals the number of channels required to
// accommodate the texture data. However in order to squarify such that
// inner dimensions stay even, we rewrite size to equal the number of
// texels. Then in the return statement we rehydrate the squarified
// dimensions to channel units.
const batchDim = getBatchDim(logShape);
let rows = 2, cols = 2;
if (logShape.length) {
[rows, cols] = getRowsCols(logShape);
}
size = batchDim * (rows / 2) * (cols / 2);
textureShape =
sizeToSquarishShape(size).map(d => d * 2);
}
else {
textureShape = sizeToSquarishShape(size);
}
}
return textureShape;
}
function isEven(n) {
return n % 2 === 0;
}
/**
* This determines whether reshaping a packed texture requires rearranging
* the data within the texture, assuming 2x2 packing.
*/
function isReshapeFree(shape1, shape2) {
shape1 = shape1.slice(-2);
shape2 = shape2.slice(-2);
if (arraysEqual(shape1, shape2)) {
return true;
}
if (!shape1.length || !shape2.length) { // One of the shapes is a scalar.
return true;
}
if (shape1[0] === 0 || shape1[1] === 0 || shape2[0] === 0 ||
shape2[1] === 0) {
return true;
}
if (shape1.length !== shape2.length) { // One of the shapes is a vector.
const shape1Cols = shape1[shape1.length - 1];
const shape2Cols = shape2[shape2.length - 1];
if (shape1Cols === shape2Cols) {
return true;
}
if (isEven(shape1Cols) && isEven(shape2Cols) &&
(shape1[0] === 1 || shape2[0] === 1)) {
return true;
}
}
return shape1[1] === shape2[1] && isEven(shape1[0]) && isEven(shape2[0]);
}
// We cache webgl params because the environment gets reset between
// unit tests and we don't want to constantly query the WebGLContext for
// MAX_TEXTURE_SIZE.
let MAX_TEXTURE_SIZE;
let MAX_TEXTURES_IN_SHADER;
function getWebGLMaxTextureSize(webGLVersion) {
if (MAX_TEXTURE_SIZE == null) {
const gl = getWebGLContext(webGLVersion);
MAX_TEXTURE_SIZE = gl.getParameter(gl.MAX_TEXTURE_SIZE);
}
return MAX_TEXTURE_SIZE;
}
function getMaxTexturesInShader(webGLVersion) {
if (MAX_TEXTURES_IN_SHADER == null) {
const gl = getWebGLContext(webGLVersion);
MAX_TEXTURES_IN_SHADER = gl.getParameter(gl.MAX_TEXTURE_IMAGE_UNITS);
}
// We cap at 16 to avoid spurious runtime "memory exhausted" error.
return Math.min(16, MAX_TEXTURES_IN_SHADER);
}
function getWebGLDisjointQueryTimerVersion(webGLVersion) {
if (webGLVersion === 0) {
return 0;
}
let queryTimerVersion;
const gl = getWebGLContext(webGLVersion);
if (hasExtension(gl, 'EXT_disjoint_timer_query_webgl2') &&
webGLVersion === 2) {
queryTimerVersion = 2;
}
else if (hasExtension(gl, 'EXT_disjoint_timer_query')) {
queryTimerVersion = 1;
}
else {
queryTimerVersion = 0;
}
return queryTimerVersion;
}
function hasExtension(gl, extensionName) {
const ext = gl.getExtension(extensionName);
return ext != null;
}
function isWebGLVersionEnabled(webGLVersion) {
try {
const gl = getWebGLContext(webGLVersion);
if (gl != null) {
return true;
}
}
catch (e) {
console.log('Error when getting WebGL context: ', e);
return false;
}
return false;
}
function isCapableOfRenderingToFloatTexture(webGLVersion) {
if (webGLVersion === 0) {
return false;
}
const gl = getWebGLContext(webGLVersion);
if (webGLVersion === 1) {
if (!hasExtension(gl, 'OES_texture_float')) {
return false;
}
}
else {
if (!hasExtension(gl, 'EXT_color_buffer_float')) {
return false;
}
}
const isFrameBufferComplete = createFloatTextureAndBindToFramebuffer(gl);
return isFrameBufferComplete;
}
/**
* Check if we can download values from a float/half-float texture.
*
* Note that for performance reasons we use binding a texture to a framebuffer
* as a proxy for ability to download float values later using readPixels. The
* texture params of this texture will not match those in readPixels exactly
* but if we are unable to bind some kind of float texture to the frameBuffer
* then we definitely will not be able to read float values from it.
*/
function isDownloadFloatTextureEnabled(webGLVersion) {
if (webGLVersion === 0) {
return false;
}
const gl = getWebGLContext(webGLVersion);
if (webGLVersion === 1) {
if (!hasExtension(gl, 'OES_texture_float')) {
return false;
}
if (!hasExtension(gl, 'WEBGL_color_buffer_float')) {
return false;
}
}
else {
if (hasExtension(gl, 'EXT_color_buffer_float')) {
return createFloatTextureAndBindToFramebuffer(gl);
}
const COLOR_BUFFER_HALF_FLOAT = 'EXT_color_buffer_half_float';
if (hasExtension(gl, COLOR_BUFFER_HALF_FLOAT)) {
const textureHalfFloatExtension = gl.getExtension(COLOR_BUFFER_HALF_FLOAT);
return createHalfFloatTextureAndBindToFramebuffer(gl, textureHalfFloatExtension);
}
return false;
}
const isFrameBufferComplete = createFloatTextureAndBindToFramebuffer(gl);
return isFrameBufferComplete;
}
function createFloatTextureAndBindToFramebuffer(gl) {
const texConfig = getTextureConfig(gl);
const texture = gl.createTexture();
gl.bindTexture(gl.TEXTURE_2D, texture);
const width = 1;
const height = 1;
gl.texImage2D(gl.TEXTURE_2D, 0, texConfig.internalFormatFloat, width, height, 0, texConfig.textureFormatFloat, texConfig.textureTypeFloat, null);
const frameBuffer = gl.createFramebuffer();
gl.bindFramebuffer(gl.FRAMEBUFFER, frameBuffer);
gl.framebufferTexture2D(gl.FRAMEBUFFER, gl.COLOR_ATTACHMENT0, gl.TEXTURE_2D, texture, 0);
const isFrameBufferComplete = gl.checkFramebufferStatus(gl.FRAMEBUFFER) === gl.FRAMEBUFFER_COMPLETE;
gl.bindTexture(gl.TEXTURE_2D, null);
gl.bindFramebuffer(gl.FRAMEBUFFER, null);
gl.deleteTexture(texture);
gl.deleteFramebuffer(frameBuffer);
return isFrameBufferComplete;
}
function createHalfFloatTextureAndBindToFramebuffer(
// tslint:disable-next-line:no-any
gl, textureHalfFloatExtension) {
const texConfig = getTextureConfig(gl, textureHalfFloatExtension);
const texture = gl.createTexture();
gl.bindTexture(gl.TEXTURE_2D, texture);
const width = 1;
const height = 1;
gl.texImage2D(gl.TEXTURE_2D, 0, texConfig.internalFormatHalfFloat, width, height, 0, texConfig.textureFormatFloat, texConfig.textureTypeHalfFloat, null);
const frameBuffer = gl.createFramebuffer();
gl.bindFramebuffer(gl.FRAMEBUFFER, frameBuffer);
gl.framebufferTexture2D(gl.FRAMEBUFFER, gl.COLOR_ATTACHMENT0, gl.TEXTURE_2D, texture, 0);
const isFrameBufferComplete = gl.checkFramebufferStatus(gl.FRAMEBUFFER) === gl.FRAMEBUFFER_COMPLETE;
gl.bindTexture(gl.TEXTURE_2D, null);
gl.bindFramebuffer(gl.FRAMEBUFFER, null);
gl.deleteTexture(texture);
gl.deleteFramebuffer(frameBuffer);
return isFrameBufferComplete;
}
function isWebGLFenceEnabled(webGLVersion) {
if (webGLVersion !== 2) {
return false;
}
const gl = getWebGLContext(webGLVersion);
// tslint:disable-next-line:no-any
const isEnabled = gl.fenceSync != null;
return isEnabled;
}
function assertNotComplex$1(tensor, opName) {
if (!Array.isArray(tensor)) {
tensor = [tensor];
}
tensor.forEach(t => {
if (t != null) {
assert$1(t.dtype !== 'complex64', () => `${opName} does not support complex64 tensors ` +
'in the WebGL backend.');
}
});
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ENV = env();
/**
* This file contains WebGL-specific flag registrations.
*/
/**
* True if WebGL is supported.
*/
ENV.registerFlag('HAS_WEBGL', () => ENV.getNumber('WEBGL_VERSION') > 0);
/** 0: No WebGL, 1: WebGL 1.0, 2: WebGL 2.0. */
ENV.registerFlag('WEBGL_VERSION', () => {
if (isWebGLVersionEnabled(2)) {
return 2;
}
else if (isWebGLVersionEnabled(1)) {
return 1;
}
return 0;
});
/** Whether to check for numerical representation problems. */
ENV.registerFlag('WEBGL_CHECK_NUMERICAL_PROBLEMS', () => false);
ENV.registerFlag('WEBGL_BUFFER_SUPPORTED', () => ENV.get('WEBGL_VERSION') === 2);
/** Whether the WebGL backend will sometimes forward ops to the CPU. */
ENV.registerFlag('WEBGL_CPU_FORWARD', () => true);
/** Whether the WebGL backend will always use f16 textures for rendering. */
ENV.registerFlag('WEBGL_FORCE_F16_TEXTURES', () => false);
/** Whether to turn all packing related flags on. */
ENV.registerFlag('WEBGL_PACK', () => ENV.getBool('HAS_WEBGL'));
/** Whether we will pack the batchnormalization op. */
ENV.registerFlag('WEBGL_PACK_NORMALIZATION', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack the clip op. */
ENV.registerFlag('WEBGL_PACK_CLIP', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack the depthwise conv op. */
ENV.registerFlag('WEBGL_PACK_DEPTHWISECONV', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack binary ops. */
ENV.registerFlag('WEBGL_PACK_BINARY_OPERATIONS', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack unary ops. */
ENV.registerFlag('WEBGL_PACK_UNARY_OPERATIONS', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack array ops. */
ENV.registerFlag('WEBGL_PACK_ARRAY_OPERATIONS', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack image ops. */
ENV.registerFlag('WEBGL_PACK_IMAGE_OPERATIONS', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack reduce ops. */
ENV.registerFlag('WEBGL_PACK_REDUCE', () => ENV.getBool('WEBGL_PACK'));
/** Whether packed WebGL kernels lazily unpack their outputs. */
ENV.registerFlag('WEBGL_LAZILY_UNPACK', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will use the im2col algorithm to speed up convolutions. */
ENV.registerFlag('WEBGL_CONV_IM2COL', () => ENV.getBool('WEBGL_PACK'));
/** Whether we will pack conv2dTranspose op. */
ENV.registerFlag('WEBGL_PACK_CONV2DTRANSPOSE', () => ENV.getBool('WEBGL_PACK'));
/** The maximum texture dimension. */
ENV.registerFlag('WEBGL_MAX_TEXTURE_SIZE', () => getWebGLMaxTextureSize(ENV.getNumber('WEBGL_VERSION')));
/** The maximum texture dimension. */
ENV.registerFlag('WEBGL_MAX_TEXTURES_IN_SHADER', () => getMaxTexturesInShader(ENV.getNumber('WEBGL_VERSION')));
/**
* The disjoint_query_timer extension version.
* 0: disabled, 1: EXT_disjoint_timer_query, 2:
* EXT_disjoint_timer_query_webgl2.
* In Firefox with WebGL 2.0,
* EXT_disjoint_timer_query_webgl2 is not available, so we must use the
* WebGL 1.0 extension.
*/
ENV.registerFlag('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION', () => {
const webGLVersion = ENV.getNumber('WEBGL_VERSION');
if (webGLVersion === 0) {
return 0;
}
return getWebGLDisjointQueryTimerVersion(webGLVersion);
});
/**
* Whether the timer object from the disjoint_query_timer extension gives
* timing information that is reliable.
*/
ENV.registerFlag('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE', () => ENV.getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION') > 0 &&
!isMobile());
/**
* Whether the device is physically capable of rendering to float32 textures.
*/
ENV.registerFlag('WEBGL_RENDER_FLOAT32_CAPABLE', () => isCapableOfRenderingToFloatTexture(ENV.getNumber('WEBGL_VERSION')));
/**
* Whether rendering to float32 textures is enabled. If disabled, renders to
* float16 textures.
*/
ENV.registerFlag('WEBGL_RENDER_FLOAT32_ENABLED', () => {
return ENV.getBool('WEBGL_FORCE_F16_TEXTURES') ?
false :
ENV.getBool('WEBGL_RENDER_FLOAT32_CAPABLE');
});
/**
* Whether downloading float textures is enabled (16 or 32 bit). If disabled,
* uses IEEE 754 encoding of the float32 values to 4 uint8 when downloading.
*/
ENV.registerFlag('WEBGL_DOWNLOAD_FLOAT_ENABLED', () => isDownloadFloatTextureEnabled(ENV.getNumber('WEBGL_VERSION')));
/** Whether the fence API is available. */
ENV.registerFlag('WEBGL_FENCE_API_ENABLED', () => isWebGLFenceEnabled(ENV.getNumber('WEBGL_VERSION')));
/**
* Tensors with size <= than this will be uploaded as uniforms, not textures.
*/
ENV.registerFlag('WEBGL_SIZE_UPLOAD_UNIFORM', () => {
// Use uniform uploads only when 32bit floats are supported. In
// 16bit
// environments there are problems with comparing a 16bit texture value
// with a 32bit uniform value.
const useUniforms = ENV.getBool('WEBGL_RENDER_FLOAT32_ENABLED');
return useUniforms ? 4 : 0;
});
/**
* If the total number of bytes allocated on the GPU is greater than this
* number, we will aggressively delete textures upon disposal with
* gl.deleteMatrixTexture, rather than making them available for reuse.
*
* Default value -1 indicates that we will never aggressively delete textures.
*/
ENV.registerFlag('WEBGL_DELETE_TEXTURE_THRESHOLD', () => {
return -1;
}, threshold => {
if (!(typeof threshold === 'number')) {
throw new Error('WEBGL_DELETE_TEXTURE_THRESHOLD must be a number but ' +
`got ${threshold}.`);
}
if (threshold < 0 && threshold !== -1) {
throw new Error(`WEBGL_DELETE_TEXTURE_THRESHOLD must be -1 (indicating never ` +
`delete) or at least 0, but got ${threshold}.`);
}
});
/**
* Trigger a manual GL command flush if the threshold of time has passed since
* previous Kernel execution. This can be useful for Andorid device where GL
* command flush are delayed un til the end of javascript task. This value is
* measured in millisecond. Typically you want to set this value to close to 1.
*
* Default value 1 for mobile chrome, and -1 for rest cases. -1 indicates that
* we will not enforce manual flush and depend on system default flush schedule.
*/
ENV.registerFlag('WEBGL_FLUSH_THRESHOLD', () => {
return isMobile() ? 1 : -1;
}, threshold => {
if (!(typeof threshold === 'number')) {
throw new Error('WEBGL_FLUSH_THRESHOLD must be a number but got ' +
`${threshold}.`);
}
if (threshold < 0 && threshold !== -1) {
throw new Error(`WEBGL_FLUSH_THRESHOLD must be -1 (indicating never ` +
`manual flush) or at least 0, but got ${threshold}.`);
}
});
/**
* Threshold for input tensor size that determines whether WebGL backend will
* delegate computation to CPU.
*
* Default value is 128.
*/
ENV.registerFlag('CPU_HANDOFF_SIZE_THRESHOLD', () => 128);
/** Whether we will use shapes uniforms. */
ENV.registerFlag('WEBGL_USE_SHAPES_UNIFORMS', () => false);
/**
* Threshold for last dimension of input tensor that determines whether
* WebGL backend for the Top K op will delegate computation to CPU. If input
* is smaller than threshold then CPU will be used
*
* Default value is 100000.
*/
ENV.registerFlag('TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD', () => 100000);
/**
* Threshold for K that determines whether
* WebGL backend for the Top K op will delegate computation to CPU. If k
* is larger than threshold then CPU will be used
*
* Default value is 128.
*/
ENV.registerFlag('TOPK_K_CPU_HANDOFF_THRESHOLD', () => 128);
/** Whether we will use the experimental conv op. */
ENV.registerFlag('WEBGL_EXP_CONV', () => false);
/**
* If the device performance is low or if no hardware GPU is available, whether
* software WebGL will be used.
*/
ENV.registerFlag('SOFTWARE_WEBGL_ENABLED', () => ENV.getBool('IS_TEST'));
/**
* For narrow texture (physical height or physical width is 1), if the length of
* any texture edges exceed the threshold, the texture will be reshaped to be
* more squarish.
*
* This flag is used to help some GPUs that could not provide correct
* interpolations for long skinny triangles. We found Mali GPU probably has this
* problem: https://github.com/tensorflow/tfjs/issues/6775.
*/
ENV.registerFlag('WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE', () => Infinity);
/**
* If the flag is set to true, the max size of the narrow texture will be auto
* computed and it will be considerred as a threshold to reshape the narrow
* texture to be more squarish.
*
* This flag is used to help some GPUs that could not provide correct
* interpolations for long skinny triangles. We found Mali GPU probably has this
* problem: https://github.com/tensorflow/tfjs/issues/6775.
*/
ENV.registerFlag('WEBGL_AUTO_SQUARIFY_NARROW_TEXTURE_SHAPE', () => false);
/**
* Whether to use the customized isnan. It's only useful for webgl2 since webgl1
* doesn't have the builtin isnan.
*/
ENV.registerFlag('WEBGL2_ISNAN_CUSTOM', () => false);
/** Experimental flag, whether enter compile only phase. */
ENV.registerFlag('ENGINE_COMPILE_ONLY', () => false);
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function getGlslDifferences() {
let version;
let attribute;
let varyingVs;
let varyingFs;
let texture2D;
let output;
let defineOutput;
let defineSpecialNaN;
let defineSpecialInf;
let defineRound;
if (env().getNumber('WEBGL_VERSION') === 2) {
version = '#version 300 es';
attribute = 'in';
varyingVs = 'out';
varyingFs = 'in';
texture2D = 'texture';
output = 'outputColor';
defineOutput = 'out vec4 outputColor;';
// Use custom isnan definition to work across differences between
// implementations on various platforms. While this should happen in ANGLE
// we still see differences between android and windows (on chrome) when
// using isnan directly. Since WebGL2 supports uint type and
// floatBitsToUinT built-in function, we could implment isnan following
// IEEE 754 rules.
// NaN defination in IEEE 754-1985 is :
// - sign = either 0 or 1.
// - biased exponent = all 1 bits.
// - fraction = anything except all 0 bits (since all 0 bits represents
// infinity).
// https://en.wikipedia.org/wiki/IEEE_754-1985#Representation_of_non-numbers
defineSpecialNaN = env().getBool('WEBGL2_ISNAN_CUSTOM') ? `
bool isnan_custom(float val) {
uint floatToUint = floatBitsToUint(val);
return (floatToUint & 0x7fffffffu) > 0x7f800000u;
}
bvec4 isnan_custom(vec4 val) {
return bvec4(isnan_custom(val.x),
isnan_custom(val.y), isnan_custom(val.z), isnan_custom(val.w));
}
#define isnan(value) isnan_custom(value)
` :
'';
// In webgl 2 we do not need to specify a custom isinf so there is no
// need for a special INFINITY constant.
defineSpecialInf = ``;
defineRound = `
#define round(value) newRound(value)
int newRound(float value) {
return int(floor(value + 0.5));
}
ivec4 newRound(vec4 value) {
return ivec4(floor(value + vec4(0.5)));
}
`;
}
else {
version = '';
attribute = 'attribute';
varyingVs = 'varying';
varyingFs = 'varying';
texture2D = 'texture2D';
output = 'gl_FragColor';
defineOutput = '';
// WebGL1 has no built in isnan so we define one here.
defineSpecialNaN = `
#define isnan(value) isnan_custom(value)
bool isnan_custom(float val) {
return (val > 0. || val < 1. || val == 0.) ? false : true;
}
bvec4 isnan_custom(vec4 val) {
return bvec4(isnan(val.x), isnan(val.y), isnan(val.z), isnan(val.w));
}
`;
defineSpecialInf = `
uniform float INFINITY;
bool isinf(float val) {
return abs(val) == INFINITY;
}
bvec4 isinf(vec4 val) {
return equal(abs(val), vec4(INFINITY));
}
`;
defineRound = `
int round(float value) {
return int(floor(value + 0.5));
}
ivec4 round(vec4 value) {
return ivec4(floor(value + vec4(0.5)));
}
`;
}
return {
version,
attribute,
varyingVs,
varyingFs,
texture2D,
output,
defineOutput,
defineSpecialNaN,
defineSpecialInf,
defineRound
};
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Produces GLSL code that derives logical coordinates from a flat
* index. The code performs integer division with each stride and decrements
* the index until the index equals the final dimension coordinate.
*/
function getLogicalCoordinatesFromFlatIndex(coords, shape, index = 'index') {
const strides = computeStrides(shape);
return strides
.map((stride, i) => {
const line1 = `int ${coords[i]} = ${index} / ${stride}`;
const line2 = i === strides.length - 1 ?
`int ${coords[i + 1]} = ${index} - ${coords[i]} * ${stride}` :
`index -= ${coords[i]} * ${stride}`;
return `${line1}; ${line2};`;
})
.join('');
}
function getOutputLogicalCoordinatesFromFlatIndexByUniform(coords, shape, index = 'index') {
const strides = computeStrides(shape);
return strides
.map((_, i) => {
const line1 = `int ${coords[i]} = ${index} / outShapeStrides[${i}]`;
const line2 = i === strides.length - 1 ?
`int ${coords[i + 1]} = ${index} - ${coords[i]} * outShapeStrides[${i}]` :
`index -= ${coords[i]} * outShapeStrides[${i}]`;
return `${line1}; ${line2};`;
})
.join('');
}
// Produces GLSL code that computes strides.
function symbolicallyComputeStrides(indicesArr, variableName) {
const numCoords = indicesArr.length;
const shape = indicesArr.map(d => `${variableName}[${d}]`);
const strides = new Array(numCoords - 1);
strides[numCoords - 2] = shape[numCoords - 1];
for (let i = numCoords - 3; i >= 0; --i) {
strides[i] = `(${strides[i + 1]} * ${shape[i + 1]})`;
}
return strides;
}
function getLogicalCoordinatesFromFlatIndexByUniform(coords, variableName, index = 'index') {
const indicesArray = coords.map((_, i) => i);
const strides = symbolicallyComputeStrides(indicesArray, variableName);
return strides
.map((_, i) => {
const line1 = `int ${coords[i]} = ${index} / ${strides[i]}`;
const line2 = i === strides.length - 1 ?
`int ${coords[i + 1]} = ${index} - ${coords[i]} * ${strides[i]}` :
`index -= ${coords[i]} * ${strides[i]}`;
return `${line1}; ${line2};`;
})
.join('');
}
/**
* Produces GLSL that computes the flat index from 3D coordinates.
*/
function getFlatIndexFrom3D(shape) {
const strides = computeStrides(shape).map(d => d.toString());
return `
int getFlatIndex(ivec3 coords) {
return coords.x * ${strides[0]} + coords.y * ${strides[1]} + coords.z;
}
`;
}
function getFlatIndexFrom3DOutput() {
return `
int getFlatIndex(ivec3 coords) {
return coords.x * outShapeStrides[0] + coords.y * outShapeStrides[1] + coords.z;
}
`;
}
const ENCODE_FLOAT_SNIPPET = `
const float FLOAT_MAX = 1.70141184e38;
const float FLOAT_MIN = 1.17549435e-38;
lowp vec4 encode_float(highp float v) {
if (isnan(v)) {
return vec4(255, 255, 255, 255);
}
highp float av = abs(v);
if(av < FLOAT_MIN) {
return vec4(0.0, 0.0, 0.0, 0.0);
} else if(v > FLOAT_MAX) {
return vec4(0.0, 0.0, 128.0, 127.0) / 255.0;
} else if(v < -FLOAT_MAX) {
return vec4(0.0, 0.0, 128.0, 255.0) / 255.0;
}
highp vec4 c = vec4(0,0,0,0);
highp float e = floor(log2(av));
highp float m = exp2(fract(log2(av))) - 1.0;
c[2] = floor(128.0 * m);
m -= c[2] / 128.0;
c[1] = floor(32768.0 * m);
m -= c[1] / 32768.0;
c[0] = floor(8388608.0 * m);
highp float ebias = e + 127.0;
c[3] = floor(ebias / 2.0);
ebias -= c[3] * 2.0;
c[2] += floor(ebias) * 128.0;
c[3] += 128.0 * step(0.0, -v);
return c / 255.0;
}
`;
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Please make sure the shaker key in makeShaderKey in gpgpu_math.ts is well
// mapped if any shader source code is changed in this file.
const { getBroadcastDims } = backend_util;
function makeShader(inputsInfo, outputShape, program) {
const prefixSnippets = [];
inputsInfo.forEach(x => {
const size = sizeFromShape(x.shapeInfo.logicalShape);
// Snippet when we decided to upload the values as uniform.
if (x.shapeInfo.isUniform) {
prefixSnippets.push(`uniform float ${x.name}${size > 1 ? `[${size}]` : ''};`);
}
else {
prefixSnippets.push(`uniform sampler2D ${x.name};`);
prefixSnippets.push(`uniform int offset${x.name};`);
}
if (program.enableShapeUniforms) {
const { uniformShape } = getUniformInfoFromShape(program.packedInputs, x.shapeInfo.logicalShape, x.shapeInfo.texShape);
switch (uniformShape.length) {
case 1:
prefixSnippets.push(`uniform int ${x.name}Shape;`);
break;
case 2:
prefixSnippets.push(`uniform ivec2 ${x.name}Shape;`);
break;
case 3:
prefixSnippets.push(`uniform ivec3 ${x.name}Shape;`);
break;
case 4:
prefixSnippets.push(`uniform ivec4 ${x.name}Shape;`);
break;
}
prefixSnippets.push(`uniform ivec2 ${x.name}TexShape;`);
}
});
if (program.enableShapeUniforms) {
switch (outputShape.logicalShape.length) {
case 1:
prefixSnippets.push(`uniform int outShape;`);
break;
case 2:
prefixSnippets.push(`uniform ivec2 outShape;`);
prefixSnippets.push(`uniform int outShapeStrides;`);
break;
case 3:
prefixSnippets.push(`uniform ivec3 outShape;`);
prefixSnippets.push(`uniform ivec2 outShapeStrides;`);
break;
case 4:
prefixSnippets.push(`uniform ivec4 outShape;`);
prefixSnippets.push(`uniform ivec3 outShapeStrides;`);
break;
}
prefixSnippets.push(`uniform ivec2 outTexShape;`);
}
if (program.customUniforms) {
program.customUniforms.forEach((d) => {
prefixSnippets.push(`uniform ${d.type} ${d.name}${d.arrayIndex ? `[${d.arrayIndex}]` : ''};`);
});
}
const inputPrefixSnippet = prefixSnippets.join('\n');
const inputSamplingSnippet = inputsInfo
.map(x => getInputSamplingSnippet(x, outputShape, program.packedInputs, program.enableShapeUniforms))
.join('\n');
const outTexShape = outputShape.texShape;
const glsl = getGlslDifferences();
const floatTextureSampleSnippet = getFloatTextureSampleSnippet(glsl);
let outputSamplingSnippet;
let floatTextureSetOutputSnippet;
let shaderPrefix = getShaderPrefix(glsl);
if (outputShape.isPacked) {
outputSamplingSnippet = getPackedOutputSamplingSnippet(outputShape.logicalShape, outTexShape, program.enableShapeUniforms);
floatTextureSetOutputSnippet = getFloatTextureSetRGBASnippet(glsl);
}
else {
outputSamplingSnippet = getOutputSamplingSnippet(outputShape.logicalShape, outTexShape, program.enableShapeUniforms);
floatTextureSetOutputSnippet = getFloatTextureSetRSnippet(glsl);
}
if (program.packedInputs) {
shaderPrefix += SHADER_PACKED_PREFIX;
}
const source = [
shaderPrefix, floatTextureSampleSnippet, floatTextureSetOutputSnippet,
inputPrefixSnippet, outputSamplingSnippet, inputSamplingSnippet,
program.userCode
].join('\n');
return source;
}
function getSamplerFromInInfo(inInfo, enableShapeUniforms = false) {
const shape = inInfo.shapeInfo.logicalShape;
switch (shape.length) {
case 0:
return getSamplerScalar(inInfo, enableShapeUniforms);
case 1:
return getSampler1D(inInfo, enableShapeUniforms);
case 2:
return getSampler2D(inInfo, enableShapeUniforms);
case 3:
return getSampler3D(inInfo, enableShapeUniforms);
case 4:
return getSampler4D(inInfo, enableShapeUniforms);
case 5:
return getSampler5D(inInfo);
case 6:
return getSampler6D(inInfo);
default:
throw new Error(`${shape.length}-D input sampling` +
` is not yet supported`);
}
}
function getPackedSamplerFromInInfo(inInfo, enableShapeUniforms) {
const shape = inInfo.shapeInfo.logicalShape;
switch (shape.length) {
case 0:
return getPackedSamplerScalar(inInfo);
case 1:
return getPackedSampler1D(inInfo, enableShapeUniforms);
case 2:
return getPackedSampler2D(inInfo, enableShapeUniforms);
case 3:
return getPackedSampler3D(inInfo, enableShapeUniforms);
default:
return getPackedSamplerND(inInfo, enableShapeUniforms);
}
}
function getInputSamplingSnippet(inInfo, outShapeInfo, usesPackedTextures = false, enableShapeUniforms) {
let res = '';
if (usesPackedTextures) {
res += getPackedSamplerFromInInfo(inInfo, enableShapeUniforms);
}
else {
res += getSamplerFromInInfo(inInfo, enableShapeUniforms);
}
const inShape = inInfo.shapeInfo.logicalShape;
const outShape = outShapeInfo.logicalShape;
if (inShape.length <= outShape.length) {
if (usesPackedTextures) {
res += getPackedSamplerAtOutputCoords(inInfo, outShapeInfo);
}
else {
res += getSamplerAtOutputCoords(inInfo, outShapeInfo);
}
}
return res;
}
function getPackedOutputSamplingSnippet(outShape, outTexShape, enableShapeUniforms) {
switch (outShape.length) {
case 0:
return getOutputScalarCoords();
case 1:
return getOutputPacked1DCoords(outShape, outTexShape, enableShapeUniforms);
case 2:
return getOutputPacked2DCoords(outShape, outTexShape, enableShapeUniforms);
case 3:
return getOutputPacked3DCoords(outShape, outTexShape, enableShapeUniforms);
default:
return getOutputPackedNDCoords(outShape, outTexShape, enableShapeUniforms);
}
}
function getOutputSamplingSnippet(outShape, outTexShape, enableShapeUniforms) {
switch (outShape.length) {
case 0:
return getOutputScalarCoords();
case 1:
return getOutput1DCoords(outShape, outTexShape, enableShapeUniforms);
case 2:
return getOutput2DCoords(outShape, outTexShape, enableShapeUniforms);
case 3:
return getOutput3DCoords(outShape, outTexShape, enableShapeUniforms);
case 4:
return getOutput4DCoords(outShape, outTexShape, enableShapeUniforms);
case 5:
return getOutput5DCoords(outShape, outTexShape);
case 6:
return getOutput6DCoords(outShape, outTexShape);
default:
throw new Error(`${outShape.length}-D output sampling is not yet supported`);
}
}
function getFloatTextureSampleSnippet(glsl) {
return `
float sampleTexture(sampler2D textureSampler, vec2 uv) {
return ${glsl.texture2D}(textureSampler, uv).r;
}
`;
}
function getFloatTextureSetRSnippet(glsl) {
return `
void setOutput(float val) {
${glsl.output} = vec4(val, 0, 0, 0);
}
`;
}
function getFloatTextureSetRGBASnippet(glsl) {
return `
void setOutput(vec4 val) {
${glsl.output} = val;
}
`;
}
function getShaderPrefix(glsl) {
const SHADER_PREFIX = `${glsl.version}
precision highp float;
precision highp int;
precision highp sampler2D;
${glsl.varyingFs} vec2 resultUV;
${glsl.defineOutput}
const vec2 halfCR = vec2(0.5, 0.5);
struct ivec5
{
int x;
int y;
int z;
int w;
int u;
};
struct ivec6
{
int x;
int y;
int z;
int w;
int u;
int v;
};
uniform float NAN;
${glsl.defineSpecialNaN}
${glsl.defineSpecialInf}
${glsl.defineRound}
int imod(int x, int y) {
return x - y * (x / y);
}
int idiv(int a, int b, float sign) {
int res = a / b;
int mod = imod(a, b);
if (sign < 0. && mod != 0) {
res -= 1;
}
return res;
}
//Based on the work of Dave Hoskins
//https://www.shadertoy.com/view/4djSRW
#define HASHSCALE1 443.8975
float random(float seed){
vec2 p = resultUV * seed;
vec3 p3 = fract(vec3(p.xyx) * HASHSCALE1);
p3 += dot(p3, p3.yzx + 19.19);
return fract((p3.x + p3.y) * p3.z);
}
${SAMPLE_1D_SNIPPET}
${SAMPLE_2D_SNIPPET}
${SAMPLE_3D_SNIPPET}
`;
return SHADER_PREFIX;
}
const SAMPLE_1D_SNIPPET = `
vec2 uvFromFlat(int texNumR, int texNumC, int index) {
int texR = index / texNumC;
int texC = index - texR * texNumC;
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
vec2 packedUVfrom1D(int texNumR, int texNumC, int index) {
int texelIndex = index / 2;
int texR = texelIndex / texNumC;
int texC = texelIndex - texR * texNumC;
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
`;
const SAMPLE_2D_SNIPPET = `
vec2 packedUVfrom2D(int texelsInLogicalRow, int texNumR,
int texNumC, int row, int col) {
int texelIndex = (row / 2) * texelsInLogicalRow + (col / 2);
int texR = texelIndex / texNumC;
int texC = texelIndex - texR * texNumC;
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
`;
const SAMPLE_3D_SNIPPET = `
vec2 packedUVfrom3D(int texNumR, int texNumC,
int texelsInBatch, int texelsInLogicalRow, int b,
int row, int col) {
int index = b * texelsInBatch + (row / 2) * texelsInLogicalRow + (col / 2);
int texR = index / texNumC;
int texC = index - texR * texNumC;
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
`;
const SHADER_PACKED_PREFIX = `
float getChannel(vec4 frag, vec2 innerDims) {
vec2 modCoord = mod(innerDims, 2.);
return modCoord.x == 0. ?
(modCoord.y == 0. ? frag.r : frag.g) :
(modCoord.y == 0. ? frag.b : frag.a);
}
float getChannel(vec4 frag, int dim) {
float modCoord = mod(float(dim), 2.);
return modCoord == 0. ? frag.r : frag.g;
}
`;
function getOutputScalarCoords() {
return `
int getOutputCoords() {
return 0;
}
`;
}
function getOutputPacked1DCoords(shape, texShape, enableShapeUniforms) {
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
if (packedTexShape[0] === 1) {
if (enableShapeUniforms) {
return `
int getOutputCoords() {
return 2 * int(resultUV.x * ceil(float(outTexShape[1]) / 2.0));
}
`;
}
return `
int getOutputCoords() {
return 2 * int(resultUV.x * ${packedTexShape[1]}.0);
}
`;
}
if (packedTexShape[1] === 1) {
if (enableShapeUniforms) {
return `
int getOutputCoords() {
return 2 * int(resultUV.y * ceil(float(outTexShape[0]) / 2.0));
}
`;
}
return `
int getOutputCoords() {
return 2 * int(resultUV.y * ${packedTexShape[0]}.0);
}
`;
}
if (enableShapeUniforms) {
return `
int getOutputCoords() {
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(packedTexShape[0], packedTexShape[1]));
return 2 * (resTexRC.x * packedTexShape[1] + resTexRC.y);
}
`;
}
return `
int getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${packedTexShape[0]}, ${packedTexShape[1]}));
return 2 * (resTexRC.x * ${packedTexShape[1]} + resTexRC.y);
}
`;
}
function getOutput1DCoords(shape, texShape, enableShapeUniforms) {
if (texShape[0] === 1) {
if (enableShapeUniforms) {
return `
int getOutputCoords() {
return int(resultUV.x * float(outTexShape[1]));
}
`;
}
return `
int getOutputCoords() {
return int(resultUV.x * ${texShape[1]}.0);
}
`;
}
if (texShape[1] === 1) {
if (enableShapeUniforms) {
return `
int getOutputCoords() {
return int(resultUV.y * float(outTexShape[0]));
}
`;
}
return `
int getOutputCoords() {
return int(resultUV.y * ${texShape[0]}.0);
}
`;
}
if (enableShapeUniforms) {
return `
int getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(outTexShape[0], outTexShape[1]));
return resTexRC.x * outTexShape[1] + resTexRC.y;
}
`;
}
return `
int getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${texShape[0]}, ${texShape[1]}));
return resTexRC.x * ${texShape[1]} + resTexRC.y;
}
`;
}
function getOutputPacked3DCoords(shape, texShape, enableShapeUniforms) {
if (enableShapeUniforms) {
return `
ivec3 getOutputCoords() {
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
int texelsInLogicalRow = int(ceil(float(outShape[2]) / 2.0));
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[1]) / 2.0));
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(packedTexShape[0], packedTexShape[1]));
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
int b = index / texelsInBatch;
index -= b * texelsInBatch;
int r = 2 * (index / texelsInLogicalRow);
int c = imod(index, texelsInLogicalRow) * 2;
return ivec3(b, r, c);
}
`;
}
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
const texelsInLogicalRow = Math.ceil(shape[2] / 2);
const texelsInBatch = texelsInLogicalRow * Math.ceil(shape[1] / 2);
return `
ivec3 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${packedTexShape[0]}, ${packedTexShape[1]}));
int index = resTexRC.x * ${packedTexShape[1]} + resTexRC.y;
int b = index / ${texelsInBatch};
index -= b * ${texelsInBatch};
int r = 2 * (index / ${texelsInLogicalRow});
int c = imod(index, ${texelsInLogicalRow}) * 2;
return ivec3(b, r, c);
}
`;
}
function getOutput3DCoords(shape, texShape, enableShapeUniforms) {
if (enableShapeUniforms) {
const coordsFromIndexSnippet = getOutputLogicalCoordinatesFromFlatIndexByUniform(['r', 'c', 'd'], shape);
return `
ivec3 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(outTexShape[0], outTexShape[1]));
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
${coordsFromIndexSnippet}
return ivec3(r, c, d);
}
`;
}
const coordsFromIndexSnippet = getLogicalCoordinatesFromFlatIndex(['r', 'c', 'd'], shape);
return `
ivec3 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${texShape[0]}, ${texShape[1]}));
int index = resTexRC.x * ${texShape[1]} + resTexRC.y;
${coordsFromIndexSnippet}
return ivec3(r, c, d);
}
`;
}
function getOutputPackedNDCoords(shape, texShape, enableShapeUniforms) {
if (enableShapeUniforms) {
// TODO: support 5d and 6d
return `
ivec4 getOutputCoords() {
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(packedTexShape[0], packedTexShape[1]));
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
int texelsInLogicalRow = int(ceil(float(outShape[3]) / 2.0));
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[2]) / 2.0));
int texelsInBatchN = texelsInBatch * outShape[1];
int b2 = index / texelsInBatchN;
index -= b2 * texelsInBatchN;
int b = index / texelsInBatch;
index -= b * texelsInBatch;
int r = 2 * (index / texelsInLogicalRow);
int c = imod(index, texelsInLogicalRow) * 2;
return ivec4(b2, b, r, c);
}
`;
}
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
const texelsInLogicalRow = Math.ceil(shape[shape.length - 1] / 2);
const texelsInBatch = texelsInLogicalRow * Math.ceil(shape[shape.length - 2] / 2);
let texelsInBatchN = texelsInBatch;
let batches = ``;
let coords = 'b, r, c';
for (let b = 2; b < shape.length - 1; b++) {
texelsInBatchN *= shape[shape.length - b - 1];
batches = `
int b${b} = index / ${texelsInBatchN};
index -= b${b} * ${texelsInBatchN};
` + batches;
coords = `b${b}, ` + coords;
}
return `
ivec${shape.length} getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${packedTexShape[0]}, ${packedTexShape[1]}));
int index = resTexRC.x * ${packedTexShape[1]} + resTexRC.y;
${batches}
int b = index / ${texelsInBatch};
index -= b * ${texelsInBatch};
int r = 2 * (index / ${texelsInLogicalRow});
int c = imod(index, ${texelsInLogicalRow}) * 2;
return ivec${shape.length}(${coords});
}
`;
}
function getOutput4DCoords(shape, texShape, enableShapeUniforms) {
if (enableShapeUniforms) {
const coordsFromIndexSnippet = getOutputLogicalCoordinatesFromFlatIndexByUniform(['r', 'c', 'd', 'd2'], shape);
return `
ivec4 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(outTexShape[0], outTexShape[1]));
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
${coordsFromIndexSnippet}
return ivec4(r, c, d, d2);
}
`;
}
const coordsFromIndexSnippet = getLogicalCoordinatesFromFlatIndex(['r', 'c', 'd', 'd2'], shape);
return `
ivec4 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${texShape[0]}, ${texShape[1]}));
int index = resTexRC.x * ${texShape[1]} + resTexRC.y;
${coordsFromIndexSnippet}
return ivec4(r, c, d, d2);
}
`;
}
function getOutput5DCoords(shape, texShape) {
const coordsFromIndexSnippet = getLogicalCoordinatesFromFlatIndex(['r', 'c', 'd', 'd2', 'd3'], shape);
return `
ivec5 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx * vec2(${texShape[0]},
${texShape[1]}));
int index = resTexRC.x * ${texShape[1]} + resTexRC.y;
${coordsFromIndexSnippet}
ivec5 outShape = ivec5(r, c, d, d2, d3);
return outShape;
}
`;
}
function getOutput6DCoords(shape, texShape) {
const coordsFromIndexSnippet = getLogicalCoordinatesFromFlatIndex(['r', 'c', 'd', 'd2', 'd3', 'd4'], shape);
return `
ivec6 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${texShape[0]}, ${texShape[1]}));
int index = resTexRC.x * ${texShape[1]} + resTexRC.y;
${coordsFromIndexSnippet}
ivec6 result = ivec6(r, c, d, d2, d3, d4);
return result;
}
`;
}
function getOutputPacked2DCoords(shape, texShape, enableShapeUniforms) {
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
if (arraysEqual(shape, texShape)) {
if (enableShapeUniforms) {
return `
ivec2 getOutputCoords() {
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
return 2 * ivec2(resultUV.yx * vec2(packedTexShape[0], packedTexShape[1]));
}
`;
}
return `
ivec2 getOutputCoords() {
return 2 * ivec2(resultUV.yx * vec2(${packedTexShape[0]}, ${packedTexShape[1]}));
}
`;
}
// texels needed to accommodate a logical row
const texelsInLogicalRow = Math.ceil(shape[1] / 2);
/**
* getOutputCoords
*
* resTexRC: The rows and columns of the texels. If you move over one
* texel to the right in the packed texture, you are moving over one column
* (not two).
*
* index: The texel index
*/
if (enableShapeUniforms) {
return `
ivec2 getOutputCoords() {
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
int texelsInLogicalRow = int(ceil(float(outShape[1]) / 2.0));
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(packedTexShape[0], packedTexShape[1]));
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
int r = 2 * (index / texelsInLogicalRow);
int c = imod(index, texelsInLogicalRow) * 2;
return ivec2(r, c);
}
`;
}
return `
ivec2 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${packedTexShape[0]}, ${packedTexShape[1]}));
int index = resTexRC.x * ${packedTexShape[1]} + resTexRC.y;
int r = 2 * (index / ${texelsInLogicalRow});
int c = imod(index, ${texelsInLogicalRow}) * 2;
return ivec2(r, c);
}
`;
}
function getOutput2DCoords(shape, texShape, enableShapeUniforms) {
if (arraysEqual(shape, texShape)) {
if (enableShapeUniforms) {
return `
ivec2 getOutputCoords() {
return ivec2(resultUV.yx * vec2(outTexShape[0], outTexShape[1]));
}
`;
}
return `
ivec2 getOutputCoords() {
return ivec2(resultUV.yx * vec2(${texShape[0]}, ${texShape[1]}));
}
`;
}
if (shape[1] === 1) {
if (enableShapeUniforms) {
return `
ivec2 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(outTexShape[0], outTexShape[1]));
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
return ivec2(index, 0);
}
`;
}
return `
ivec2 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${texShape[0]}, ${texShape[1]}));
int index = resTexRC.x * ${texShape[1]} + resTexRC.y;
return ivec2(index, 0);
}
`;
}
if (shape[0] === 1) {
if (enableShapeUniforms) {
return `
ivec2 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(outTexShape[0], outTexShape[1]));
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
return ivec2(0, index);
}
`;
}
return `
ivec2 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${texShape[0]}, ${texShape[1]}));
int index = resTexRC.x * ${texShape[1]} + resTexRC.y;
return ivec2(0, index);
}
`;
}
if (enableShapeUniforms) {
return `
ivec2 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(outTexShape[0], outTexShape[1]));
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
int r = index / outShape[1];
int c = index - r * outShape[1];
return ivec2(r, c);
}
`;
}
return `
ivec2 getOutputCoords() {
ivec2 resTexRC = ivec2(resultUV.yx *
vec2(${texShape[0]}, ${texShape[1]}));
int index = resTexRC.x * ${texShape[1]} + resTexRC.y;
int r = index / ${shape[1]};
int c = index - r * ${shape[1]};
return ivec2(r, c);
}
`;
}
function getFlatOffsetUniformName(texName) {
return `offset${texName}`;
}
function getPackedSamplerScalar(inputInfo) {
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const glsl = getGlslDifferences();
return `
vec4 ${funcName}() {
return ${glsl.texture2D}(${texName}, halfCR);
}
`;
}
function getSamplerScalar(inputInfo, enableShapeUniforms) {
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
if (inputInfo.shapeInfo.isUniform) {
return `float ${funcName}() {return ${texName};}`;
}
const [texNumR, texNumC] = inputInfo.shapeInfo.texShape;
if (texNumR === 1 && texNumC === 1) {
return `
float ${funcName}() {
return sampleTexture(${texName}, halfCR);
}
`;
}
const offset = getFlatOffsetUniformName(texName);
if (enableShapeUniforms) {
return `
float ${funcName}() {
vec2 uv = uvFromFlat(${texName}TexShape[0], ${texName}TexShape[1], ${offset});
return sampleTexture(${texName}, uv);
}
`;
}
const [tNumR, tNumC] = inputInfo.shapeInfo.texShape;
return `
float ${funcName}() {
vec2 uv = uvFromFlat(${tNumR}, ${tNumC}, ${offset});
return sampleTexture(${texName}, uv);
}
`;
}
function getPackedSampler1D(inputInfo, enableShapeUniforms) {
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const texShape = inputInfo.shapeInfo.texShape;
const glsl = getGlslDifferences();
if (enableShapeUniforms) {
return `
vec4 ${funcName}(int index) {
ivec2 packedTexShape = ivec2(ceil(float(${texName}TexShape[0]) / 2.0), ceil(float(${texName}TexShape[1]) / 2.0));
vec2 uv = packedUVfrom1D(
packedTexShape[0], packedTexShape[1], index);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
return `
vec4 ${funcName}(int index) {
vec2 uv = packedUVfrom1D(
${packedTexShape[0]}, ${packedTexShape[1]}, index);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
function getSampler1D(inputInfo, enableShapeUniforms) {
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
if (inputInfo.shapeInfo.isUniform) {
// Uniform arrays will be less than 65505 (no risk of float16 overflow).
return `
float ${funcName}(int index) {
${getUniformSampler(inputInfo)}
}
`;
}
const texShape = inputInfo.shapeInfo.texShape;
const tNumR = texShape[0];
const tNumC = texShape[1];
if (tNumC === 1 && tNumR === 1) {
return `
float ${funcName}(int index) {
return sampleTexture(${texName}, halfCR);
}
`;
}
const offset = getFlatOffsetUniformName(texName);
if (tNumC === 1) {
if (enableShapeUniforms) {
return `
float ${funcName}(int index) {
vec2 uv = vec2(0.5, (float(index + ${offset}) + 0.5) / float(${texName}TexShape[0]));
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int index) {
vec2 uv = vec2(0.5, (float(index + ${offset}) + 0.5) / ${tNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
if (tNumR === 1) {
if (enableShapeUniforms) {
return `
float ${funcName}(int index) {
vec2 uv = vec2((float(index + ${offset}) + 0.5) / float(${texName}TexShape[1]), 0.5);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int index) {
vec2 uv = vec2((float(index + ${offset}) + 0.5) / ${tNumC}.0, 0.5);
return sampleTexture(${texName}, uv);
}
`;
}
if (enableShapeUniforms) {
return `
float ${funcName}(int index) {
vec2 uv = uvFromFlat(${texName}TexShape[0], ${texName}TexShape[1], index + ${offset});
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int index) {
vec2 uv = uvFromFlat(${tNumR}, ${tNumC}, index + ${offset});
return sampleTexture(${texName}, uv);
}
`;
}
function getPackedSampler2D(inputInfo, enableShapeUniforms) {
const shape = inputInfo.shapeInfo.logicalShape;
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const texShape = inputInfo.shapeInfo.texShape;
const texNumR = texShape[0];
const texNumC = texShape[1];
const glsl = getGlslDifferences();
if (texShape != null && arraysEqual(shape, texShape)) {
if (enableShapeUniforms) {
return `
vec4 ${funcName}(int row, int col) {
vec2 uv = (vec2(col, row) + halfCR) / vec2(${texName}TexShape[1], ${texName}TexShape[0]);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
return `
vec4 ${funcName}(int row, int col) {
vec2 uv = (vec2(col, row) + halfCR) / vec2(${texNumC}.0, ${texNumR}.0);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
if (enableShapeUniforms) {
return `
vec4 ${funcName}(int row, int col) {
ivec2 packedTexShape = ivec2(ceil(float(${texName}TexShape[0]) / 2.0), ceil(float(${texName}TexShape[1]) / 2.0));
int valuesPerRow = int(ceil(float(${texName}Shape[1]) / 2.0));
vec2 uv = packedUVfrom2D(valuesPerRow, packedTexShape[0], packedTexShape[1], row, col);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
const valuesPerRow = Math.ceil(shape[1] / 2);
return `
vec4 ${funcName}(int row, int col) {
vec2 uv = packedUVfrom2D(${valuesPerRow}, ${packedTexShape[0]}, ${packedTexShape[1]}, row, col);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
function getSampler2D(inputInfo, enableShapeUniforms) {
const shape = inputInfo.shapeInfo.logicalShape;
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const texShape = inputInfo.shapeInfo.texShape;
if (texShape != null && arraysEqual(shape, texShape)) {
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col) {
vec2 uv = (vec2(col, row) + halfCR) / vec2(${texName}TexShape[1], ${texName}TexShape[0]);
return sampleTexture(${texName}, uv);
}
`;
}
const texNumR = texShape[0];
const texNumC = texShape[1];
return `
float ${funcName}(int row, int col) {
vec2 uv = (vec2(col, row) + halfCR) / vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
const { newShape, keptDims } = squeezeShape(shape);
const squeezedShape = newShape;
if (squeezedShape.length < shape.length) {
const newInputInfo = squeezeInputInfo(inputInfo, squeezedShape);
const params = ['row', 'col'];
return `
${getSamplerFromInInfo(newInputInfo, enableShapeUniforms)}
float ${funcName}(int row, int col) {
return ${funcName}(${getSqueezedParams(params, keptDims)});
}
`;
}
if (inputInfo.shapeInfo.isUniform) {
// Uniform arrays will be less than 65505 (no risk of float16 overflow).
return `
float ${funcName}(int row, int col) {
int index = round(dot(vec2(row, col), vec2(${shape[1]}, 1)));
${getUniformSampler(inputInfo)}
}
`;
}
const texNumR = texShape[0];
const texNumC = texShape[1];
const offset = getFlatOffsetUniformName(texName);
if (texNumC === 1) {
// index is used directly as physical (no risk of float16 overflow).
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col) {
float index = dot(vec3(row, col, ${offset}), vec3(${texName}Shape[1], 1, 1));
vec2 uv = vec2(0.5, (index + 0.5) / float(${texName}TexShape[0]));
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col) {
float index = dot(vec3(row, col, ${offset}), vec3(${shape[1]}, 1, 1));
vec2 uv = vec2(0.5, (index + 0.5) / ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
if (texNumR === 1) {
// index is used directly as physical (no risk of float16 overflow).
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col) {
float index = dot(vec3(row, col, ${offset}), vec3(${texName}Shape[1], 1, 1));
vec2 uv = vec2((index + 0.5) / float(${texName}TexShape[1]), 0.5);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col) {
float index = dot(vec3(row, col, ${offset}), vec3(${shape[1]}, 1, 1));
vec2 uv = vec2((index + 0.5) / ${texNumC}.0, 0.5);
return sampleTexture(${texName}, uv);
}
`;
}
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * ${texName}Shape[1] + col + ${offset};
vec2 uv = uvFromFlat(${texName}TexShape[0], ${texName}TexShape[1], index);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * ${shape[1]} + col + ${offset};
vec2 uv = uvFromFlat(${texNumR}, ${texNumC}, index);
return sampleTexture(${texName}, uv);
}
`;
}
function getPackedSampler3D(inputInfo, enableShapeUniforms) {
const shape = inputInfo.shapeInfo.logicalShape;
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const texShape = inputInfo.shapeInfo.texShape;
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
if (shape[0] === 1) {
const squeezedShape = shape.slice(1);
const keptDims = [1, 2];
const newInputInfo = squeezeInputInfo(inputInfo, squeezedShape);
const params = ['b', 'row', 'col'];
return `
${getPackedSamplerFromInInfo(newInputInfo, enableShapeUniforms)}
vec4 ${funcName}(int b, int row, int col) {
return ${funcName}(${getSqueezedParams(params, keptDims)});
}
`;
}
const glsl = getGlslDifferences();
if (enableShapeUniforms) {
return `
vec4 ${funcName}(int b, int row, int col) {
ivec2 packedTexShape = ivec2(ceil(float(${texName}TexShape[0]) / 2.0), ceil(float(${texName}TexShape[1]) / 2.0));
int valuesPerRow = int(ceil(float(${texName}Shape[2]) / 2.0));
int texelsInBatch = valuesPerRow * int(ceil(float(${texName}Shape[1]) / 2.0));
vec2 uv = packedUVfrom3D(
packedTexShape[0], packedTexShape[1], texelsInBatch, valuesPerRow, b, row, col);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
const texNumR = packedTexShape[0];
const texNumC = packedTexShape[1];
const valuesPerRow = Math.ceil(shape[2] / 2);
const texelsInBatch = valuesPerRow * Math.ceil(shape[1] / 2);
return `
vec4 ${funcName}(int b, int row, int col) {
vec2 uv = packedUVfrom3D(
${texNumR}, ${texNumC}, ${texelsInBatch}, ${valuesPerRow}, b, row, col);
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
function getSampler3D(inputInfo, enableShapeUniforms) {
const shape = inputInfo.shapeInfo.logicalShape;
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const stride0 = shape[1] * shape[2];
const stride1 = shape[2];
const { newShape, keptDims } = squeezeShape(shape);
const squeezedShape = newShape;
if (squeezedShape.length < shape.length) {
const newInputInfo = squeezeInputInfo(inputInfo, squeezedShape);
const params = ['row', 'col', 'depth'];
return `
${getSamplerFromInInfo(newInputInfo, enableShapeUniforms)}
float ${funcName}(int row, int col, int depth) {
return ${funcName}(${getSqueezedParams(params, keptDims)});
}
`;
}
if (inputInfo.shapeInfo.isUniform) {
// Uniform arrays will be less than 65505 (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth) {
int index = round(dot(vec3(row, col, depth),
vec3(${stride0}, ${stride1}, 1)));
${getUniformSampler(inputInfo)}
}
`;
}
const texShape = inputInfo.shapeInfo.texShape;
const texNumR = texShape[0];
const texNumC = texShape[1];
const flatOffset = inputInfo.shapeInfo.flatOffset;
if (texNumC === stride0 && flatOffset == null) {
// texC is used directly as physical (no risk of float16 overflow).
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col, int depth) {
int stride1 = ${texName}Shape[2];
float texR = float(row);
float texC = dot(vec2(col, depth), vec2(stride1, 1));
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texName}TexShape[1], ${texName}TexShape[0]);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col, int depth) {
float texR = float(row);
float texC = dot(vec2(col, depth), vec2(${stride1}, 1));
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
if (texNumC === stride1 && flatOffset == null) {
// texR is used directly as physical (no risk of float16 overflow).
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col, int depth) {
float texR = dot(vec2(row, col), vec2(${texName}Shape[1], 1));
float texC = float(depth);
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${texName}TexShape[1], ${texName}TexShape[0]);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col, int depth) {
float texR = dot(vec2(row, col), vec2(${shape[1]}, 1));
float texC = float(depth);
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
const offset = getFlatOffsetUniformName(texName);
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col, int depth) {
// Explicitly use integer operations as dot() only works on floats.
int stride0 = ${texName}Shape[1] * ${texName}Shape[2];
int stride1 = ${texName}Shape[2];
int index = row * stride0 + col * stride1 + depth + ${offset};
vec2 uv = uvFromFlat(${texName}TexShape[0], ${texName}TexShape[1], index);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col, int depth) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * ${stride0} + col * ${stride1} + depth + ${offset};
vec2 uv = uvFromFlat(${texNumR}, ${texNumC}, index);
return sampleTexture(${texName}, uv);
}
`;
}
function getPackedSamplerND(inputInfo, enableShapeUniforms) {
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const glsl = getGlslDifferences();
if (enableShapeUniforms) {
// TODO: support 5d and 6d
return `
vec4 ${funcName}(int b2, int b, int row, int col) {
int valuesPerRow = int(ceil(float(${texName}Shape[3]) / 2.0));
int texelsInBatch = valuesPerRow * int(ceil(float(${texName}Shape[2]) / 2.0));
int index = b * texelsInBatch + (row / 2) * valuesPerRow + (col / 2);
texelsInBatch *= ${texName}Shape[1];
index = b2 * texelsInBatch + index;
ivec2 packedTexShape = ivec2(ceil(float(${texName}TexShape[0]) / 2.0), ceil(float(${texName}TexShape[1]) / 2.0));
int texR = index / packedTexShape[1];
int texC = index - texR * packedTexShape[1];
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(packedTexShape[1], packedTexShape[0]); return ${glsl.texture2D}(${texName}, uv);
}
`;
}
const shape = inputInfo.shapeInfo.logicalShape;
const rank = shape.length;
const texShape = inputInfo.shapeInfo.texShape;
const packedTexShape = [Math.ceil(texShape[0] / 2), Math.ceil(texShape[1] / 2)];
const texNumR = packedTexShape[0];
const texNumC = packedTexShape[1];
const valuesPerRow = Math.ceil(shape[rank - 1] / 2);
let texelsInBatch = valuesPerRow * Math.ceil(shape[rank - 2] / 2);
let params = `int b, int row, int col`;
let index = `b * ${texelsInBatch} + (row / 2) * ${valuesPerRow} + (col / 2)`;
for (let b = 2; b < rank - 1; b++) {
params = `int b${b}, ` + params;
texelsInBatch *= shape[rank - b - 1];
index = `b${b} * ${texelsInBatch} + ` + index;
}
return `
vec4 ${funcName}(${params}) {
int index = ${index};
int texR = index / ${texNumC};
int texC = index - texR * ${texNumC};
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${texNumC}, ${texNumR});
return ${glsl.texture2D}(${texName}, uv);
}
`;
}
function getSampler4D(inputInfo, enableShapeUniforms) {
const shape = inputInfo.shapeInfo.logicalShape;
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const stride2 = shape[3];
const stride1 = shape[2] * stride2;
const stride0 = shape[1] * stride1;
const { newShape, keptDims } = squeezeShape(shape);
if (newShape.length < shape.length) {
const newInputInfo = squeezeInputInfo(inputInfo, newShape);
const params = ['row', 'col', 'depth', 'depth2'];
return `
${getSamplerFromInInfo(newInputInfo, enableShapeUniforms)}
float ${funcName}(int row, int col, int depth, int depth2) {
return ${funcName}(${getSqueezedParams(params, keptDims)});
}
`;
}
if (inputInfo.shapeInfo.isUniform) {
// Uniform arrays will be less than 65505 (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth, int depth2) {
int index = round(dot(vec4(row, col, depth, depth2),
vec4(${stride0}, ${stride1}, ${stride2}, 1)));
${getUniformSampler(inputInfo)}
}
`;
}
const flatOffset = inputInfo.shapeInfo.flatOffset;
const texShape = inputInfo.shapeInfo.texShape;
const texNumR = texShape[0];
const texNumC = texShape[1];
const stride2Str = `int stride2 = ${texName}Shape[3];`;
const stride1Str = `int stride1 = ${texName}Shape[2] * stride2;`;
const stride0Str = `int stride0 = ${texName}Shape[1] * stride1;`;
if (texNumC === stride0 && flatOffset == null) {
// texC is used directly as physical (no risk of float16 overflow).
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col, int depth, int depth2) {
${stride2Str}
${stride1Str}
float texR = float(row);
float texC =
dot(vec3(col, depth, depth2),
vec3(stride1, stride2, 1));
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texName}TexShape[1], ${texName}TexShape[0]);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col, int depth, int depth2) {
float texR = float(row);
float texC =
dot(vec3(col, depth, depth2),
vec3(${stride1}, ${stride2}, 1));
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
if (texNumC === stride2 && flatOffset == null) {
// texR is used directly as physical (no risk of float16 overflow).
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col, int depth, int depth2) {
float texR = dot(vec3(row, col, depth),
vec3(${texName}Shape[1] * ${texName}Shape[2], ${texName}Shape[2], 1));
float texC = float(depth2);
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texName}TexShape[1], ${texName}TexShape[0]);
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col, int depth, int depth2) {
float texR = dot(vec3(row, col, depth),
vec3(${shape[1] * shape[2]}, ${shape[2]}, 1));
float texC = float(depth2);
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
const offset = getFlatOffsetUniformName(texName);
if (enableShapeUniforms) {
return `
float ${funcName}(int row, int col, int depth, int depth2) {
// Explicitly use integer operations as dot() only works on floats.
${stride2Str}
${stride1Str}
${stride0Str}
int index = row * stride0 + col * stride1 +
depth * stride2 + depth2;
vec2 uv = uvFromFlat(${texName}TexShape[0], ${texName}TexShape[1], index + ${offset});
return sampleTexture(${texName}, uv);
}
`;
}
return `
float ${funcName}(int row, int col, int depth, int depth2) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * ${stride0} + col * ${stride1} +
depth * ${stride2} + depth2;
vec2 uv = uvFromFlat(${texNumR}, ${texNumC}, index + ${offset});
return sampleTexture(${texName}, uv);
}
`;
}
function getSampler5D(inputInfo) {
const shape = inputInfo.shapeInfo.logicalShape;
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const stride3 = shape[4];
const stride2 = shape[3] * stride3;
const stride1 = shape[2] * stride2;
const stride0 = shape[1] * stride1;
const { newShape, keptDims } = squeezeShape(shape);
if (newShape.length < shape.length) {
const newInputInfo = squeezeInputInfo(inputInfo, newShape);
const params = ['row', 'col', 'depth', 'depth2', 'depth3'];
return `
${getSamplerFromInInfo(newInputInfo)}
float ${funcName}(int row, int col, int depth, int depth2, int depth3) {
return ${funcName}(${getSqueezedParams(params, keptDims)});
}
`;
}
if (inputInfo.shapeInfo.isUniform) {
// Uniform arrays will be less than 65505 (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth, int depth2, int depth3) {
float index = dot(
vec4(row, col, depth, depth2),
vec4(${stride0}, ${stride1}, ${stride2}, ${stride3})) +
depth3;
${getUniformSampler(inputInfo)}
}
`;
}
const flatOffset = inputInfo.shapeInfo.flatOffset;
const texShape = inputInfo.shapeInfo.texShape;
const texNumR = texShape[0];
const texNumC = texShape[1];
if (texNumC === stride0 && flatOffset == null) {
// texC is used directly as physical (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth, int depth2, int depth3) {
int texR = row;
float texC = dot(vec4(col, depth, depth2, depth3),
vec4(${stride1}, ${stride2}, ${stride3}, 1));
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
if (texNumC === stride3 && flatOffset == null) {
// texR is used directly as physical (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth, int depth2, int depth3) {
float texR = dot(
vec4(row, col, depth, depth2),
vec4(${shape[1] * shape[2] * shape[3]},
${shape[2] * shape[3]}, ${shape[3]}, 1));
int texC = depth3;
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
const offset = getFlatOffsetUniformName(texName);
return `
float ${funcName}(int row, int col, int depth, int depth2, int depth3) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * ${stride0} + col * ${stride1} + depth * ${stride2} +
depth2 * ${stride3} + depth3 + ${offset};
vec2 uv = uvFromFlat(${texNumR}, ${texNumC}, index);
return sampleTexture(${texName}, uv);
}
`;
}
function getSampler6D(inputInfo) {
const shape = inputInfo.shapeInfo.logicalShape;
const texName = inputInfo.name;
const funcName = 'get' + texName.charAt(0).toUpperCase() + texName.slice(1);
const { newShape, keptDims } = squeezeShape(shape);
if (newShape.length < shape.length) {
const newInputInfo = squeezeInputInfo(inputInfo, newShape);
const params = ['row', 'col', 'depth', 'depth2', 'depth3', 'depth4'];
return `
${getSamplerFromInInfo(newInputInfo)}
float ${funcName}(int row, int col, int depth,
int depth2, int depth3, int depth4) {
return ${funcName}(${getSqueezedParams(params, keptDims)});
}
`;
}
const stride4 = shape[5];
const stride3 = shape[4] * stride4;
const stride2 = shape[3] * stride3;
const stride1 = shape[2] * stride2;
const stride0 = shape[1] * stride1;
if (inputInfo.shapeInfo.isUniform) {
// Uniform arrays will be less than 65505 (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth,
int depth2, int depth3, int depth4) {
int index = round(dot(
vec4(row, col, depth, depth2),
vec4(${stride0}, ${stride1}, ${stride2}, ${stride3})) +
dot(
vec2(depth3, depth4),
vec2(${stride4}, 1)));
${getUniformSampler(inputInfo)}
}
`;
}
const flatOffset = inputInfo.shapeInfo.flatOffset;
const texShape = inputInfo.shapeInfo.texShape;
const texNumR = texShape[0];
const texNumC = texShape[1];
if (texNumC === stride0 && flatOffset == null) {
// texC is used directly as physical (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth,
int depth2, int depth3, int depth4) {
int texR = row;
float texC = dot(vec4(col, depth, depth2, depth3),
vec4(${stride1}, ${stride2}, ${stride3}, ${stride4})) +
float(depth4);
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
if (texNumC === stride4 && flatOffset == null) {
// texR is used directly as physical (no risk of float16 overflow).
return `
float ${funcName}(int row, int col, int depth,
int depth2, int depth3, int depth4) {
float texR = dot(vec4(row, col, depth, depth2),
vec4(${shape[1] * shape[2] * shape[3] * shape[4]},
${shape[2] * shape[3] * shape[4]},
${shape[3] * shape[4]},
${shape[4]})) + float(depth3);
int texC = depth4;
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${texNumC}.0, ${texNumR}.0);
return sampleTexture(${texName}, uv);
}
`;
}
const offset = getFlatOffsetUniformName(texName);
return `
float ${funcName}(int row, int col, int depth,
int depth2, int depth3, int depth4) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * ${stride0} + col * ${stride1} + depth * ${stride2} +
depth2 * ${stride3} + depth3 * ${stride4} + depth4 + ${offset};
vec2 uv = uvFromFlat(${texNumR}, ${texNumC}, index);
return sampleTexture(${texName}, uv);
}
`;
}
function getUniformSampler(inputInfo) {
const texName = inputInfo.name;
const inSize = sizeFromShape(inputInfo.shapeInfo.logicalShape);
if (inSize < 2) {
return `return ${texName};`;
}
return `
for (int i = 0; i < ${inSize}; i++) {
if (i == index) {
return ${texName}[i];
}
}
`;
}
function getPackedSamplerAtOutputCoords(inputInfo, outShapeInfo) {
const texName = inputInfo.name;
const texFuncSnippet = texName.charAt(0).toUpperCase() + texName.slice(1);
const funcName = 'get' + texFuncSnippet + 'AtOutCoords';
const inRank = inputInfo.shapeInfo.logicalShape.length;
const outRank = outShapeInfo.logicalShape.length;
const broadcastDims = getBroadcastDims(inputInfo.shapeInfo.logicalShape, outShapeInfo.logicalShape);
const type = getCoordsDataType(outRank);
const rankDiff = outRank - inRank;
let coordsSnippet;
const fields = ['x', 'y', 'z', 'w', 'u', 'v'];
if (inRank === 0) {
coordsSnippet = '';
}
else if (outRank < 2 && broadcastDims.length >= 1) {
coordsSnippet = 'coords = 0;';
}
else {
coordsSnippet =
broadcastDims.map(d => `coords.${fields[d + rankDiff]} = 0;`)
.join('\n');
}
let unpackedCoordsSnippet = '';
if (outRank < 2 && inRank > 0) {
unpackedCoordsSnippet = 'coords';
}
else {
unpackedCoordsSnippet = inputInfo.shapeInfo.logicalShape
.map((s, i) => `coords.${fields[i + rankDiff]}`)
.join(', ');
}
let output = `return outputValue;`;
const inSize = sizeFromShape(inputInfo.shapeInfo.logicalShape);
const isInputScalar = inSize === 1;
const outSize = sizeFromShape(outShapeInfo.logicalShape);
const isOutputScalar = outSize === 1;
if (inRank === 1 && !isInputScalar && !isOutputScalar) {
output = `
return vec4(outputValue.xy, outputValue.xy);
`;
}
else if (isInputScalar && !isOutputScalar) {
if (outRank === 1) {
output = `
return vec4(outputValue.x, outputValue.x, 0., 0.);
`;
}
else {
output = `
return vec4(outputValue.x);
`;
}
}
else if (broadcastDims.length) {
const rows = inRank - 2;
const cols = inRank - 1;
if (broadcastDims.indexOf(rows) > -1 && broadcastDims.indexOf(cols) > -1) {
output = `return vec4(outputValue.x);`;
}
else if (broadcastDims.indexOf(rows) > -1) {
output = `return vec4(outputValue.x, outputValue.y, ` +
`outputValue.x, outputValue.y);`;
}
else if (broadcastDims.indexOf(cols) > -1) {
output = `return vec4(outputValue.xx, outputValue.zz);`;
}
}
return `
vec4 ${funcName}() {
${type} coords = getOutputCoords();
${coordsSnippet}
vec4 outputValue = get${texFuncSnippet}(${unpackedCoordsSnippet});
${output}
}
`;
}
function getSamplerAtOutputCoords(inputInfo, outShapeInfo) {
const texName = inputInfo.name;
const texFuncSnippet = texName.charAt(0).toUpperCase() + texName.slice(1);
const funcName = 'get' + texFuncSnippet + 'AtOutCoords';
const outTexShape = outShapeInfo.texShape;
const inTexShape = inputInfo.shapeInfo.texShape;
const inRank = inputInfo.shapeInfo.logicalShape.length;
const outRank = outShapeInfo.logicalShape.length;
if (!inputInfo.shapeInfo.isUniform && inRank === outRank &&
inputInfo.shapeInfo.flatOffset == null &&
arraysEqual(inTexShape, outTexShape)) {
return `
float ${funcName}() {
return sampleTexture(${texName}, resultUV);
}
`;
}
const type = getCoordsDataType(outRank);
const broadcastDims = getBroadcastDims(inputInfo.shapeInfo.logicalShape, outShapeInfo.logicalShape);
const rankDiff = outRank - inRank;
let coordsSnippet;
const fields = ['x', 'y', 'z', 'w', 'u', 'v'];
if (inRank === 0) {
coordsSnippet = '';
}
else if (outRank < 2 && broadcastDims.length >= 1) {
coordsSnippet = 'coords = 0;';
}
else {
coordsSnippet =
broadcastDims.map(d => `coords.${fields[d + rankDiff]} = 0;`)
.join('\n');
}
let unpackedCoordsSnippet = '';
if (outRank < 2 && inRank > 0) {
unpackedCoordsSnippet = 'coords';
}
else {
unpackedCoordsSnippet = inputInfo.shapeInfo.logicalShape
.map((s, i) => `coords.${fields[i + rankDiff]}`)
.join(', ');
}
return `
float ${funcName}() {
${type} coords = getOutputCoords();
${coordsSnippet}
return get${texFuncSnippet}(${unpackedCoordsSnippet});
}
`;
}
function getCoordsDataType(rank) {
if (rank <= 1) {
return 'int';
}
else if (rank === 2) {
return 'ivec2';
}
else if (rank === 3) {
return 'ivec3';
}
else if (rank === 4) {
return 'ivec4';
}
else if (rank === 5) {
return 'ivec5';
}
else if (rank === 6) {
return 'ivec6';
}
else {
throw Error(`GPU for rank ${rank} is not yet supported`);
}
}
function getUniformInfoFromShape(isPacked, shape, texShape) {
const { newShape, keptDims } = squeezeShape(shape);
const rank = shape.length;
const useSqueezePackedShape = isPacked && rank === 3 && shape[0] === 1;
const squeezeShape$1 = useSqueezePackedShape ? shape.slice(1) : newShape;
const useSqueezeShape = (!isPacked && rank > 1 && !arraysEqual(shape, texShape) &&
newShape.length < rank) ||
useSqueezePackedShape;
const uniformShape = useSqueezeShape ? squeezeShape$1 : shape;
return { useSqueezeShape, uniformShape, keptDims };
}
/** Returns a new input info (a copy) that has a squeezed logical shape. */
function squeezeInputInfo(inInfo, squeezedShape) {
// Deep copy.
const newInputInfo = JSON.parse(JSON.stringify(inInfo));
newInputInfo.shapeInfo.logicalShape = squeezedShape;
return newInputInfo;
}
function getSqueezedParams(params, keptDims) {
return keptDims.map(d => params[d]).join(', ');
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function compileProgram(gpgpu, program, inputs, output) {
const inputInfos = inputs.map((input, i) => {
const shapeInfo = {
logicalShape: input.shape,
texShape: input.isUniform ? null : input.texData.texShape,
isUniform: input.isUniform,
isPacked: input.isUniform ? false : input.texData.isPacked,
flatOffset: null
};
if (input.texData != null && input.texData.slice != null &&
input.texData.slice.flatOffset > 0) {
shapeInfo.flatOffset = input.texData.slice.flatOffset;
}
return { name: program.variableNames[i], shapeInfo };
});
const inShapeInfos = inputInfos.map(x => x.shapeInfo);
const outShapeInfo = {
logicalShape: output.shape,
texShape: output.texData.texShape,
isUniform: false,
isPacked: output.texData.isPacked,
flatOffset: null
};
const source = makeShader(inputInfos, outShapeInfo, program);
const fragmentShader = createFragmentShader(gpgpu.gl, source);
const webGLProgram = gpgpu.createProgram(fragmentShader);
if (!env().get('ENGINE_COMPILE_ONLY')) {
gpgpu.buildVao(webGLProgram);
return Object.assign({ program,
fragmentShader,
source,
webGLProgram,
inShapeInfos,
outShapeInfo }, getUniformLocations(gpgpu, program, webGLProgram));
}
else {
return {
program,
fragmentShader,
source,
webGLProgram,
inShapeInfos,
outShapeInfo,
variablesLocations: null,
customUniformLocations: null,
infLoc: null,
nanLoc: null,
outShapeLocation: null,
outShapeStridesLocation: null,
outTexShapeLocation: null
};
}
}
function getUniformLocations(gpgpu, program, webGLProgram) {
const variablesLocations = [];
const customUniformLocations = [];
let outShapeLocation;
let outTexShapeLocation;
let outShapeStridesLocation;
let infLoc = null;
let nanLoc = null;
// Add special uniforms (NAN, INFINITY)
nanLoc = gpgpu.getUniformLocation(webGLProgram, 'NAN', false);
if (env().getNumber('WEBGL_VERSION') === 1) {
infLoc = gpgpu.getUniformLocation(webGLProgram, 'INFINITY', false);
}
// Add user-defined uniforms
const shouldThrow = false;
for (const varName of program.variableNames) {
const varLocs = {
name: varName,
uniform: gpgpu.getUniformLocation(webGLProgram, varName, shouldThrow),
offset: gpgpu.getUniformLocation(webGLProgram, `offset${varName}`, shouldThrow),
};
if (program.enableShapeUniforms) {
varLocs.shape = gpgpu.getUniformLocation(webGLProgram, `${varName}Shape`, shouldThrow);
varLocs.texShape = gpgpu.getUniformLocation(webGLProgram, `${varName}TexShape`, shouldThrow);
}
variablesLocations.push(varLocs);
}
if (program.enableShapeUniforms) {
outShapeLocation =
gpgpu.getUniformLocation(webGLProgram, 'outShape', shouldThrow);
outShapeStridesLocation =
gpgpu.getUniformLocation(webGLProgram, 'outShapeStrides', shouldThrow);
outTexShapeLocation =
gpgpu.getUniformLocation(webGLProgram, 'outTexShape', shouldThrow);
}
if (program.customUniforms) {
for (const d of program.customUniforms) {
customUniformLocations.push(gpgpu.getUniformLocation(webGLProgram, d.name, shouldThrow));
}
}
return {
variablesLocations,
customUniformLocations,
infLoc,
nanLoc,
outShapeLocation,
outShapeStridesLocation,
outTexShapeLocation
};
}
function validateBinaryAndProgram(shapeInfos, inputs) {
if (shapeInfos.length !== inputs.length) {
throw Error(`Binary was compiled with ${shapeInfos.length} inputs, but ` +
`was executed with ${inputs.length} inputs`);
}
shapeInfos.forEach((s, i) => {
const shapeA = s.logicalShape;
const input = inputs[i];
const shapeB = input.shape;
if (!arraysEqual(shapeA, shapeB)) {
throw Error(`Binary was compiled with different shapes than ` +
`the current args. Shapes ${shapeA} and ${shapeB} must match`);
}
// The input is uploaded as uniform.
if (s.isUniform && input.isUniform) {
return;
}
const texShapeA = s.texShape;
const texShapeB = input.isUniform ? null : input.texData.texShape;
if (!arraysEqual(texShapeA, texShapeB)) {
throw Error(`Binary was compiled with different texture shapes than the` +
` current args. Shape ${texShapeA} and ${texShapeB} must match`);
}
});
}
function runProgram(gpgpu, binary, inputs, output, customUniformValues) {
if (!binary.program.enableShapeUniforms) {
validateBinaryAndProgram(binary.inShapeInfos, inputs);
validateBinaryAndProgram([binary.outShapeInfo], [output]);
}
const outTex = output.texData.texture;
const outTexShape = output.texData.texShape;
if (output.texData.isPacked) {
gpgpu.setOutputPackedMatrixTexture(outTex.texture, outTexShape[0], outTexShape[1]);
}
else {
gpgpu.setOutputMatrixTexture(outTex.texture, outTexShape[0], outTexShape[1]);
}
gpgpu.setProgram(binary.webGLProgram);
gpgpu.bindVertexArray(binary.webGLProgram.vao);
// Set special uniforms (NAN, INFINITY)
if (env().getNumber('WEBGL_VERSION') === 1) {
if (binary.infLoc !== null) {
gpgpu.gl.uniform1f(binary.infLoc, Infinity);
}
}
if (binary.nanLoc !== null) {
gpgpu.gl.uniform1f(binary.nanLoc, NaN);
}
// Set user-defined inputs
for (let i = 0; i < inputs.length; ++i) {
const input = inputs[i];
const { uniform: varLoc, offset: varOffsetLoc, shape: varShapeLoc, texShape: varTexShapeLoc, } = binary.variablesLocations[i];
if (varShapeLoc) {
const { uniformShape } = getUniformInfoFromShape(binary.program.packedInputs, input.shape, input.texData.texShape);
switch (uniformShape.length) {
case 1:
gpgpu.gl.uniform1iv(varShapeLoc, new Int32Array(uniformShape));
break;
case 2:
gpgpu.gl.uniform2iv(varShapeLoc, new Int32Array(uniformShape));
break;
case 3:
gpgpu.gl.uniform3iv(varShapeLoc, new Int32Array(uniformShape));
break;
case 4:
gpgpu.gl.uniform4iv(varShapeLoc, new Int32Array(uniformShape));
break;
}
}
if (varTexShapeLoc) {
gpgpu.gl.uniform2i(varTexShapeLoc, input.texData.texShape[0], input.texData.texShape[1]);
}
if (varLoc == null) {
// The compiler inferred that this variable is not used in this shader.
continue;
}
if (input.isUniform) {
// Upload the values of the tensor as uniform.
if (sizeFromShape(input.shape) < 2) {
gpgpu.gl.uniform1f(varLoc, input.uniformValues[0]);
}
else {
let vals = input.uniformValues;
if (!(vals instanceof Float32Array)) {
vals = new Float32Array(vals);
}
gpgpu.gl.uniform1fv(varLoc, vals);
}
continue;
}
// If the input was sliced, upload the flat offset index.
if (input.texData.slice != null && varOffsetLoc != null) {
gpgpu.gl.uniform1i(varOffsetLoc, input.texData.slice.flatOffset);
}
gpgpu.setInputMatrixTexture(input.texData.texture.texture, varLoc, i);
}
const outShapeLoc = binary.outShapeLocation;
if (outShapeLoc) {
switch (output.shape.length) {
case 1:
gpgpu.gl.uniform1iv(outShapeLoc, new Int32Array(output.shape));
break;
case 2:
gpgpu.gl.uniform2iv(outShapeLoc, new Int32Array(output.shape));
break;
case 3:
gpgpu.gl.uniform3iv(outShapeLoc, new Int32Array(output.shape));
break;
case 4:
gpgpu.gl.uniform4iv(outShapeLoc, new Int32Array(output.shape));
break;
}
}
if (binary.outShapeStridesLocation) {
const strides = computeStrides(output.shape);
switch (output.shape.length) {
case 2:
gpgpu.gl.uniform1iv(binary.outShapeStridesLocation, new Int32Array(strides));
break;
case 3:
gpgpu.gl.uniform2iv(binary.outShapeStridesLocation, new Int32Array(strides));
break;
case 4:
gpgpu.gl.uniform3iv(binary.outShapeStridesLocation, new Int32Array(strides));
break;
}
}
if (binary.outTexShapeLocation) {
gpgpu.gl.uniform2i(binary.outTexShapeLocation, output.texData.texShape[0], output.texData.texShape[1]);
}
if (binary.program.customUniforms && customUniformValues) {
for (let i = 0; i < binary.program.customUniforms.length; ++i) {
const d = binary.program.customUniforms[i];
const customLoc = binary.customUniformLocations[i];
const customValue = customUniformValues[i];
if (d.type === 'float') {
gpgpu.gl.uniform1fv(customLoc, customValue);
}
else if (d.type === 'vec2') {
gpgpu.gl.uniform2fv(customLoc, customValue);
}
else if (d.type === 'vec3') {
gpgpu.gl.uniform3fv(customLoc, customValue);
}
else if (d.type === 'vec4') {
gpgpu.gl.uniform4fv(customLoc, customValue);
}
else if (d.type === 'int') {
gpgpu.gl.uniform1iv(customLoc, customValue);
}
else if (d.type === 'ivec2') {
gpgpu.gl.uniform2iv(customLoc, customValue);
}
else if (d.type === 'ivec3') {
gpgpu.gl.uniform3iv(customLoc, customValue);
}
else if (d.type === 'ivec4') {
gpgpu.gl.uniform4iv(customLoc, customValue);
}
else {
throw Error(`uniform type ${d.type} is not supported yet.`);
}
}
}
gpgpu.executeProgram();
}
function makeShaderKey(program, inputs, output) {
let keyInputs = '';
inputs.concat(output).forEach(x => {
const hasOffset = x.texData != null && x.texData.slice != null &&
x.texData.slice.flatOffset > 0;
// TODO: Remove the condition of !x.isUniform.
if (program.enableShapeUniforms && !x.isUniform) {
const xTexShape = x.texData.texShape;
const { useSqueezeShape, uniformShape, keptDims } = getUniformInfoFromShape(program.packedInputs, x.shape, xTexShape);
let rank1 = '', rank2 = '', rank34 = '';
if (uniformShape.length === 1 && program.packedInputs) {
const packedTexShape = [Math.ceil(xTexShape[0] / 2), Math.ceil(xTexShape[1] / 2)];
rank1 = `${packedTexShape[0] > 1}_${packedTexShape[1] > 1}`;
}
else if (uniformShape.length === 2 && !program.packedInputs) {
rank2 = `${uniformShape[0] > 1}_${uniformShape[1] > 1}`;
}
else if (uniformShape.length > 2 && !program.packedInputs) {
const strides = computeStrides(uniformShape);
rank34 = `${strides[0] === xTexShape[1]}_${strides[strides.length - 1] === xTexShape[1]}`;
}
const xRank = x.shape.length;
const isLogicalShapTexShapeEqual = uniformShape.length === 2 && arraysEqual(x.shape, xTexShape);
const isScalar = sizeFromShape(x.shape) === 1;
const broadcastDims = getBroadcastDims$1(x.shape, output.shape);
const isInOutTexShapeEqual = !program.packedInputs &&
xRank === output.shape.length &&
arraysEqual(xTexShape, output.texData.texShape);
const isTexShapeGreaterThanOne = program.packedInputs || uniformShape.length > 2 ?
'' :
`${xTexShape[0] > 1}_${xTexShape[1] > 1}`;
// These key components are needed due to shader_compiler is embedding
// them in the shader.
// |xRank| is used to determine the coords length. See
// get[Packed]SamplerAtOutputCoords.
// |isInOutTexShapeEqual| is used to determine whether going to an
// optimization path in getSamplerAtOutputCoords.
// |useSqueezeShape| is extracted from squeezeInputInfo of
// getSampler[2|3|4]D/getPackedSampler3D.
// |isScalar| is extracted from isInputScalar/isOutputScalar in
// getPackedSamplerAtOutputCoords.
// |broadcastDims| is extracted from get[Packed]SamplerAtOutputCoords.
// |isLogicalShapTexShapeEqual| is used in
// getOutput[Packed]2DCoords/get[Packed]Sampler2D.
// |rank1| is used in getOutputPacked1DCoords.
// |rank2| is used in getOutput2DCoords.
// |rank34| is used in getSampler3D/getSampler4D.
// |isTexShapeGreaterThanOne| are used in
// getSampler[Scalar|1D|2D]/getOutput1DCoords.
keyInputs += `${xRank}_${isInOutTexShapeEqual}_${useSqueezeShape ? keptDims : ''}_${uniformShape.length}_${isScalar}_${broadcastDims}_${isLogicalShapTexShapeEqual}_${rank1}_${rank2}_${rank34}_${isTexShapeGreaterThanOne}_${hasOffset}`;
}
else {
const texShape = x.isUniform ? 'uniform' : x.texData.texShape;
keyInputs += `${x.shape}_${texShape}_${hasOffset}`;
}
});
const keyUserCode = program.userCode;
let key = program.constructor.name;
// Fast string concat. See https://jsperf.com/string-concatenation/14.
key += '_' + keyInputs + '_' + keyUserCode +
`${env().getNumber('WEBGL_VERSION')}`;
return key;
}
function useShapeUniforms(rank) {
// TODO: Remove the limitaion of rank <= 4.
return env().getBool('WEBGL_USE_SHAPES_UNIFORMS') && rank <= 4;
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class DecodeMatrixProgram {
constructor(outputShape) {
this.variableNames = ['A'];
this.packedInputs = false;
this.packedOutput = true;
this.outPackingScheme = PackingScheme.DENSE;
this.customUniforms = [{ name: 'texShape', type: 'ivec2' }];
const glsl = getGlslDifferences();
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
this.userCode = `
ivec3 outCoordsFromFlatIndex(int index) {
${this.enableShapeUniforms ?
getOutputLogicalCoordinatesFromFlatIndexByUniform(['r', 'c', 'd'], outputShape) :
getLogicalCoordinatesFromFlatIndex(['r', 'c', 'd'], outputShape)}
return ivec3(r, c, d);
}
void main() {
ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);
vec4 result = vec4(0.);
for (int i=0; i<4; i++) {
int flatIndex = index + i;
ivec3 rc = outCoordsFromFlatIndex(flatIndex);
result[i] = getA(rc.x, rc.y, rc.z);
}
${glsl.output} = result;
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class DecodeMatrixPackedProgram {
constructor(outputShape) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
this.outPackingScheme = PackingScheme.DENSE;
this.customUniforms = [{ name: 'texShape', type: 'ivec2' }];
const glsl = getGlslDifferences();
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
this.userCode = `
ivec3 outCoordsFromFlatIndex(int index) {
${this.enableShapeUniforms ?
getOutputLogicalCoordinatesFromFlatIndexByUniform(['r', 'c', 'd'], outputShape) :
getLogicalCoordinatesFromFlatIndex(['r', 'c', 'd'], outputShape)}
return ivec3(r, c, d);
}
void main() {
ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);
vec4 result = vec4(0.);
for (int i=0; i<4; i++) {
int flatIndex = index + i;
ivec3 rc = outCoordsFromFlatIndex(flatIndex);
result[i] = getChannel(getA(rc.x, rc.y, rc.z), vec2(rc.y, rc.z));
}
${glsl.output} = result;
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class EncodeFloatProgram {
constructor(outputShape) {
this.variableNames = ['A'];
this.outTexUsage = TextureUsage.DOWNLOAD;
const glsl = getGlslDifferences();
this.outputShape = outputShape;
this.userCode = `
${ENCODE_FLOAT_SNIPPET}
void main() {
float x = getAAtOutCoords();
${glsl.output} = encode_float(x);
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class EncodeFloatPackedProgram {
constructor(outputShape) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = false;
this.outTexUsage = TextureUsage.DOWNLOAD;
const glsl = getGlslDifferences();
this.outputShape = outputShape;
this.userCode = `
${ENCODE_FLOAT_SNIPPET}
void main() {
ivec3 coords = getOutputCoords();
float x = getChannel(getAAtOutCoords(), vec2(coords.y, coords.z));
${glsl.output} = encode_float(x);
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const CHANNEL_CHAR_TO_INDEX_MAP = {
'R': 0,
'G': 1,
'B': 2,
'A': 3
};
class EncodeMatrixProgram {
constructor(outputShape, inputIsUnsignedByte = false, usedChannels = 'RGBA') {
this.variableNames = ['A'];
this.customUniforms = [{ name: 'texShape', type: 'ivec2' }];
const glsl = getGlslDifferences();
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
let output = `result`;
if (inputIsUnsignedByte) {
output = `floor(result * 255. + 0.5)`;
}
let mainLoop = '';
for (let usedChannelIndex = 0; usedChannelIndex < usedChannels.length; usedChannelIndex++) {
const curChannel = usedChannels[usedChannelIndex];
mainLoop += `
if(offset == ${usedChannelIndex}) {
result = values[${CHANNEL_CHAR_TO_INDEX_MAP[curChannel]}];
}`;
}
this.userCode = `
${this.enableShapeUniforms ? getFlatIndexFrom3DOutput() :
getFlatIndexFrom3D(outputShape)}
void main() {
ivec3 coords = getOutputCoords();
int flatIndex = getFlatIndex(coords);
float result = 0.;
int offset = imod(flatIndex, ${usedChannels.length});
flatIndex = idiv(flatIndex, ${usedChannels.length}, 1.);
int r = flatIndex / texShape[1];
if (r < texShape[0]) {
int c = imod(flatIndex, texShape[1]);
vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
vec4 values = ${glsl.texture2D}(A, uv);
${mainLoop}
}
${glsl.output} = vec4(${output}, 0., 0., 0.);
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/*
This is how the shader encodes a tensor with shape = [2, 3, 5]
(indices are [batch, row, col]).
000|001 002|003 004|xxx 020|021 022|023 024|xxx
------- ------- ------- ------- ------- -------
010|011 012|013 014|xxx xxx|xxx xxx|xxx xxx|xxx
100|101 102|103 104|xxx 120|121 122|123 124|xxx
------- ------- ------- ------- ------- -------
110|111 112|113 114|xxx xxx|xxx xxx|xxx xxx|xxx
Single texels contain only values from the same batch, and from adjacent rows
and columns.
*/
class EncodeMatrixPackedProgram {
constructor(outputShape, inputIsUnsignedByte = false) {
this.variableNames = ['A'];
this.packedInputs = false;
this.packedOutput = true;
this.customUniforms = [{ name: 'texShape', type: 'ivec2' }];
const glsl = getGlslDifferences();
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
let mainLoop = '';
let output = 'result';
if (inputIsUnsignedByte) {
output = 'floor(result * 255. + 0.5)';
}
for (let row = 0; row <= 1; row++) {
for (let col = 0; col <= 1; col++) {
const channel = row * 2 + col;
mainLoop += `
localCoords = coords;
if(localCoords[2] + ${col} < ${this.enableShapeUniforms ? 'outShape[2]' : `${outputShape[2]}`}) {
localCoords[2] += ${col};
if (localCoords[1] + ${row} < ${this.enableShapeUniforms ? 'outShape[1]' : `${outputShape[1]}`}) {
localCoords[1] += ${row};
flatIndex = getFlatIndex(localCoords);
offset = imod(flatIndex, 4);
flatIndex = idiv(flatIndex, 4, 1.);
int r = flatIndex / texShape[1];
int c = imod(flatIndex, texShape[1]);
vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
values = ${glsl.texture2D}(A, uv);
if (offset == 0) {
result[${channel}] = values[0];
} else if (offset == 1) {
result[${channel}] = values[1];
} else if (offset == 2) {
result[${channel}] = values[2];
} else {
result[${channel}] = values[3];
}
}
}
`;
}
}
this.userCode = `
${this.enableShapeUniforms ? getFlatIndexFrom3DOutput() :
getFlatIndexFrom3D(outputShape)}
void main() {
ivec3 coords = getOutputCoords();
vec4 result = vec4(0.);
int flatIndex, r, c, offset;
ivec3 localCoords;
vec2 uv;
vec4 values;
${mainLoop}
${glsl.output} = ${output};
}
`;
}
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function createVertexShader(gl) {
const glsl = getGlslDifferences();
const vertexShaderSource = `${glsl.version}
precision highp float;
${glsl.attribute} vec3 clipSpacePos;
${glsl.attribute} vec2 uv;
${glsl.varyingVs} vec2 resultUV;
void main() {
gl_Position = vec4(clipSpacePos, 1);
resultUV = uv;
}`;
return createVertexShader$1(gl, vertexShaderSource);
}
function createVertexBuffer(gl) {
// [x y z u v] * [upper-left, lower-left, upper-right, lower-right]
const vertexArray = new Float32Array([-1, 1, 0, 0, 1, -1, -1, 0, 0, 0, 1, 1, 0, 1, 1, 1, -1, 0, 1, 0]);
return createStaticVertexBuffer(gl, vertexArray);
}
function createIndexBuffer(gl) {
// OpenGL (and WebGL) have "CCW == front" winding
const triangleVertexIndices = new Uint16Array([0, 1, 2, 2, 1, 3]);
return createStaticIndexBuffer(gl, triangleVertexIndices);
}
function createAndConfigureTexture(gl, width, height, internalFormat, textureFormat, textureType) {
validateTextureSize(width, height);
const texture = createTexture(gl);
const tex2d = gl.TEXTURE_2D;
callAndCheck(gl, () => gl.bindTexture(tex2d, texture));
callAndCheck(gl, () => gl.texParameteri(tex2d, gl.TEXTURE_WRAP_S, gl.CLAMP_TO_EDGE));
callAndCheck(gl, () => gl.texParameteri(tex2d, gl.TEXTURE_WRAP_T, gl.CLAMP_TO_EDGE));
callAndCheck(gl, () => gl.texParameteri(tex2d, gl.TEXTURE_MIN_FILTER, gl.NEAREST));
callAndCheck(gl, () => gl.texParameteri(tex2d, gl.TEXTURE_MAG_FILTER, gl.NEAREST));
if (env().getNumber('WEBGL_VERSION') === 1) {
callAndCheck(gl, () => gl.texImage2D(tex2d, 0, internalFormat, width, height, 0, textureFormat, textureType, null));
}
else {
callAndCheck(gl, () => gl
.texStorage2D(tex2d, 1, internalFormat, width, height));
}
callAndCheck(gl, () => gl.bindTexture(gl.TEXTURE_2D, null));
return { texture, texShape: [height, width] };
}
function getInternalFormatForFloat32MatrixTexture(textureConfig) {
return textureConfig.internalFormatFloat;
}
function createFloat32MatrixTexture(gl, rows, columns, textureConfig) {
const [width, height] = getUnpackedMatrixTextureShapeWidthHeight(rows, columns);
return createAndConfigureTexture(gl, width, height, getInternalFormatForFloat32MatrixTexture(textureConfig), textureConfig.textureFormatFloat, gl.FLOAT);
}
function getInternalFormatForFloat16MatrixTexture(textureConfig) {
return textureConfig.internalFormatHalfFloat;
}
function createFloat16MatrixTexture(gl, rows, columns, textureConfig) {
const [width, height] = getUnpackedMatrixTextureShapeWidthHeight(rows, columns);
return createAndConfigureTexture(gl, width, height, getInternalFormatForFloat16MatrixTexture(textureConfig), textureConfig.textureFormatFloat, textureConfig.textureTypeHalfFloat);
}
function getInternalFormatForUnsignedBytesMatrixTexture(textureConfig) {
return textureConfig.downloadTextureFormat;
}
function createUnsignedBytesMatrixTexture(gl, rows, columns, textureConfig) {
const [width, height] = getUnpackedMatrixTextureShapeWidthHeight(rows, columns);
return createAndConfigureTexture(gl, width, height, getInternalFormatForUnsignedBytesMatrixTexture(textureConfig), gl.RGBA, gl.UNSIGNED_BYTE);
}
function getInternalFormatForPackedMatrixTexture(textureConfig) {
return textureConfig.internalFormatPackedFloat;
}
function createPackedMatrixTexture(gl, rows, columns, textureConfig) {
const [width, height] = getPackedMatrixTextureShapeWidthHeight(rows, columns);
return createAndConfigureTexture(gl, width, height, getInternalFormatForPackedMatrixTexture(textureConfig), gl.RGBA, gl.FLOAT);
}
function getInternalFormatForFloat16PackedMatrixTexture(textureConfig) {
return textureConfig.internalFormatPackedHalfFloat;
}
function createFloat16PackedMatrixTexture(gl, rows, columns, textureConfig) {
const [width, height] = getPackedMatrixTextureShapeWidthHeight(rows, columns);
return createAndConfigureTexture(gl, width, height, getInternalFormatForFloat16PackedMatrixTexture(textureConfig), gl.RGBA, textureConfig.textureTypeHalfFloat);
}
function bindVertexProgramAttributeStreams(gl, program, vertexBuffer) {
const posOffset = 0; // x is the first buffer element
const uvOffset = 3 * 4; // uv comes after [x y z]
const stride = (3 * 4) + (2 * 4); // xyz + uv, each entry is 4-byte float.
callAndCheck(gl, () => gl.bindBuffer(gl.ARRAY_BUFFER, vertexBuffer));
const success = bindVertexBufferToProgramAttribute(gl, program, 'clipSpacePos', vertexBuffer, 3, stride, posOffset);
return success &&
bindVertexBufferToProgramAttribute(gl, program, 'uv', vertexBuffer, 2, stride, uvOffset);
}
function uploadDenseMatrixToTexture(gl, texture, width, height, data, textureConfig) {
callAndCheck(gl, () => gl.bindTexture(gl.TEXTURE_2D, texture));
let dataForUpload, texelDataType, internalFormat;
if (data instanceof Uint8Array) {
dataForUpload = new Uint8Array(width * height * 4);
texelDataType = gl.UNSIGNED_BYTE;
internalFormat = gl.RGBA;
}
else {
dataForUpload = new Float32Array(width * height * 4);
texelDataType = gl.FLOAT;
internalFormat = textureConfig.internalFormatPackedFloat;
}
dataForUpload.set(data);
if (env().getNumber('WEBGL_VERSION') === 2) {
callAndCheck(gl, () => gl.texSubImage2D(gl.TEXTURE_2D, 0, 0, 0, width, height, gl.RGBA, texelDataType, dataForUpload));
}
else {
callAndCheck(gl, () => gl.texImage2D(gl.TEXTURE_2D, 0, internalFormat, width, height, 0, gl.RGBA, texelDataType, dataForUpload));
}
callAndCheck(gl, () => gl.bindTexture(gl.TEXTURE_2D, null));
}
function uploadPixelDataToTexture(gl, texture, pixels) {
callAndCheck(gl, () => gl.bindTexture(gl.TEXTURE_2D, texture));
if (pixels.data instanceof Uint8Array) {
if (env().getNumber('WEBGL_VERSION') === 2) {
callAndCheck(gl, () => gl.texSubImage2D(gl.TEXTURE_2D, 0, 0, 0, pixels.width, pixels.height, gl.RGBA, gl.UNSIGNED_BYTE, pixels.data));
}
else {
callAndCheck(gl, () => gl.texImage2D(gl.TEXTURE_2D, 0, gl.RGBA, pixels.width, pixels.height, 0, gl.RGBA, gl.UNSIGNED_BYTE, pixels.data));
}
}
else {
if (env().getNumber('WEBGL_VERSION') === 2) {
callAndCheck(gl, () => gl.texSubImage2D(gl.TEXTURE_2D, 0, 0, 0, gl.RGBA, gl.UNSIGNED_BYTE, pixels));
}
else {
callAndCheck(gl, () => gl.texImage2D(gl.TEXTURE_2D, 0, gl.RGBA, gl.RGBA, gl.UNSIGNED_BYTE, pixels));
}
}
callAndCheck(gl, () => gl.bindTexture(gl.TEXTURE_2D, null));
}
function createBufferFromOutputTexture(gl2, rows, columns, textureConfig) {
// Create and bind the buffer.
const buffer = gl2.createBuffer();
callAndCheck(gl2, () => gl2.bindBuffer(gl2.PIXEL_PACK_BUFFER, buffer));
// Initialize the buffer to the size of the texture in bytes.
const bytesPerFloat = 4;
const valuesPerTexel = 4;
const bufferSizeBytes = bytesPerFloat * valuesPerTexel * rows * columns;
callAndCheck(gl2, () => gl2.bufferData(gl2.PIXEL_PACK_BUFFER, bufferSizeBytes, gl2.STREAM_READ));
// Enqueue a command on the GPU command queue to copy of texture into the
// buffer.
callAndCheck(gl2, () => gl2.readPixels(0, 0, columns, rows, gl2.RGBA, gl2.FLOAT, 0));
callAndCheck(gl2, () => gl2.bindBuffer(gl2.PIXEL_PACK_BUFFER, null));
return buffer;
}
function downloadFloat32MatrixFromBuffer(gl, buffer, size) {
const gl2 = gl;
const downloadTarget = new Float32Array(size);
gl2.bindBuffer(gl2.PIXEL_PACK_BUFFER, buffer);
gl2.getBufferSubData(gl2.PIXEL_PACK_BUFFER, 0, downloadTarget);
gl2.bindBuffer(gl2.PIXEL_PACK_BUFFER, null);
return downloadTarget;
}
function downloadByteEncodedFloatMatrixFromOutputTexture(gl, rows, columns, textureConfig) {
const [w, h] = getUnpackedMatrixTextureShapeWidthHeight(rows, columns);
const numChannels = 4;
const downloadTarget = new Uint8Array(getUnpackedArraySizeFromMatrixSize(rows * columns, numChannels));
callAndCheck(gl, () => gl.readPixels(0, 0, w, h, textureConfig.downloadTextureFormat, gl.UNSIGNED_BYTE, downloadTarget));
// By wrapping the buffer in a Float32Array, we use native browser IEEE 754
// decoding of the 4 bytes that back each 32 bit float.
return new Float32Array(downloadTarget.buffer);
}
function downloadPackedMatrixFromBuffer(gl, buffer, batch, rows, cols, physicalRows, physicalCols, textureConfig) {
const gl2 = gl;
const downloadTarget = new Float32Array(getPackedRGBAArraySizeFromMatrixShape(physicalRows, physicalCols));
gl2.bindBuffer(gl2.PIXEL_PACK_BUFFER, buffer);
gl2.getBufferSubData(gl2.PIXEL_PACK_BUFFER, 0, downloadTarget);
gl2.bindBuffer(gl2.PIXEL_PACK_BUFFER, null);
return downloadTarget;
}
function downloadMatrixFromPackedOutputTexture(gl, physicalRows, physicalCols) {
const packedRGBA = new Float32Array(physicalRows * physicalCols * 4);
callAndCheck(gl, () => gl.readPixels(0, 0, physicalCols, physicalRows, gl.RGBA, gl.FLOAT, packedRGBA));
return packedRGBA;
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class GPGPUContext {
constructor(gl) {
this.outputTexture = null;
this.program = null;
this.disposed = false;
this.itemsToPoll = [];
const glVersion = env().getNumber('WEBGL_VERSION');
if (gl != null) {
this.gl = gl;
setWebGLContext(glVersion, gl);
}
else {
this.gl = getWebGLContext(glVersion);
}
gl = this.gl;
if (env().getNumber('WEBGL_VERSION') === 2) {
const gl2 = gl;
this.createVertexArray = () => {
return callAndCheck(gl2, () => gl2.createVertexArray());
};
this.bindVertexArray = (vao) => {
return callAndCheck(gl2, () => gl2.bindVertexArray(vao));
};
this.deleteVertexArray = (vao) => {
return callAndCheck(gl2, () => gl2.deleteVertexArray(vao));
};
this.getVertexArray = () => {
return callAndCheck(gl2, () => gl2.getParameter(gl2.VERTEX_ARRAY_BINDING));
};
}
else if (gl != null) {
const ext = gl.getExtension('OES_vertex_array_object');
if (ext == null) {
throw new Error('All WebGL1 implementations are expected to offer' +
' OES_vertex_array_object.');
}
this.createVertexArray = () => {
return callAndCheck(gl, () => ext.createVertexArrayOES());
};
this.bindVertexArray = (vao) => {
return callAndCheck(gl, () => ext.bindVertexArrayOES(vao));
};
this.deleteVertexArray = (vao) => {
return callAndCheck(gl, () => ext.deleteVertexArrayOES(vao));
};
this.getVertexArray = () => {
return callAndCheck(gl, () => gl.getParameter(ext.VERTEX_ARRAY_BINDING_OES));
};
}
// WebGL 2.0 enables texture floats without an extension.
let COLOR_BUFFER_FLOAT = 'WEBGL_color_buffer_float';
const COLOR_BUFFER_HALF_FLOAT = 'EXT_color_buffer_half_float';
this.parallelCompilationExtension =
this.gl.getExtension('KHR_parallel_shader_compile');
if (env().getNumber('WEBGL_VERSION') === 1) {
const TEXTURE_FLOAT = 'OES_texture_float';
const TEXTURE_HALF_FLOAT = 'OES_texture_half_float';
this.textureFloatExtension =
getExtensionOrThrow(this.gl, TEXTURE_FLOAT);
if (hasExtension(this.gl, TEXTURE_HALF_FLOAT)) {
this.textureHalfFloatExtension =
getExtensionOrThrow(this.gl, TEXTURE_HALF_FLOAT);
}
else if (env().get('WEBGL_FORCE_F16_TEXTURES')) {
throw new Error('GL context does not support half float textures, yet the ' +
'environment flag WEBGL_FORCE_F16_TEXTURES is set to true.');
}
this.colorBufferFloatExtension = this.gl.getExtension(COLOR_BUFFER_FLOAT);
if (hasExtension(this.gl, COLOR_BUFFER_HALF_FLOAT)) {
this.colorBufferHalfFloatExtension =
getExtensionOrThrow(this.gl, COLOR_BUFFER_HALF_FLOAT);
}
else if (env().get('WEBGL_FORCE_F16_TEXTURES')) {
throw new Error('GL context does not support color renderable half floats, yet ' +
'the environment flag WEBGL_FORCE_F16_TEXTURES is set to true.');
}
}
else {
COLOR_BUFFER_FLOAT = 'EXT_color_buffer_float';
if (hasExtension(this.gl, COLOR_BUFFER_FLOAT)) {
this.colorBufferFloatExtension =
this.gl.getExtension(COLOR_BUFFER_FLOAT);
}
else if (hasExtension(this.gl, COLOR_BUFFER_HALF_FLOAT)) {
this.colorBufferHalfFloatExtension =
this.gl.getExtension(COLOR_BUFFER_HALF_FLOAT);
}
else {
throw new Error('GL context does not support color renderable floats');
}
}
this.vertexBuffer = createVertexBuffer(this.gl);
this.indexBuffer = createIndexBuffer(this.gl);
this.framebuffer = createFramebuffer(this.gl);
this.textureConfig =
getTextureConfig(this.gl, this.textureHalfFloatExtension);
}
get debug() {
return env().getBool('DEBUG');
}
dispose() {
if (this.disposed) {
return;
}
if (this.program != null) {
console.warn('Disposing a GPGPUContext that still has a bound WebGLProgram.' +
' This is probably a resource leak, delete the program with ' +
'GPGPUContext.deleteProgram before disposing.');
}
if (this.outputTexture != null) {
console.warn('Disposing a GPGPUContext that still has a bound output matrix ' +
'texture. This is probably a resource leak, delete the output ' +
'matrix texture with GPGPUContext.deleteMatrixTexture before ' +
'disposing.');
}
const gl = this.gl;
callAndCheck(gl, () => gl.finish());
callAndCheck(gl, () => gl.bindFramebuffer(gl.FRAMEBUFFER, null));
callAndCheck(gl, () => gl.deleteFramebuffer(this.framebuffer));
callAndCheck(gl, () => gl.bindBuffer(gl.ARRAY_BUFFER, null));
callAndCheck(gl, () => gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, null));
callAndCheck(gl, () => gl.deleteBuffer(this.indexBuffer));
this.disposed = true;
}
createFloat32MatrixTexture(rows, columns) {
this.throwIfDisposed();
return createFloat32MatrixTexture(this.gl, rows, columns, this.textureConfig);
}
createFloat16MatrixTexture(rows, columns) {
this.throwIfDisposed();
return createFloat16MatrixTexture(this.gl, rows, columns, this.textureConfig);
}
createUnsignedBytesMatrixTexture(rows, columns) {
this.throwIfDisposed();
return createUnsignedBytesMatrixTexture(this.gl, rows, columns, this.textureConfig);
}
uploadPixelDataToTexture(texture, pixels) {
this.throwIfDisposed();
uploadPixelDataToTexture(this.gl, texture, pixels);
}
uploadDenseMatrixToTexture(texture, width, height, data) {
this.throwIfDisposed();
uploadDenseMatrixToTexture(this.gl, texture, width, height, data, this.textureConfig);
}
createFloat16PackedMatrixTexture(rows, columns) {
this.throwIfDisposed();
return createFloat16PackedMatrixTexture(this.gl, rows, columns, this.textureConfig);
}
createPackedMatrixTexture(rows, columns) {
this.throwIfDisposed();
return createPackedMatrixTexture(this.gl, rows, columns, this.textureConfig);
}
deleteMatrixTexture(texture) {
this.throwIfDisposed();
if (this.outputTexture === texture) {
unbindColorTextureFromFramebuffer(this.gl, this.framebuffer);
this.outputTexture = null;
}
callAndCheck(this.gl, () => this.gl.deleteTexture(texture));
}
downloadByteEncodedFloatMatrixFromOutputTexture(texture, rows, columns) {
return this.downloadMatrixDriver(texture, () => downloadByteEncodedFloatMatrixFromOutputTexture(this.gl, rows, columns, this.textureConfig));
}
downloadPackedMatrixFromBuffer(buffer, batch, rows, columns, physicalRows, physicalCols) {
return downloadPackedMatrixFromBuffer(this.gl, buffer, batch, rows, columns, physicalRows, physicalCols);
}
downloadFloat32MatrixFromBuffer(buffer, size) {
return downloadFloat32MatrixFromBuffer(this.gl, buffer, size);
}
createBufferFromTexture(texture, rows, columns) {
this.bindTextureToFrameBuffer(texture);
const result = createBufferFromOutputTexture(this.gl, rows, columns);
this.unbindTextureToFrameBuffer();
return result;
}
createAndWaitForFence() {
const fenceContext = this.createFence(this.gl);
return this.pollFence(fenceContext);
}
createFence(gl) {
let query;
let isFencePassed;
if (env().getBool('WEBGL_FENCE_API_ENABLED')) {
const gl2 = gl;
const sync = gl2.fenceSync(gl2.SYNC_GPU_COMMANDS_COMPLETE, 0);
gl.flush();
isFencePassed = () => {
const status = gl2.clientWaitSync(sync, 0, 0);
return status === gl2.ALREADY_SIGNALED ||
status === gl2.CONDITION_SATISFIED;
};
query = sync;
}
else if (env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION') > 0) {
query = this.beginQuery();
this.endQuery();
isFencePassed = () => this.isQueryAvailable(query, env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION'));
}
else {
// If we have no way to fence, return true immediately. This will fire in
// WebGL 1.0 when there is no disjoint query timer. In this case, because
// the fence passes immediately, we'll immediately ask for a download of
// the texture, which will cause the UI thread to hang.
isFencePassed = () => true;
}
return { query, isFencePassed };
}
downloadMatrixFromPackedTexture(texture, physicalRows, physicalCols) {
return this.downloadMatrixDriver(texture, () => downloadMatrixFromPackedOutputTexture(this.gl, physicalRows, physicalCols));
}
createProgram(fragmentShader) {
this.throwIfDisposed();
const gl = this.gl;
if (this.vertexShader == null) {
this.vertexShader = createVertexShader(gl);
}
const program = createProgram(gl);
callAndCheck(gl, () => gl.attachShader(program, this.vertexShader));
callAndCheck(gl, () => gl.attachShader(program, fragmentShader));
linkProgram(gl, program);
const program2 = Object.assign(program, { vao: this.createVertexArray() });
if (this.debug) {
validateProgram(gl, program2);
}
return program2;
}
buildVao(program) {
this.setProgram(program);
this.bindVertexArray(program.vao);
const gl = this.gl;
// Bind index buffer, and vertex buffers based on program attrib
// locations.
callAndCheck(gl, () => gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, this.indexBuffer));
bindVertexProgramAttributeStreams(gl, program, this.vertexBuffer);
}
deleteProgram(program) {
this.throwIfDisposed();
if (program === this.program) {
this.program = null;
}
if (program != null) {
callAndCheck(this.gl, () => this.gl.deleteProgram(program));
this.deleteVertexArray(program.vao);
}
}
setProgram(program) {
this.throwIfDisposed();
this.program = program;
if (this.program != null) {
if (this.debug) {
validateProgram(this.gl, this.program);
}
}
callAndCheck(this.gl, () => this.gl.useProgram(program));
}
getUniformLocation(program, uniformName, shouldThrow = true) {
this.throwIfDisposed();
if (shouldThrow) {
return getProgramUniformLocationOrThrow(this.gl, program, uniformName);
}
else {
return getProgramUniformLocation(this.gl, program, uniformName);
}
}
getAttributeLocation(program, attribute) {
this.throwIfDisposed();
return callAndCheck(this.gl, () => this.gl.getAttribLocation(program, attribute));
}
getUniformLocationNoThrow(program, uniformName) {
this.throwIfDisposed();
return this.gl.getUniformLocation(program, uniformName);
}
setInputMatrixTexture(inputMatrixTexture, uniformLocation, textureUnit) {
this.throwIfDisposed();
this.throwIfNoProgram();
bindTextureToProgramUniformSampler(this.gl, inputMatrixTexture, uniformLocation, textureUnit);
}
setOutputMatrixTexture(outputMatrixTexture, rows, columns) {
this.setOutputMatrixTextureDriver(outputMatrixTexture, columns, rows);
}
setOutputPackedMatrixTexture(outputPackedMatrixTexture, rows, columns) {
this.throwIfDisposed();
const [width, height] = getPackedMatrixTextureShapeWidthHeight(rows, columns);
this.setOutputMatrixTextureDriver(outputPackedMatrixTexture, width, height);
}
setOutputMatrixWriteRegion(startRow, numRows, startColumn, numColumns) {
this.setOutputMatrixWriteRegionDriver(startColumn, startRow, numColumns, numRows);
}
setOutputPackedMatrixWriteRegion(startRow, numRows, startColumn, numColumns) {
throw new Error('setOutputPackedMatrixWriteRegion not implemented.');
}
debugValidate() {
if (this.program != null) {
validateProgram(this.gl, this.program);
}
validateFramebuffer(this.gl);
}
executeProgram() {
this.throwIfDisposed();
this.throwIfNoProgram();
const gl = this.gl;
if (this.debug) {
const boundVao = this.getVertexArray();
console.assert(boundVao === this.program.vao, 'VAO changed between setProgram and executeProgram!');
this.debugValidate();
}
callAndCheck(gl, () => gl.drawElements(gl.TRIANGLES, 6, gl.UNSIGNED_SHORT, 0));
}
blockUntilAllProgramsCompleted() {
this.throwIfDisposed();
callAndCheck(this.gl, () => this.gl.finish());
}
getQueryTimerExtension() {
if (this.disjointQueryTimerExtension == null) {
this.disjointQueryTimerExtension =
getExtensionOrThrow(this.gl, env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION') === 2 ?
'EXT_disjoint_timer_query_webgl2' :
'EXT_disjoint_timer_query');
}
return this.disjointQueryTimerExtension;
}
getQueryTimerExtensionWebGL2() {
return this.getQueryTimerExtension();
}
getQueryTimerExtensionWebGL1() {
return this.getQueryTimerExtension();
}
beginQuery() {
if (env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION') === 2) {
const gl2 = this.gl;
const ext = this.getQueryTimerExtensionWebGL2();
const query = gl2.createQuery();
gl2.beginQuery(ext.TIME_ELAPSED_EXT, query);
return query;
}
const ext = this.getQueryTimerExtensionWebGL1();
const query = ext.createQueryEXT();
ext.beginQueryEXT(ext.TIME_ELAPSED_EXT, query);
return query;
}
endQuery() {
if (env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION') === 2) {
const gl2 = this.gl;
const ext = this.getQueryTimerExtensionWebGL2();
gl2.endQuery(ext.TIME_ELAPSED_EXT);
return;
}
const ext = this.getQueryTimerExtensionWebGL1();
ext.endQueryEXT(ext.TIME_ELAPSED_EXT);
}
async waitForQueryAndGetTime(query) {
await repeatedTry(() => this.disposed || // while testing contexts are created / disposed
// in rapid succession, so without this check we
// may poll for the query timer indefinitely
this.isQueryAvailable(query, env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION')));
return this.getQueryTime(query, env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION'));
}
getQueryTime(query, queryTimerVersion) {
if (queryTimerVersion === 0) {
return null;
}
if (queryTimerVersion === 2) {
const gl2 = this.gl;
const timeElapsedNanos = gl2.getQueryParameter(query, gl2.QUERY_RESULT);
// Return milliseconds.
return timeElapsedNanos / 1000000;
}
else {
const ext = this.getQueryTimerExtensionWebGL1();
const timeElapsedNanos = ext.getQueryObjectEXT(query, ext.QUERY_RESULT_EXT);
// Return milliseconds.
return timeElapsedNanos / 1000000;
}
}
isQueryAvailable(query, queryTimerVersion) {
if (queryTimerVersion === 0) {
return true;
}
if (queryTimerVersion === 2) {
const gl2 = this.gl;
const ext = this.getQueryTimerExtensionWebGL2();
const available = gl2.getQueryParameter(query, gl2.QUERY_RESULT_AVAILABLE);
if (this.disjoint == null) {
this.disjoint = this.gl.getParameter(ext.GPU_DISJOINT_EXT);
}
return available && !this.disjoint;
}
else {
const ext = this.getQueryTimerExtensionWebGL1();
const available = ext.getQueryObjectEXT(query, ext.QUERY_RESULT_AVAILABLE_EXT);
if (this.disjoint == null) {
this.disjoint = this.gl.getParameter(ext.GPU_DISJOINT_EXT);
}
return available && !this.disjoint;
}
}
pollFence(fenceContext) {
return new Promise(resolve => {
this.addItemToPoll(() => fenceContext.isFencePassed(), () => resolve());
});
}
pollItems() {
// Find the last query that has finished.
const index = linearSearchLastTrue(this.itemsToPoll.map(x => x.isDoneFn));
for (let i = 0; i <= index; ++i) {
const { resolveFn } = this.itemsToPoll[i];
resolveFn();
}
this.itemsToPoll = this.itemsToPoll.slice(index + 1);
}
addItemToPoll(isDoneFn, resolveFn) {
this.itemsToPoll.push({ isDoneFn, resolveFn });
if (this.itemsToPoll.length > 1) {
// We already have a running loop that polls.
return;
}
// Start a new loop that polls.
let scheduleFn = undefined;
if ('setTimeoutCustom' in env().platform) {
scheduleFn = env().platform.setTimeoutCustom.bind(env().platform);
}
repeatedTry(() => {
this.pollItems();
// End the loop if no more items to poll.
return this.itemsToPoll.length === 0;
}, () => 0, null, scheduleFn);
}
bindTextureToFrameBuffer(texture) {
this.throwIfDisposed();
bindColorTextureToFramebuffer(this.gl, texture, this.framebuffer);
if (this.debug) {
validateFramebuffer(this.gl);
}
}
unbindTextureToFrameBuffer() {
if (this.outputTexture != null) {
bindColorTextureToFramebuffer(this.gl, this.outputTexture, this.framebuffer);
if (this.debug) {
validateFramebuffer(this.gl);
}
}
else {
unbindColorTextureFromFramebuffer(this.gl, this.framebuffer);
}
}
downloadMatrixDriver(texture, downloadAndDecode) {
this.bindTextureToFrameBuffer(texture);
const result = downloadAndDecode();
this.unbindTextureToFrameBuffer();
return result;
}
setOutputMatrixTextureDriver(outputMatrixTextureMaybePacked, width, height) {
this.throwIfDisposed();
const gl = this.gl;
bindColorTextureToFramebuffer(gl, outputMatrixTextureMaybePacked, this.framebuffer);
if (this.debug) {
validateFramebuffer(gl);
}
this.outputTexture = outputMatrixTextureMaybePacked;
callAndCheck(gl, () => gl.viewport(0, 0, width, height));
callAndCheck(gl, () => gl.scissor(0, 0, width, height));
}
setOutputMatrixWriteRegionDriver(x, y, width, height) {
this.throwIfDisposed();
callAndCheck(this.gl, () => this.gl.scissor(x, y, width, height));
}
throwIfDisposed() {
if (this.disposed) {
throw new Error('Attempted to use disposed GPGPUContext.');
}
}
throwIfNoProgram() {
if (this.program == null) {
throw new Error('No GPU program is currently set.');
}
}
}
/**
* Finds the index of the last true element using linear search.
* Note: We can't do binary search because Chrome expects us to explicitly
* test all fences before download:
* https://github.com/tensorflow/tfjs/issues/1145
*/
function linearSearchLastTrue(arr) {
let i = 0;
for (; i < arr.length; ++i) {
const isDone = arr[i]();
if (!isDone) {
break;
}
}
return i - 1;
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function assertNotComplex(tensor, opName) {
if (!Array.isArray(tensor)) {
tensor = [tensor];
}
tensor.forEach(t => {
if (t != null) {
assert$1(t.dtype !== 'complex64', () => `${opName} does not support complex64 tensors in the CPU backend.`);
}
});
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function simpleAbsImpl(vals) {
const resultValues = new Float32Array(vals.length);
for (let i = 0; i < vals.length; ++i) {
resultValues[i] = Math.abs(vals[i]);
}
return resultValues;
}
const abs$1 = (args) => {
const { x } = args.inputs;
const cpuBackend = args.backend;
assertNotComplex(x, 'abs');
let resultValues = new Float32Array(sizeFromShape(x.shape));
const values = cpuBackend.data.get(x.dataId).values;
resultValues = simpleAbsImpl(values);
return cpuBackend.makeOutput(resultValues, x.shape, x.dtype);
};
const absConfig$1 = {
kernelName: Abs,
backendName: 'cpu',
kernelFunc: abs$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Template that creates implementation for binary ops. Supports broadcast.
*/
function createSimpleBinaryKernelImpl(op) {
return (aShape, bShape, aVals, bVals, dtype) => {
const newShape = assertAndGetBroadcastShape(aShape, bShape);
const resultRank = newShape.length;
const resultStrides = computeStrides(newShape);
const resultSize = sizeFromShape(newShape);
const result = getTypedArrayFromDType(dtype, resultSize);
const aRank = aShape.length;
const bRank = bShape.length;
const aStrides = computeStrides(aShape);
const bStrides = computeStrides(bShape);
const aBroadcastDims = getBroadcastDims$1(aShape, newShape);
const bBroadcastDims = getBroadcastDims$1(bShape, newShape);
if (aBroadcastDims.length + bBroadcastDims.length === 0) {
for (let i = 0; i < result.length; ++i) {
result[i] = op(aVals[i % aVals.length], bVals[i % bVals.length]);
}
}
else {
for (let i = 0; i < result.length; ++i) {
const loc = indexToLoc(i, resultRank, resultStrides);
const aLoc = loc.slice(-aRank);
aBroadcastDims.forEach(d => aLoc[d] = 0);
const aIndex = locToIndex(aLoc, aRank, aStrides);
const bLoc = loc.slice(-bRank);
bBroadcastDims.forEach(d => bLoc[d] = 0);
const bIndex = locToIndex(bLoc, bRank, bStrides);
result[i] = op(aVals[aIndex], bVals[bIndex]);
}
}
return [result, newShape];
};
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function complex$1(args) {
const { inputs, backend } = args;
const { real, imag } = inputs;
const realVals = backend.data.get(real.dataId).values;
const imagVals = backend.data.get(imag.dataId).values;
const complexInfo = backend.makeTensorInfo(real.shape, 'complex64');
const complex = backend.data.get(complexInfo.dataId);
// The complex tensor owns the underlying real and imag tensorInfos, only the
// complex tensor tracks refCount, when complexData is disposed the
// underlying tensorData will be disposed.
complex.complexTensorInfos = {
real: backend.makeTensorInfo(real.shape, 'float32', realVals),
imag: backend.makeTensorInfo(imag.shape, 'float32', imagVals)
};
return complexInfo;
}
const complexConfig$1 = {
kernelName: Complex,
backendName: 'cpu',
kernelFunc: complex$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Generates a tensorInfo with all zeros value.
* @param backend cpu backend.
* @param shape Shape for the zeros tensor.
* @param dtype Optional. If set, the result has this dtype.
*/
function zeros(backend, shape, dtype = 'float32') {
if (dtype === 'complex64') {
const real = zeros(backend, shape, 'float32');
const imag = zeros(backend, shape, 'float32');
return complex$1({ inputs: { real, imag }, backend });
}
const values = makeZerosTypedArray(sizeFromShape(shape), dtype);
return backend.makeTensorInfo(shape, dtype, values);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function identity$1(args) {
const { inputs, backend } = args;
const { x } = inputs;
backend.incRef(x.dataId);
return { dataId: x.dataId, shape: x.shape, dtype: x.dtype };
}
const identityConfig$1 = {
kernelName: Identity$1,
backendName: 'cpu',
kernelFunc: identity$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function real$1(args) {
const { inputs, backend } = args;
const { input } = inputs;
const real = backend.data.get(input.dataId).complexTensorInfos.real;
const realVal = backend.data.get(real.dataId).values;
// When complex tensor is disposed, its underlying parts will be disposed too.
// Make new tensor out of the real value of the complex. This makes sure the
// value is still accessible even if complex tensor is disposed.
return backend.makeTensorInfo(real.shape, real.dtype, realVal);
}
const realConfig$1 = {
kernelName: Real,
backendName: 'cpu',
kernelFunc: real$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function castImpl(values, shape, inputType, dtype) {
if (dtype === 'int32') {
const resultValues = Int32Array.from(values);
return [shape, 'int32', resultValues];
}
if (dtype === 'bool') {
// This is essentially the result of notEqual(x, 0). We avoid using
// kernel notEqual to avoid circular dependency, i.e. binary_utils ->
// cast -> notEqual -> binary_utils.
const zero = toTypedArray([0], inputType);
const [resultData, resultShape] = createSimpleBinaryKernelImpl((a, b) => (a !== b) ? 1 : 0)(shape, [], values, zero, 'bool');
return [resultShape, 'bool', resultData];
}
throw new Error(`Error in Cast: failed to cast ${inputType} to ${dtype}`);
}
function cast$2(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { dtype } = attrs;
// Casting to complex64.
if (dtype === 'complex64') {
if (x.dtype === 'complex64') {
return identity$1({ inputs: { x }, backend });
}
const zerosTensorInfo = zeros(backend, x.shape, x.dtype);
const floatX = cast$2({ inputs: { x }, backend, attrs: { dtype: 'float32' } });
const result = complex$1({ inputs: { real: floatX, imag: zerosTensorInfo }, backend });
backend.disposeIntermediateTensorInfo(zerosTensorInfo);
backend.disposeIntermediateTensorInfo(floatX);
return result;
}
// Casting from complex64
if (x.dtype === 'complex64') {
const realPart = real$1({ inputs: { input: x }, backend });
const result = cast$2({ inputs: { x: realPart }, backend, attrs: { dtype } });
backend.disposeIntermediateTensorInfo(realPart);
return result;
}
if (!hasEncodingLoss(x.dtype, dtype)) {
// We don't change the underlying data, since we cast to higher
// precision.
const result = identity$1({ inputs: { x }, backend });
return { dataId: result.dataId, shape: result.shape, dtype };
}
const values = backend.data.get(x.dataId).values;
const [resultShape, resultType, resultData] = castImpl(values, x.shape, x.dtype, dtype);
return backend.makeTensorInfo(resultShape, resultType, resultData);
}
const castConfig$1 = {
kernelName: Cast,
backendName: 'cpu',
kernelFunc: cast$2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Template that creates a `KernelFunc` for binary ops.
* @param name Kernel name.
* @param binaryKernelImpl A `SimpleBinaryKernelImpl` for the kernel.
* @param binaryKernelComplexImpl Optional. If exists, represents a
* `ComplexBinaryKernelImpl` for the kernel, will be used when input dtype
* is `complex64`.
* @param dtype Optional. If set, the result has this dtype. Otherwise, the
* result has the same dtype as the first input. This is mainly used in
* comparison kernels, such as Equal, Less, Greater, etc.
*/
function binaryKernelFunc$1(name, simpleImpl, complexImpl, dtype) {
if (complexImpl == null) {
return ({ inputs, backend }) => {
const { a, b } = inputs;
const cpuBackend = backend;
assertNotComplex([a, b], name);
const aVals = cpuBackend.data.get(a.dataId).values;
const bVals = cpuBackend.data.get(b.dataId).values;
const decodedAVals = a.dtype === 'string' ?
// tslint:disable-next-line: no-any
fromUint8ToStringArray(aVals) :
aVals;
const decodedBVals = a.dtype === 'string' ?
// tslint:disable-next-line: no-any
fromUint8ToStringArray(bVals) :
bVals;
const $dtype = dtype || a.dtype;
const [resultData, resultShape] = simpleImpl(a.shape, b.shape, decodedAVals, decodedBVals, $dtype);
return cpuBackend.makeTensorInfo(resultShape, $dtype, resultData);
};
}
return ({ inputs, backend }) => {
const { a, b } = inputs;
const cpuBackend = backend;
if (a.dtype === 'complex64' || b.dtype === 'complex64') {
const $aComplex = cast$2({ inputs: { x: a }, backend: cpuBackend, attrs: { dtype: 'complex64' } });
const $aComplexVals = cpuBackend.data.get($aComplex.dataId);
const aReal = $aComplexVals.complexTensorInfos.real;
const aImag = $aComplexVals.complexTensorInfos.imag;
const aRealVals = cpuBackend.data.get(aReal.dataId).values;
const aImagVals = cpuBackend.data.get(aImag.dataId).values;
const $bComplex = cast$2({ inputs: { x: b }, backend: cpuBackend, attrs: { dtype: 'complex64' } });
const $bComplexVals = cpuBackend.data.get($bComplex.dataId);
const bReal = $bComplexVals.complexTensorInfos.real;
const bImag = $bComplexVals.complexTensorInfos.imag;
const bRealVals = cpuBackend.data.get(bReal.dataId).values;
const bImagVals = cpuBackend.data.get(bImag.dataId).values;
const [resultRealData, resultImagData, resultShape] = complexImpl(a.shape, b.shape, aRealVals, aImagVals, bRealVals, bImagVals);
const resultReal = cpuBackend.makeTensorInfo(resultShape, 'float32', resultRealData);
const resultImag = cpuBackend.makeTensorInfo(resultShape, 'float32', resultImagData);
const result = complex$1({ inputs: { real: resultReal, imag: resultImag }, backend: cpuBackend });
cpuBackend.disposeIntermediateTensorInfo($aComplex);
cpuBackend.disposeIntermediateTensorInfo($bComplex);
cpuBackend.disposeIntermediateTensorInfo(resultReal);
cpuBackend.disposeIntermediateTensorInfo(resultImag);
return result;
}
else {
const aVals = cpuBackend.data.get(a.dataId).values;
const bVals = cpuBackend.data.get(b.dataId).values;
const $dtype = dtype || a.dtype;
const [resultData, resultShape] = simpleImpl(a.shape, b.shape, aVals, bVals, $dtype);
return cpuBackend.makeTensorInfo(resultShape, $dtype, resultData);
}
};
}
/**
* Template that creates the complex type implementation for binary ops.
* Supports broadcast.
*/
function createComplexBinaryKernelImpl(op) {
return (aShape, bShape, aRealVals, aImagVals, bRealVals, bImagVals) => {
const resultShape = assertAndGetBroadcastShape(aShape, bShape);
const resultSize = sizeFromShape(resultShape);
const resultRank = resultShape.length;
const resultStrides = computeStrides(resultShape);
const resultRealVals = getTypedArrayFromDType('float32', resultSize);
const resultImagVals = getTypedArrayFromDType('float32', resultSize);
const aBroadcastDims = getBroadcastDims$1(aShape, resultShape);
const bBroadcastDims = getBroadcastDims$1(bShape, resultShape);
const aVals = mergeRealAndImagArrays(aRealVals, aImagVals);
const bVals = mergeRealAndImagArrays(bRealVals, bImagVals);
const aRank = aShape.length;
const aStrides = computeStrides(aShape);
const bRank = bShape.length;
const bStrides = computeStrides(bShape);
if (aBroadcastDims.length + bBroadcastDims.length === 0) {
for (let i = 0; i < resultRealVals.length; i++) {
const aIdx = i % aVals.length;
const bIdx = i % bVals.length;
const result = op(aVals[aIdx * 2], aVals[aIdx * 2 + 1], bVals[bIdx * 2], bVals[bIdx * 2 + 1]);
resultRealVals[i] = result.real;
resultImagVals[i] = result.imag;
}
}
else {
for (let i = 0; i < resultRealVals.length; i++) {
const loc = indexToLoc(i, resultRank, resultStrides);
const aLoc = loc.slice(-aRank);
aBroadcastDims.forEach(d => aLoc[d] = 0);
const aIndex = locToIndex(aLoc, aRank, aStrides);
const bLoc = loc.slice(-bRank);
bBroadcastDims.forEach(d => bLoc[d] = 0);
const bIndex = locToIndex(bLoc, bRank, bStrides);
const opResult = op(aVals[aIndex * 2], aVals[aIndex * 2 + 1], bVals[bIndex * 2], bVals[bIndex * 2 + 1]);
resultRealVals[i] = opResult.real;
resultImagVals[i] = opResult.imag;
}
}
return [resultRealVals, resultImagVals, resultShape];
};
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const addImpl = createSimpleBinaryKernelImpl(((a, b) => a + b));
const addComplexImpl = createComplexBinaryKernelImpl(((aReal, aImag, bReal, bImag) => {
return { real: aReal + bReal, imag: aImag + bImag };
}));
const add = binaryKernelFunc$1(Add, addImpl, addComplexImpl);
const addConfig$1 = {
kernelName: Add,
backendName: 'cpu',
kernelFunc: add
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function bincountImpl(xVals, weightsVals, weightsDtype, weightsShape, size) {
const weightsSize = sizeFromShape(weightsShape);
const outVals = makeZerosTypedArray(size, weightsDtype);
for (let i = 0; i < xVals.length; i++) {
const value = xVals[i];
if (value < 0) {
throw new Error('Input x must be non-negative!');
}
if (value >= size) {
continue;
}
if (weightsSize > 0) {
outVals[value] += weightsVals[i];
}
else {
outVals[value] += 1;
}
}
return outVals;
}
function bincountReduceImpl(xBuf, weightsBuf, size, binaryOutput = false) {
const numRows = xBuf.shape[0];
const numCols = xBuf.shape[1];
const outBuf = buffer([numRows, size], weightsBuf.dtype);
for (let i = 0; i < numRows; i++) {
for (let j = 0; j < numCols; j++) {
const value = xBuf.get(i, j);
if (value < 0) {
throw new Error('Input x must be non-negative!');
}
if (value >= size) {
continue;
}
if (binaryOutput) {
outBuf.set(1, i, value);
}
else {
if (weightsBuf.size > 0) {
outBuf.set(outBuf.get(i, value) + weightsBuf.get(i, j), i, value);
}
else {
outBuf.set(outBuf.get(i, value) + 1, i, value);
}
}
}
}
return outBuf;
}
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const bitwiseAndImpl = createSimpleBinaryKernelImpl(((a, b) => a & b));
const bitwiseAnd$1 = binaryKernelFunc$1(BitwiseAnd, bitwiseAndImpl);
const bitwiseAndConfig$1 = {
kernelName: BitwiseAnd,
backendName: 'cpu',
kernelFunc: bitwiseAnd$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Template that creates implementation for unary op.
*/
function createSimpleUnaryImpl(op) {
return (values, dtype, attrs) => {
const newValues = getArrayFromDType(dtype, values.length);
for (let i = 0; i < values.length; ++i) {
newValues[i] = op(values[i], attrs);
}
return newValues;
};
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Template that creates a `KernelFunc` for unary ops.
* @param name Kernel name.
* @param op A `SimpleUnaryOperation` for the kernel.
* @param dtype Optional. If set, the result has this dtype. Otherwise, the
* result has the same dtype as the input. This is mainly used in certain
* kernels that return bool type, such as isFinite, isInf, etc.
*/
function unaryKernelFunc$1(name, op, dtype) {
const impl = createSimpleUnaryImpl(op);
return unaryKernelFuncFromImpl(name, impl, dtype);
}
/**
* Template that creates a `KernelFunc` for unary ops from the given
* `SimpleUnaryImpl`..
* @param name Kernel name.
* @param unaryImpl A `SimpleUnaryImpl` that implements the op.
* @param dtype Optional. If set, the result has this dtype. Otherwise, the
* result has the same dtype as the input. This is mainly used in certain
* kernels that return bool type, such as isFinite, isInf, etc.
*/
function unaryKernelFuncFromImpl(name, unaryImpl, dtype) {
return ({ inputs, attrs, backend }) => {
const { x } = inputs;
assertNotComplex(x, name);
const cpuBackend = backend;
const values = cpuBackend.data.get(x.dataId).values;
let decoded;
if (x.dtype === 'string') {
if (!Array.isArray(values)) {
throw new Error('String tensor\'s value was not an instance of Array');
}
decoded = fromUint8ToStringArray(values);
}
else {
decoded = values;
}
const $dtype = dtype || x.dtype;
const newValues = unaryImpl(decoded, $dtype, attrs);
return cpuBackend.makeTensorInfo(x.shape, $dtype, newValues);
};
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ceilImpl = createSimpleUnaryImpl((xi) => Math.ceil(xi));
const ceil$1 = unaryKernelFuncFromImpl(Ceil, ceilImpl);
const ceilConfig$1 = {
kernelName: Ceil,
backendName: 'cpu',
kernelFunc: ceil$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function concatImpl$1(inputs, outShape, dtype, simplyConcat) {
const outVals = getArrayFromDType(dtype, sizeFromShape(outShape));
if (simplyConcat && dtype !== 'string') {
// Use built-in TypedArray.set() method for speed.
let offset = 0;
inputs.forEach(input => {
const size = sizeFromShape(input.shape);
outVals.set(input.vals, offset);
offset += size;
});
}
else {
let colOffset = 0;
inputs.forEach(input => {
const decodedData = dtype === 'string' ?
fromUint8ToStringArray(input.vals) :
input.vals;
let tIdx = 0;
for (let row = 0; row < input.shape[0]; ++row) {
const resIdx = row * outShape[1] + colOffset;
for (let col = 0; col < input.shape[1]; ++col) {
outVals[resIdx + col] = decodedData[tIdx++];
}
}
colOffset += input.shape[1];
});
}
return outVals;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const equalImpl = createSimpleBinaryKernelImpl((a, b) => (a === b) ? 1 : 0);
const equal$1 = binaryKernelFunc$1(Equal, equalImpl, null /* complexImpl */, 'bool');
const equalConfig$1 = {
kernelName: Equal,
backendName: 'cpu',
kernelFunc: equal$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const expImpl = createSimpleUnaryImpl((xi) => Math.exp(xi));
const exp$1 = unaryKernelFuncFromImpl(Exp, expImpl, 'float32');
const expConfig$1 = {
kernelName: Exp,
backendName: 'cpu',
kernelFunc: exp$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const expm1Impl = createSimpleUnaryImpl((xi) => Math.expm1(xi));
const expm1$1 = unaryKernelFuncFromImpl(Expm1, expm1Impl);
const expm1Config$1 = {
kernelName: Expm1,
backendName: 'cpu',
kernelFunc: expm1$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const floorImpl = createSimpleUnaryImpl((xi) => Math.floor(xi));
const floor$1 = unaryKernelFuncFromImpl(Floor, floorImpl);
const floorConfig$1 = {
kernelName: Floor,
backendName: 'cpu',
kernelFunc: floor$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const floorDivImpl = createSimpleBinaryKernelImpl((a, b) => Math.floor(a / b));
const floorDiv$1 = binaryKernelFunc$1(FloorDiv, floorDivImpl, null /* complexImpl */, 'int32');
const floorDivConfig$1 = {
kernelName: FloorDiv,
backendName: 'cpu',
kernelFunc: floorDiv$1
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function gatherNdImpl(indicesData, paramsBuf, dtype, numSlices, sliceRank, sliceSize, strides, paramsShape, paramsSize) {
const outBuf = buffer([numSlices, sliceSize], dtype);
for (let i = 0; i < numSlices; i++) {
const index = [];
let flattenIndex = 0;
for (let j = 0; j < sliceRank; j++) {
const dim = indicesData[i * sliceRank + j];
flattenIndex += dim * strides[j];
index.push(dim);
}
if (flattenIndex < 0 || flattenIndex >= paramsSize / sliceSize) {
throw new Error(`Invalid indices: ${index} does not index into ${paramsShape}`);
}
for (let k = 0; k < sliceSize; k++) {
outBuf.values[i * sliceSize + k] =
paramsBuf.get(...paramsBuf.indexToLoc(flattenIndex * sliceSize + k));
}
}
return outBuf;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function gatherV2Impl(xBuf, indicesBuf, flattenOutputShape) {
const outBuf = buffer(flattenOutputShape, xBuf.dtype);
for (let i = 0; i < outBuf.size; ++i) {
const newLoc = outBuf.indexToLoc(i);
const originalLoc = newLoc.slice();
const batchIdx = originalLoc[0];
const indicesIdx = originalLoc[2];
const indicesIndex = indicesBuf.locToIndex([batchIdx, indicesIdx]);
originalLoc[2] = indicesBuf.values[indicesIndex];
const originalIndex = xBuf.locToIndex(originalLoc);
if (0 <= originalIndex && originalIndex < xBuf.values.length) {
outBuf.values[i] = xBuf.values[originalIndex];
} // Else, index is out of bounds, so leave the default zero val in outBuf.
}
return outBuf;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const greaterImpl = createSimpleBinaryKernelImpl((a, b) => (a > b) ? 1 : 0);
const greater$1 = binaryKernelFunc$1(Greater, greaterImpl, null /* complexImpl */, 'bool');
const greaterConfig$1 = {
kernelName: Greater,
backendName: 'cpu',
kernelFunc: greater$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const greaterEqualImpl = createSimpleBinaryKernelImpl((a, b) => (a >= b) ? 1 : 0);
const greaterEqual$1 = binaryKernelFunc$1(GreaterEqual, greaterEqualImpl, null /* complexImpl */, 'bool');
const greaterEqualConfig$1 = {
kernelName: GreaterEqual,
backendName: 'cpu',
kernelFunc: greaterEqual$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const lessImpl = createSimpleBinaryKernelImpl((a, b) => (a < b) ? 1 : 0);
const less$1 = binaryKernelFunc$1(Less, lessImpl, null /* complexImpl */, 'bool');
const lessConfig$1 = {
kernelName: Less,
backendName: 'cpu',
kernelFunc: less$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const lessEqualImpl = createSimpleBinaryKernelImpl((a, b) => (a <= b) ? 1 : 0);
const lessEqual$1 = binaryKernelFunc$1(LessEqual, lessEqualImpl, null /* complexImpl */, 'bool');
const lessEqualConfig$1 = {
kernelName: LessEqual,
backendName: 'cpu',
kernelFunc: lessEqual$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function linSpaceImpl(start, stop, num) {
const step = (stop - start) / (num - 1);
const values = makeZerosTypedArray(num, 'float32');
values[0] = start;
for (let i = 1; i < values.length; i++) {
values[i] = values[i - 1] + step;
}
return values;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const logImpl = createSimpleUnaryImpl((xi) => Math.log(xi));
const log$1 = unaryKernelFuncFromImpl(Log, logImpl);
const logConfig$1 = {
kernelName: Log,
backendName: 'cpu',
kernelFunc: log$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxImpl$1(aVals, reduceSize, outShape, dtype) {
const vals = getTypedArrayFromDType(dtype, sizeFromShape(outShape));
for (let i = 0; i < vals.length; ++i) {
const offset = i * reduceSize;
let max = aVals[offset];
for (let j = 0; j < reduceSize; ++j) {
const value = aVals[offset + j];
if (Number.isNaN(value) ||
value > max) { // comparison with NaN always return false
max = value;
}
}
vals[i] = max;
}
return vals;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const maximumImpl = createSimpleBinaryKernelImpl(((aValue, bValue) => Math.max(aValue, bValue)));
const maximum$1 = binaryKernelFunc$1(Maximum, maximumImpl);
const maximumConfig$1 = {
kernelName: Maximum,
backendName: 'cpu',
kernelFunc: maximum$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const minimumImpl = createSimpleBinaryKernelImpl(((aValue, bValue) => Math.min(aValue, bValue)));
const minimum$1 = binaryKernelFunc$1(Minimum, minimumImpl);
const minimumConfig$1 = {
kernelName: Minimum,
backendName: 'cpu',
kernelFunc: minimum$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const multiplyImpl = createSimpleBinaryKernelImpl(((aValue, bValue) => aValue * bValue));
const multiplyComplexImpl = createComplexBinaryKernelImpl(((aReal, aImag, bReal, bImag) => {
return {
real: aReal * bReal - aImag * bImag,
imag: aReal * bImag + aImag * bReal
};
}));
const multiply$1 = binaryKernelFunc$1(Multiply, multiplyImpl, multiplyComplexImpl);
const multiplyConfig$1 = {
kernelName: Multiply,
backendName: 'cpu',
kernelFunc: multiply$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function negImpl(xVals, xShape, xDtype) {
const minusOne = createScalarValue(-1, xDtype);
return multiplyImpl([], xShape, minusOne, xVals, xDtype);
}
function neg$1(args) {
const { inputs, backend } = args;
const { x } = inputs;
assertNotComplex(x, 'neg');
const xVals = backend.data.get(x.dataId).values;
const [res, newShape] = negImpl(xVals, x.shape, x.dtype);
return backend.makeTensorInfo(newShape, x.dtype, res);
}
const negConfig$1 = {
kernelName: Neg,
backendName: 'cpu',
kernelFunc: neg$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const notEqualImpl = createSimpleBinaryKernelImpl(((a, b) => (a !== b) ? 1 : 0));
const notEqual$1 = binaryKernelFunc$1(NotEqual, notEqualImpl, null /* complexOp */, 'bool');
const notEqualConfig$1 = {
kernelName: NotEqual,
backendName: 'cpu',
kernelFunc: notEqual$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function transposeImpl$1(xVals, xShape, dtype, perm, newShape) {
const xRank = xShape.length;
const xSize = sizeFromShape(xShape);
const xStrides = computeStrides(xShape);
const newStrides = computeStrides(newShape);
const result = getTypedArrayFromDType(dtype, sizeFromShape(newShape));
for (let i = 0; i < xSize; ++i) {
const loc = indexToLoc(i, xRank, xStrides);
// Permute location.
const newLoc = new Array(loc.length);
for (let i = 0; i < newLoc.length; i++) {
newLoc[i] = loc[perm[i]];
}
const newIndex = locToIndex(newLoc, xRank, newStrides);
result[newIndex] = xVals[i];
}
return result;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function transpose$1(args) {
const { inputs, attrs, backend } = args;
const { x } = inputs;
const { perm } = attrs;
assertNotComplex(x, 'transpose');
const xRank = x.shape.length;
const newShape = new Array(xRank);
for (let i = 0; i < newShape.length; i++) {
newShape[i] = x.shape[perm[i]];
}
const values = backend.data.get(x.dataId).values;
const result = transposeImpl$1(values, x.shape, x.dtype, perm, newShape);
const dataId = backend.write(result, newShape, x.dtype);
return { dataId, shape: newShape, dtype: x.dtype };
}
const transposeConfig$1 = {
kernelName: Transpose,
backendName: 'cpu',
kernelFunc: transpose$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function prodImpl(xShape, xDtype, xVals, reductionAxes) {
const [outShape, reduceShape] = computeOutAndReduceShapes(xShape, reductionAxes);
const outDtype = upcastType(xDtype, 'int32');
const outVals = makeZerosTypedArray(sizeFromShape(outShape), outDtype);
const reduceSize = sizeFromShape(reduceShape);
for (let i = 0; i < outVals.length; ++i) {
const offset = i * reduceSize;
let prod = 1;
for (let j = 0; j < reduceSize; ++j) {
prod *= xVals[offset + j];
}
outVals[i] = prod;
}
return { outVals, outShape, outDtype };
}
function prod$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
assertNotComplex(x, 'prod');
const xRank = x.shape.length;
const axes = parseAxisParam(axis, x.shape);
const permutation = getAxesPermutation(axes, xRank);
let reductionAxes = axes;
let permutedX = x;
const intermediateTensorInfos = [];
if (permutation != null) {
permutedX = transpose$1({ inputs: { x }, backend, attrs: { perm: permutation } });
intermediateTensorInfos.push(permutedX);
reductionAxes = getInnerMostAxes(reductionAxes.length, xRank);
}
const xVals = backend.data.get(permutedX.dataId).values;
const { outVals, outShape, outDtype } = prodImpl(permutedX.shape, permutedX.dtype, xVals, reductionAxes);
let resultShape = outShape;
if (keepDims) {
resultShape = expandShapeToKeepDim(outShape, axes);
}
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return backend.makeTensorInfo(resultShape, outDtype, outVals);
}
const prodConfig$1 = {
kernelName: Prod,
backendName: 'cpu',
kernelFunc: prod$1
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function validateIndices(indices, indicesShape, numParams) {
indices.forEach((index, i) => {
if (index < 0 || index >= numParams) {
const locString = indexToLoc(i, indicesShape.length, computeStrides(indicesShape))
.join(',');
throw new Error(`indices[${locString}] = ${index} is not in [0, ${numParams})`);
}
});
}
function validateSplits(paramsNestedSplits, numParamsDenseValues) {
// Validate
for (let dim = 0; dim < paramsNestedSplits.length; ++dim) {
const splits = paramsNestedSplits[dim];
const lastSplit = (dim === paramsNestedSplits.length - 1) ?
numParamsDenseValues :
paramsNestedSplits[dim + 1].length;
if (splits.length === 0) {
throw new Error('Ragged splits may not be empty');
}
if (splits[0] < 0) {
throw new Error('Ragged splits must be non-negative');
}
if (splits[splits.length - 1] > lastSplit) {
throw new Error('Ragged splits must not point past values');
}
for (let i = 1; i < splits.length; ++i) {
if (splits[i - 1] > splits[i]) {
throw new Error('Ragged splits must be sorted in ascending order');
}
}
}
}
// Construct the `splits` output tensors, encoded using a nested vector.
// Also find the slices of values that need to be copied, and store them
// in `valueSlices`. The total number of values that will be copied (which
// we need for allocating the output values tensor) is stored in `numValues`.
function makeSplits(indices, indicesShape, paramsNestedSplits, numParamsDenseValues) {
const valueSlices = [];
let numValues = 0;
const numSplits = indicesShape.length - 1 + paramsNestedSplits.length;
const outSplits = new Array(numSplits).fill(null).map(() => [0]);
validateSplits(paramsNestedSplits, numParamsDenseValues);
// Add `splits` that come from all but the last dimension of the dense
// Tensor `indices`. In particular, for each dimension D, we add a
// splits tensor whose values are:
// range(reduceProd(splits.shape[:D]) + 1) * splits.shape[D+1]
// E.g., if indices.shape=[2, 3, 4] then we will add splits tensors:
// [0, 3, 6] # length=2+1, stride=3
// [0, 4, 8, 12, 16, 20, 24] # length=2*3+1, stride=4
let nrows = 1;
for (let dim = 0; dim < indicesShape.length - 1; ++dim) {
nrows *= indicesShape[dim];
const rowLength = indicesShape[dim + 1];
for (let i = 1; i < nrows + 1; ++i) {
outSplits[dim].push(i * rowLength);
}
}
// Add `splits` that come from `paramsNestedSplits`. Starting with the
// outermost ragged dimension (i.e., the first `splits` tensor), we work
// our way in, finding the range of values that should be copied. As we
// go, we update the output `splits` for each dimension with the appropriate
// values. In particular, the *lengths* of the slices from `param_splits`
// should be copied to generate corresponding slice lengths in the output
// splits. E.g., if we are copying a ragged row with length 4, then we
// should add a new split point to outSplits that is 4 greater than the
// previous split point in outSplits.
for (let i = 0; i < indices.length; ++i) {
let start = indices[i];
let limit = indices[i] + 1;
// Copy splits.
for (let dim = 0; dim < paramsNestedSplits.length; ++dim) {
const splits = paramsNestedSplits[dim];
const outDim = dim + indicesShape.length - 1;
if (outDim >= 0) {
const outSplitsOutDim = outSplits[outDim];
const delta = outSplitsOutDim[outSplitsOutDim.length - 1] - splits[start];
for (let j = start; j < limit; ++j) {
outSplits[outDim].push(splits[j + 1] + delta);
}
}
start = splits[start];
limit = splits[limit];
}
if (limit !== start) {
valueSlices.push([start, limit]);
numValues += limit - start;
}
}
return { outSplits, valueSlices, numValues };
}
function getSplits(outSplits) {
const splitsOut = [];
for (let i = 0; i < outSplits.length; ++i) {
const numSplits = outSplits[i].length;
const splits = getArrayFromDType('int32', numSplits);
splitsOut.push(splits);
outSplits[i].forEach((value, j) => splits[j] = value);
}
return splitsOut;
}
function computeFlatOuterDims(orig, numOutDims) {
const outDims = orig.slice(0, numOutDims);
while (outDims.length < numOutDims) {
outDims.push(1);
}
for (let inDim = numOutDims; inDim < orig.length; inDim++) {
outDims[numOutDims - 1] *= orig[inDim];
}
return outDims;
}
// For each slice in `(start, limit)` in `valueSlices`, append
// `paramsDenseValues[start,...,limit] to `values`. `valueSize` indicates
// the number of scalars contained in each value paramsDenseValues[i].
function writeValueSlices(paramsDenseValues, paramsDenseValuesShape, valueSlices, valueSize, values, valuesShape) {
const denseM = computeFlatOuterDims(paramsDenseValuesShape, 2)[1];
const valuesM = computeFlatOuterDims(valuesShape, 2)[1];
let outPos = 0;
for (const slice of valueSlices) {
for (let i = slice[0]; i < slice[1]; ++i) {
for (let j = 0; j < valueSize; ++j) {
values[outPos * valuesM + j] = paramsDenseValues[i * denseM + j];
}
++outPos;
}
}
}
function getValues(paramsDenseValues, paramsDenseValuesShape, paramsDenseValuesDType, valueSlices, numValues) {
const valuesShape = paramsDenseValuesShape.slice();
valuesShape[0] = numValues;
const valuesOut = getArrayFromDType(paramsDenseValuesDType, sizeFromShape(valuesShape));
const numElements = paramsDenseValues.length;
const valueSize = numElements === 0 ? 0 : (numElements / paramsDenseValuesShape[0]);
writeValueSlices(paramsDenseValues, paramsDenseValuesShape, valueSlices, valueSize, valuesOut, valuesShape);
return [valuesOut, valuesShape];
}
function raggedGatherImpl(paramsNestedSplits, paramsNestedSplitsShapes, paramsDenseValues, paramsDenseValuesShape, paramsDenseValuesDType, indices, indicesShape, outputRaggedRank) {
if (paramsNestedSplits.length === 0) {
throw new Error('paramsNestedSplits must be non empty');
}
if (paramsNestedSplitsShapes[0].length === 0) {
throw new Error('Split tensors must not be scalars');
}
const numParams = paramsNestedSplitsShapes[0][0] - 1;
validateIndices(indices, indicesShape, numParams);
if (paramsDenseValuesShape.length === 0) {
throw new Error('params.rank must be nonzero');
}
const numParamsDenseValues = paramsDenseValuesShape[0];
// Calculate the `splits`, and store the value slices that we need to
// copy in `valueSlices`.
const { outSplits, valueSlices, numValues } = makeSplits(indices, indicesShape, paramsNestedSplits, numParamsDenseValues);
// Write the output tensors.
const outputNestedSplits = getSplits(outSplits);
const outputDenseValues = getValues(paramsDenseValues, paramsDenseValuesShape, paramsDenseValuesDType, valueSlices, numValues);
return [outputNestedSplits, outputDenseValues[0], outputDenseValues[1]];
}
/**
* @license
* Copyright 2022 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const INT32_MAX = 2147483647;
function raggedRangeImpl(starts, startsShape, startsDType, limits, limitsShape, deltas, deltasShape) {
// Check input tensor shapes.
if (startsShape.length > 1) {
throw new Error('starts must be a scalar or vector');
}
if (limitsShape.length > 1) {
throw new Error('limits must be a scalar or vector');
}
if (deltasShape.length > 1) {
throw new Error('deltas must be a scalar or vector');
}
// Determine which tensors we need to broadcast.
const broadcastStarts = startsShape.length === 0;
const broadcastLimits = limitsShape.length === 0;
const broadcastDeltas = deltasShape.length === 0;
// nRows (number of output rows) is the size of the non-broadcast inputs,
// or 1 if all inputs are scalars.
const inSizes = [];
if (!broadcastStarts) {
inSizes.push(startsShape[0]);
}
if (!broadcastLimits) {
inSizes.push(limitsShape[0]);
}
if (!broadcastDeltas) {
inSizes.push(deltasShape[0]);
}
for (let i = 1; i < inSizes.length; ++i) {
if (inSizes[i] !== inSizes[i - 1]) {
throw new Error('starts, limits, and deltas must have the same shape');
}
}
const nRows = inSizes.length === 0 ? 1 : inSizes[0];
// Construct the rtNestedSplits tensor.
const rtNestedSplits = getArrayFromDType('int32', nRows + 1);
rtNestedSplits[0] = 0;
for (let row = 0; row < nRows; ++row) {
const start = broadcastStarts ? starts[0] : starts[row];
const limit = broadcastLimits ? limits[0] : limits[row];
const delta = broadcastDeltas ? deltas[0] : deltas[row];
if (delta === 0) {
throw new Error('Requires delta != 0');
}
let size; // The number of elements in the specified range.
if (((delta > 0) && (limit < start)) || ((delta < 0) && (limit > start))) {
size = 0;
}
else {
size = Math.ceil(Math.abs((limit - start) / delta));
if (size > INT32_MAX) {
throw new Error(`Requires ((limit - start) / delta) <= ${INT32_MAX}`);
}
}
rtNestedSplits[row + 1] = rtNestedSplits[row] + size;
}
const nVals = rtNestedSplits[nRows];
// Construct the rtDenseValues tensor.
const rtDenseValues = getArrayFromDType(startsDType, nVals);
let valueIndex = 0;
for (let row = 0; row < nRows; ++row) {
const rowSize = rtNestedSplits[row + 1] - rtNestedSplits[row];
let value = broadcastStarts ? starts[0] : starts[row];
const delta = broadcastDeltas ? deltas[0] : deltas[row];
for (let i = 0; i < rowSize; ++i) {
rtDenseValues[valueIndex++] = value;
value += delta;
}
}
return [rtNestedSplits, rtDenseValues];
}
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
var RowPartitionType = RowPartitionType$1;
// Based on
// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc
class RaggedTensorToTensorOp {
constructor(shape, shapeShape, values, valuesShape, valuesDType, defaultValue, defaultValueShape, rowPartitionValues, rowPartitionValuesShapes, rowPartitionTypeStrings) {
this.shape = shape;
this.shapeShape = shapeShape;
this.values = values;
this.valuesShape = valuesShape;
this.valuesDType = valuesDType;
this.defaultValue = defaultValue;
this.defaultValueShape = defaultValueShape;
this.rowPartitionValues = rowPartitionValues;
this.rowPartitionValuesShapes = rowPartitionValuesShapes;
this.rowPartitionTypes =
getRowPartitionTypesHelper(rowPartitionTypeStrings);
this.raggedRank = getRaggedRank(this.rowPartitionTypes);
}
getRowPartitionTypeByDimension(dimension) {
if (this.rowPartitionTypes[0] === RowPartitionType.FIRST_DIM_SIZE) {
return this.rowPartitionTypes[dimension + 1];
}
else {
return this.rowPartitionTypes[dimension];
}
}
// Returns the relationship between dimension and dimension + 1.
getRowPartitionTensor(dimension) {
if (this.rowPartitionTypes[0] === RowPartitionType.FIRST_DIM_SIZE) {
return this.rowPartitionValues[dimension + 1];
}
else {
return this.rowPartitionValues[dimension];
}
}
getMaxWidth(dimension) {
const rowPartitionTensor = this.getRowPartitionTensor(dimension - 1);
switch (this.getRowPartitionTypeByDimension(dimension - 1)) {
case RowPartitionType.VALUE_ROWIDS:
return RaggedTensorToTensorOp.getMaxWidthValueRowID(rowPartitionTensor);
case RowPartitionType.ROW_SPLITS:
return RaggedTensorToTensorOp.getMaxWidthRowSplit(rowPartitionTensor);
default:
throw new Error(`Cannot handle partition type ${RowPartitionType[this.getRowPartitionTypeByDimension(dimension - 1)]}`);
}
}
static getMaxWidthRowSplit(rowSplit) {
const tensorLength = rowSplit.length;
if (tensorLength === 0 || tensorLength === 1) {
return 0;
}
let maxWidth = 0;
for (let i = 0; i < tensorLength - 1; ++i) {
const currentWidth = rowSplit[i + 1] - rowSplit[i];
if (currentWidth > maxWidth) {
maxWidth = currentWidth;
}
}
return maxWidth;
}
static getMaxWidthValueRowID(valueRowIds) {
const indexLength = valueRowIds.length;
if (indexLength === 0) {
return 0;
}
let firstEqualIndex = 0;
let firstEqualIndexValue = valueRowIds[0];
let maxWidth = 0;
for (let i = 1; i < indexLength; ++i) {
const value = valueRowIds[i];
if (value !== firstEqualIndexValue) {
firstEqualIndexValue = value;
maxWidth = Math.max(i - firstEqualIndex, maxWidth);
firstEqualIndex = i;
}
}
return Math.max(indexLength - firstEqualIndex, maxWidth);
}
tensorShapeFromTensor(t, tShape, isPartial = true) {
if (tShape.length === 0) {
if (t[0] === -1) {
return [];
}
throw new Error(`The only valid scalar shape tensor is the fully unknown shape specified as -1.`);
}
// MakePartialShape/MakeShapeHelper.
return makeShape(t, isPartial);
}
calculateOutputSize(firstDim) {
const valueShape = this.valuesShape;
const defaultValueShape = this.defaultValueShape;
validateDefaultValueShape(defaultValueShape, valueShape);
const shape = this.tensorShapeFromTensor(this.shape, this.shapeShape);
const outputShape = combineRaggedTensorToTensorShapes(this.raggedRank, shape, valueShape);
const result = outputShape;
if (result[0] < 0) {
result[0] = firstDim;
}
for (let i = 1; i <= this.raggedRank; ++i) {
if (result[i] < 0) {
result[i] = this.getMaxWidth(i);
}
}
return result;
}
/**
* The outputIndex represents the index in the output tensor
* where the first element of a particular dimension would be written.
* If it is -1, it indicates that the index is out of scope.
* Example, given firstDimension = 10, firstDimensionOutput = 6,
* and outputIndexMultiplier = 100:
* result = [0 100 200 300 400 500 -1 -1 -1 -1]
* If firstDimensionOutput = 11 instead, then:
* result = [0 100 200 300 400 500 600 700 800 900]
*/
calculateFirstParentOutputIndex(firstDimension, outputIndexMultiplier, firstDimensionOutput) {
const minDimension = Math.min(firstDimension, firstDimensionOutput);
const result = [];
let currentOutputIndex = 0;
for (let i = 0; i < minDimension; ++i, currentOutputIndex += outputIndexMultiplier) {
result.push(currentOutputIndex);
}
for (let i = minDimension; i < firstDimension; ++i) {
result.push(-1);
}
assert$1(result.length === firstDimension, () => 'Final length of result must be equal to firstDimension.');
return result;
}
calculateOutputIndexRowSplit(rowSplit, parentOutputIndex, outputIndexMultiplier, outputSize) {
const rowSplitSize = rowSplit.length;
const result = [];
for (let i = 0; i < rowSplitSize - 1; ++i) {
const rowLength = rowSplit[i + 1] - rowSplit[i];
let realLength = Math.min(outputSize, rowLength);
let parentOutputIndexCurrent = parentOutputIndex[i];
if (parentOutputIndexCurrent === -1) {
realLength = 0;
}
for (let j = 0; j < realLength; ++j) {
result.push(parentOutputIndexCurrent);
parentOutputIndexCurrent += outputIndexMultiplier;
}
for (let j = 0; j < rowLength - realLength; ++j) {
result.push(-1);
}
}
if (rowSplitSize > 0 && result.length !== rowSplit[rowSplitSize - 1]) {
throw new Error('Invalid row split size.');
}
return result;
}
// Calculate the output index of the first element of a list.
// The parentOutputIndex is the same computation for the previous list.
// -1 indicates an element or list that is out of range.
// The outputIndexMultiplier is the number of output indices one moves
// forward for each column.
// E.g., given:
// valueRowIds:[0 1 2 2 2 3 5 5 6]
// parentOutputIndex:[1000 1100 2000 2100 -1 3000 4000]
// outputIndexMultiplier: 10
// outputSize: 2
// You get:
// result = [1000 1100 2000 2010 -1 2100 -1 -1 3000]
// result[0] = parentOutputIndex[valueRowIds[0]]
// result[1] = parentOutputIndex[valueRowIds[1]]
// result[2] = parentOutputIndex[valueRowIds[2]]
// result[3] = parentOutputIndex[valueRowIds[2] + 10]
// result[4] = -1 because it is the third element the size is 2.
// result[5] = parentOutputIndex[valueRowIds[3]]
// result[6] = -1 because parentOutputIndex[valueRowIds[6]] == -1
// result[7] = -1 because parentOutputIndex[valueRowIds[6]] == -1
// result[8] = parentOutputIndex[valueRowIds[7]]
calculateOutputIndexValueRowID(valueRowIds, parentOutputIndex, outputIndexMultiplier, outputSize) {
const indexSize = valueRowIds.length;
const result = [];
if (indexSize === 0) {
return [];
}
let currentOutputColumn = 0;
let currentValueRowId = valueRowIds[0];
if (currentValueRowId >= parentOutputIndex.length) {
throw new Error(`Got currentValueRowId=${currentValueRowId}, which is not less than ${parentOutputIndex.length}`);
}
let currentOutputIndex = parentOutputIndex[currentValueRowId];
result.push(currentOutputIndex);
for (let i = 1; i < indexSize; ++i) {
const nextValueRowId = valueRowIds[i];
if (nextValueRowId === currentValueRowId) {
if (currentOutputIndex >= 0) {
++currentOutputColumn;
if (currentOutputColumn < outputSize) {
currentOutputIndex += outputIndexMultiplier;
}
else {
currentOutputIndex = -1;
}
}
}
else {
currentOutputColumn = 0;
currentValueRowId = nextValueRowId;
if (nextValueRowId >= parentOutputIndex.length) {
throw new Error(`Got nextValueRowId=${nextValueRowId} which is not less than ${parentOutputIndex.length}`);
}
currentOutputIndex = parentOutputIndex[nextValueRowId];
}
result.push(currentOutputIndex);
}
if (result.length !== valueRowIds.length) {
throw new Error('Invalid row ids.');
}
return result;
}
calculateOutputIndex(dimension, parentOutputIndex, outputIndexMultiplier, outputSize) {
const rowPartitionTensor = this.getRowPartitionTensor(dimension);
const partitionType = this.getRowPartitionTypeByDimension(dimension);
switch (partitionType) {
case RowPartitionType.VALUE_ROWIDS:
return this.calculateOutputIndexValueRowID(rowPartitionTensor, parentOutputIndex, outputIndexMultiplier, outputSize);
case RowPartitionType.ROW_SPLITS:
if (rowPartitionTensor.length - 1 > parentOutputIndex.length) {
throw new Error(`Row partition size is greater than output size: ${rowPartitionTensor.length - 1} > ${parentOutputIndex.length}`);
}
return this.calculateOutputIndexRowSplit(rowPartitionTensor, parentOutputIndex, outputIndexMultiplier, outputSize);
default:
throw new Error(`Unsupported partition type: ${RowPartitionType[partitionType]}`);
}
}
getFirstDimensionSize() {
const firstPartitionTensor = this.rowPartitionValues[0];
if (this.rowPartitionTypes.length === 0) {
throw new Error('No row_partition_types given.');
}
const firstPartitionType = this.rowPartitionTypes[0];
switch (firstPartitionType) {
case RowPartitionType.FIRST_DIM_SIZE:
return firstPartitionTensor[0];
case RowPartitionType.VALUE_ROWIDS:
throw new Error('Cannot handle VALUE_ROWIDS in first dimension.');
case RowPartitionType.ROW_SPLITS:
return this.rowPartitionValuesShapes[0][0] - 1;
default:
throw new Error(`Cannot handle type ${RowPartitionType[firstPartitionType]}`);
}
}
compute() {
const firstPartitionTensor = this.rowPartitionValues[0];
if (firstPartitionTensor.length <= 0) {
throw new Error('Invalid first partition input. ' +
'Tensor requires at least one element.');
}
const firstDimension = this.getFirstDimensionSize();
const outputSize = this.calculateOutputSize(firstDimension);
const multiplier = new Array(this.raggedRank + 1);
multiplier[multiplier.length - 1] = 1;
for (let i = multiplier.length - 2; i >= 0; --i) {
multiplier[i] = multiplier[i + 1] * outputSize[i + 1];
}
// Full size of the tensor.
const outputShape = makeShape(outputSize, false);
const outputTensor = getArrayFromDType(this.valuesDType, sizeFromShape(outputShape));
const fullSize = multiplier[0] * outputSize[0];
if (fullSize > 0) {
let outputIndex = this.calculateFirstParentOutputIndex(firstDimension, multiplier[0], outputSize[0]);
for (let i = 1; i <= this.raggedRank; ++i) {
const newOutputIndex = this.calculateOutputIndex(i - 1, outputIndex, multiplier[i], outputSize[i]);
outputIndex = newOutputIndex;
}
this.setOutput(this.raggedRank, outputIndex, outputTensor, outputShape);
}
return [outputShape, outputTensor];
}
setOutput(raggedRank, outputIndex, outputTensor, outputShape) {
if (outputTensor.length === 0) {
return;
}
const valuesBase = this.values;
const outputBase = outputTensor;
let elementShape = outputShape.slice();
elementShape = elementShape.slice(raggedRank + 1);
const valueElementSize = sizeFromShape(elementShape);
const outputIndexSize = outputIndex.length;
// Broadcast the default value to value_element_size. (We can skip this
// if defaultValueTensor.size == 1, since we use fill when that's true.)
let defaultValue = this.defaultValue;
if (defaultValue.length !== valueElementSize && defaultValue.length !== 1) {
const srcShape = this.defaultValueShape;
tidy(() => {
const defaultValueTensor = reshape$2(defaultValue, srcShape);
const bCastDefault = broadcastTo(defaultValueTensor, elementShape);
defaultValue = bCastDefault.dataSync();
});
}
// Loop through the outputIndex array, finding contiguous regions that
// should be copied. Once we find the end of a contiguous region, copy it
// and add any necessary padding (with defaultValue).
let srcStart = 0; // Start of contiguous region (in values)
let dstStart = 0; // Destination for contiguous region (in output)
let dstEnd = 0; // Destination for contiguous region (in output)
for (let srcI = 0; srcI <= outputIndexSize; ++srcI) {
// dstI is the destination where the value at srcI should be copied.
let dstI = srcI < outputIndexSize ? outputIndex[srcI] : -1;
// If we're still in a contiguous region, then update dstEnd go to the
// next srcI.
if (dstI === dstEnd) {
++dstEnd;
continue;
}
// We found the end of contiguous region. This can be because we found
// a gap (dstI > dstEnd), or a source value that shouldn't be copied
// because it's out-of-bounds (dstI == -1), or the end of the tensor
// (dstI === -1).
if (dstStart < dstEnd) {
// Copy the contiguous region.
const src = valuesBase.subarray(srcStart * valueElementSize);
const dst = outputBase.subarray(dstStart * valueElementSize);
const nVals = (dstEnd - dstStart) * valueElementSize;
copyArray(dst, src, nVals);
}
// Add any necessary padding (w/ defaultValue).
if (srcI >= outputIndexSize) {
// We reached the end of values: pad to the end of output.
const outputSize = outputTensor.length;
dstI = Math.floor(outputSize / valueElementSize);
}
if (dstI > dstEnd) {
if (this.defaultValue.length === 1) {
outputBase
.subarray(dstEnd * valueElementSize, dstI * valueElementSize)
.fill(this.defaultValue[0]);
dstEnd = dstI;
}
else {
while (dstI > dstEnd) {
const dst = outputBase.slice(dstEnd * valueElementSize);
copyArray(dst, defaultValue, valueElementSize);
++dstEnd;
}
}
}
// Update indices.
if (dstI < 0) {
// srcI should be skipped -- leave it out of the contiguous region.
srcStart = srcI + 1;
dstStart = dstEnd;
}
else {
// srcI should be copied -- include it in the contiguous region.
srcStart = srcI;
dstStart = dstEnd;
dstEnd = dstStart + 1;
}
}
}
}
function copyArray(dst, src, size) {
for (let i = 0; i < size; i++) {
dst[i] = src[i];
}
}
function makeShape(shape, isPartial) {
const out = [];
for (let dim of shape) {
if (dim < 0) {
if (!isPartial) {
throw new Error(`Dimension ${dim} must be >= 0`);
}
if (dim < -1) {
throw new Error(`Dimension ${dim} must be >= -1`);
}
dim = -1;
}
out.push(dim);
}
return out;
}
function raggedTensorToTensorImpl(shape, shapesShape, values, valuesShape, valuesDType, defaultValue, defaultValueShape, rowPartitionValues, rowPartitionValuesShapes, rowPartitionTypes) {
return new RaggedTensorToTensorOp(shape, shapesShape, values, valuesShape, valuesDType, defaultValue, defaultValueShape, rowPartitionValues, rowPartitionValuesShapes, rowPartitionTypes)
.compute();
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function rangeImpl(start, stop, step, dtype) {
const sameStartStop = start === stop;
const increasingRangeNegativeStep = start < stop && step < 0;
const decreasingRangePositiveStep = stop < start && step > 1;
if (sameStartStop || increasingRangeNegativeStep ||
decreasingRangePositiveStep) {
return makeZerosTypedArray(0, dtype);
}
const numElements = Math.abs(Math.ceil((stop - start) / step));
const values = makeZerosTypedArray(numElements, dtype);
if (stop < start && step === 1) {
// Auto adjust the step's sign if it hasn't been set
// (or was set to 1)
step = -1;
}
values[0] = start;
for (let i = 1; i < values.length; i++) {
values[i] = values[i - 1] + step;
}
return values;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const rsqrtImpl = createSimpleUnaryImpl((xi) => 1 / Math.sqrt(xi));
const rsqrt$1 = unaryKernelFuncFromImpl(Rsqrt, rsqrtImpl);
const rsqrtConfig$1 = {
kernelName: Rsqrt,
backendName: 'cpu',
kernelFunc: rsqrt$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function scatterImpl(indices, updates, shape, outputSize, sliceSize, numUpdates, sliceRank, strides, defaultValue, sumDupeIndices) {
const flattenShape = [outputSize / sliceSize, sliceSize];
const indicesData = indices.values;
const updatesData = updates.values;
if (outputSize === 0) {
return buffer(shape, updates.dtype);
}
const outBuf = (defaultValue instanceof TensorBuffer) ?
defaultValue :
buffer(flattenShape, updates.dtype);
if (typeof defaultValue === 'string') {
outBuf.values.fill(defaultValue);
}
else if (typeof defaultValue === 'number') {
outBuf.values.fill(defaultValue);
}
else if (typeof defaultValue === 'boolean') {
outBuf.values.fill(+defaultValue);
}
for (let i = 0; i < numUpdates; i++) {
const index = [];
let flattenIndex = 0;
for (let j = 0; j < sliceRank; j++) {
const dim = indicesData[i * sliceRank + j];
index.push(dim);
flattenIndex += dim * strides[j];
}
if (flattenIndex < 0 || flattenIndex >= outputSize / sliceSize) {
throw new Error(`Invalid indices: ${index} does not index into ${shape}`);
}
for (let k = 0; k < sliceSize; k++) {
if (sumDupeIndices) {
outBuf.values[flattenIndex * sliceSize + k] +=
updatesData[i * sliceSize + k];
}
else {
outBuf.values[flattenIndex * sliceSize + k] = updates.rank === 0 ?
updatesData[0] :
updatesData[i * sliceSize + k];
}
}
}
return outBuf;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sigmoidImpl = createSimpleUnaryImpl((xi) => 1 / (1 + Math.exp(-xi)));
const sigmoid$1 = unaryKernelFunc$1(Sigmoid$1, (xi) => 1 / (1 + Math.exp(-xi)));
const sigmoidConfig$1 = {
kernelName: Sigmoid$1,
backendName: 'cpu',
kernelFunc: sigmoid$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sliceImpl(vals, begin, size, shape, dtype) {
const isContinous = isSliceContinous(shape, begin, size);
const length = sizeFromShape(size);
const xStrides = computeStrides(shape);
if (isContinous) {
const flatOffset = computeFlatOffset(begin, xStrides);
if (dtype === 'string') {
return vals.slice(flatOffset, flatOffset + length);
}
return vals.subarray(flatOffset, flatOffset + length);
}
const decodedData = dtype === 'string' ?
fromUint8ToStringArray(vals) :
vals;
const inBuf = buffer(shape, dtype, decodedData);
const outBuf = buffer(size, dtype);
for (let i = 0; i < outBuf.size; ++i) {
const outLoc = outBuf.indexToLoc(i);
const inLoc = outLoc.map((idx, j) => idx + begin[j]);
outBuf.set(inBuf.get(...inLoc), ...outLoc);
}
if (dtype === 'string') {
return fromStringArrayToUint8(outBuf.values);
}
return outBuf.values;
}
function slice$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { begin, size } = attrs;
assertNotComplex(x, 'slice');
const [$begin, $size] = parseSliceParams(x, begin, size);
assertParamsValid(x, $begin, $size);
const vals = backend.data.get(x.dataId).values;
const outVals = sliceImpl(vals, $begin, $size, x.shape, x.dtype);
return backend.makeTensorInfo($size, x.dtype, outVals);
}
const sliceConfig$1 = {
kernelName: Slice,
backendName: 'cpu',
kernelFunc: slice$1
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseFillEmptyRowsImpl(indices, indicesShape, indicesDType, values, valuesDType, denseShape, defaultValue) {
const indicesCount = indicesShape[0];
const denseRows = denseShape[0];
const emptyRowIndicator = new Array(denseRows);
const reverseIndexMap = new Array(indicesCount);
const rank = indicesShape[1];
if (denseRows === 0) {
if (indicesCount !== 0) {
throw new Error(getSparseFillEmptyRowsIndicesDenseShapeMismatch(indicesCount));
}
const outputIndices = getArrayFromDType(indicesDType, 0);
const outputValues = getArrayFromDType(valuesDType, 0);
return [
outputIndices, [0, rank], outputValues, emptyRowIndicator, reverseIndexMap
];
}
let rowsAreOrdered = true;
let lastIndicesRow = 0;
const csrOffset = new Array(denseRows).fill(0);
for (let i = 0; i < indicesCount; ++i) {
// indices is a 2d tensor with shape of [N, rank]
const row = indices[i * rank];
if (row < 0) {
throw new Error(getSparseFillEmptyRowsNegativeIndexErrorMessage(i, row));
}
if (row >= denseRows) {
throw new Error(getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(i, row, denseRows));
}
++csrOffset[row];
rowsAreOrdered = rowsAreOrdered && (row >= lastIndicesRow);
lastIndicesRow = row;
}
let allRowsFull = true;
for (let row = 0; row < denseRows; ++row) {
// csrOffset here describes the number of elements in this dense row
const rowEmpty = (csrOffset[row] === 0);
emptyRowIndicator[row] = rowEmpty;
allRowsFull = allRowsFull && !rowEmpty;
// In filled version, each row has at least one element.
csrOffset[row] = Math.max(csrOffset[row], 1);
// Update csrOffset to represent the number of elements up to and
// including denseRows + 1:
// csrOffset[0] == #{elements of row 0}
// csrOffset[1] == #{elements of row 1} + #{elements of row 0}
// ..
// csrOffset[i] == starting index for elements in row i + 1.
if (row > 0) {
csrOffset[row] += csrOffset[row - 1];
}
}
if (allRowsFull && rowsAreOrdered) {
const outputIndices = indices;
const outputValues = values;
for (let i = 0; i < indicesCount; ++i) {
reverseIndexMap[i] = i;
}
return [
outputIndices, [indicesCount, rank], outputValues, emptyRowIndicator,
reverseIndexMap
];
}
else {
const fullIndicesCount = csrOffset[denseRows - 1];
const outputIndices = getArrayFromDType(indicesDType, fullIndicesCount * rank);
const outputValues = getArrayFromDType(valuesDType, fullIndicesCount);
const filledCount = new Array(denseRows).fill(0);
// Fill in values for rows that are not missing
for (let i = 0; i < indicesCount; ++i) {
// indices is a 2d tensor with shape of [N, rank]
const row = indices[i * rank];
const offset = filledCount[row];
const outputI = ((row === 0) ? 0 : csrOffset[row - 1]) + offset;
filledCount[row]++; // Increment the filled count for this row.
for (let j = 0; j < rank; ++j) {
// indices and outputIndices are 2d tensors with shape of [N, rank]
outputIndices[outputI * rank + j] = indices[i * rank + j];
}
outputValues[outputI] = values[i];
// We'll need this reverse index map to backprop correctly.
reverseIndexMap[i] = outputI;
}
// Fill in values for rows that are missing
for (let row = 0; row < denseRows; ++row) {
const rowCount = filledCount[row];
if (rowCount === 0) { // We haven't filled this row
const startingIndex = (row === 0) ? 0 : csrOffset[row - 1];
// Remaining index values were set to zero already.
// Just need to set the row index in the right location.
// outputIndices is a 2d tensor with shape of [N, rank]
outputIndices[startingIndex * rank + 0] = row;
for (let col = 1; col < rank; ++col) {
outputIndices[startingIndex * rank + col] = 0;
}
outputValues[startingIndex] = defaultValue;
}
}
return [
outputIndices, [fullIndicesCount, rank], outputValues, emptyRowIndicator,
reverseIndexMap
];
}
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseReshapeImpl(inputIndices, inputIndicesShape, inputDType, inputShape, targetShape) {
const denseSize = sizeFromShape(inputShape);
const nnz = inputIndicesShape[0];
const outputRank = targetShape.length;
// Compute the output shape. Determine product of specified dimensions, and
// find the index of the unspecified one.
const outputShape = [];
let product = 1;
let unknownIndex = -1;
for (let d = 0; d < outputRank; ++d) {
const size = targetShape[d];
if (size === -1) {
if (unknownIndex !== -1) {
throw new Error(getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(unknownIndex, d));
}
unknownIndex = d;
outputShape.push(1);
}
else {
if (size < 0) {
throw new Error(getSparseReshapeNegativeOutputDimErrorMessage(d, size));
}
product *= size;
outputShape.push(size);
}
}
if (unknownIndex !== -1) {
if (product <= 0) {
throw new Error(getSparseReshapeEmptyTensorZeroOutputDimErrorMessage());
}
const missing = Math.trunc(denseSize / product);
if (product * missing !== denseSize) {
throw new Error(getSparseReshapeInputOutputMultipleErrorMessage(inputShape, outputShape));
}
outputShape[unknownIndex] = missing;
}
const outputSize = sizeFromShape(outputShape);
if (outputSize !== denseSize) {
throw new Error(getSparseReshapeInputOutputMismatchErrorMessage(inputShape, outputShape));
}
const inputRank = inputShape.length;
const inputStrides = [];
if (inputRank > 0) {
inputStrides[inputRank - 1] = 1;
for (let d = inputRank - 2; d >= 0; --d) {
inputStrides[d] = inputStrides[d + 1] * inputShape[d + 1];
}
}
const outputStrides = [];
if (outputRank > 0) {
outputStrides[outputRank - 1] = 1;
for (let d = outputRank - 2; d >= 0; --d) {
outputStrides[d] = outputStrides[d + 1] * outputShape[d + 1];
}
}
const newIndices = getArrayFromDType(inputDType, nnz * outputRank);
for (let i = 0; i < nnz; ++i) {
let id = 0;
for (let j = 0; j < inputRank; ++j) {
// inputIndices is a 2d tensor with shape of [nnz, inputRank]
id += inputIndices[i * inputRank + j] * inputStrides[j];
}
for (let j = 0; j < outputRank; ++j) {
// newIndices is a 2d tensor with shape of [nnz, outputRank]
newIndices[i * outputRank + j] = Math.trunc(id / outputStrides[j]);
id %= outputStrides[j];
}
}
return [newIndices, [nnz, outputRank], outputShape];
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseSegmentReductionImpl(input, inputShape, inputDType, indices, segmentIds, isMean = false, defaultValue = 0) {
const numIndices = indices.length;
// Flatten the array to two dimensions
const inputFlat = [inputShape[0], input.length / inputShape[0]];
const numCol = inputFlat[1];
// Note that the current implementation assumes that segmentIds values are
// sorted.
const lastSegmentIdPlusOne = numIndices > 0 ? segmentIds[numIndices - 1] + 1 : 0;
const outputRows = lastSegmentIdPlusOne;
if (outputRows < 0) {
throw new Error(getSparseSegmentReductionNegativeSegmentIdsErrorMessage());
}
const outputShape = inputShape.slice();
outputShape[0] = outputRows;
const outputLength = outputShape.reduce((product, value) => product * value, 1);
// Output array is initialized with the value 0 by default.
const output = getArrayFromDType(inputDType, outputLength);
// Note that we do not initialize the output buffer with a default value, so
// we need to explicitly set missing indices to the default value.
if (numIndices === 0) {
if (outputRows > 0) {
output.fill(defaultValue);
}
return [output, outputShape];
}
if (outputRows <= 0) {
throw new Error(getSparseSegmentReductionNegativeSegmentIdsErrorMessage());
}
let start = 0, end = 1;
// Index from which the output is not initialized.
let uninitializedIndex = 0;
let outIndex = segmentIds[start];
while (true) {
// We initialize nextIndex to 0 to avoid may be uninitialized warning
let nextIndex = 0;
if (end < numIndices) {
nextIndex = segmentIds[end];
if (outIndex === nextIndex) {
++end;
continue;
}
// We have a new segment here. Verify that the segment ids are growing.
if (outIndex >= nextIndex) {
throw new Error(getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage());
}
}
if (outIndex < 0 || outIndex >= outputRows) {
throw new Error(getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(outIndex, outputRows));
}
// If there is a gap between two indices, we need to set that gap to the
// default value.
if (outIndex > uninitializedIndex) {
output.fill(defaultValue, uninitializedIndex * numCol, outIndex * numCol);
}
for (let i = start; i < end; ++i) {
const index = indices[i];
if (index < 0 || index >= inputFlat[0]) {
throw new Error(getSparseSegmentReductionIndicesOutOfRangeErrorMessage(i, indices[i], inputFlat[0]));
}
for (let j = 0; j < numCol; j++) {
output[outIndex * numCol + j] += input[index * numCol + j];
}
}
if (isMean) {
for (let j = 0; j < numCol; j++) {
output[outIndex * numCol + j] /= end - start;
}
}
start = end;
++end;
uninitializedIndex = outIndex + 1;
outIndex = nextIndex;
if (end > numIndices) {
break;
}
}
// Fill the gap at the end with the default value.
if (uninitializedIndex < outputRows) {
output.fill(defaultValue, uninitializedIndex * numCol, outputRows * numCol);
}
return [output, outputShape];
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sqrtImpl = createSimpleUnaryImpl((xi) => Math.sqrt(xi));
const sqrt$1 = unaryKernelFunc$1(Sqrt, (xi) => Math.sqrt(xi));
const sqrtConfig$1 = {
kernelName: Sqrt,
backendName: 'cpu',
kernelFunc: sqrt$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const squaredDifferenceImpl = createSimpleBinaryKernelImpl(((a, b) => {
const diff = a - b;
return diff * diff;
}));
const squaredDifference$1 = binaryKernelFunc$1(SquaredDifference, squaredDifferenceImpl);
const squaredDifferenceConfig$1 = {
kernelName: SquaredDifference,
backendName: 'cpu',
kernelFunc: squaredDifference$1
};
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const staticRegexReplaceImpl = createSimpleUnaryImpl((x, attrs) => {
const { pattern, replaceGlobal, rewrite } = attrs;
// TODO(mattSoulanille): Don't create a regex each time.
return x.replace(new RegExp(pattern, replaceGlobal ? 'g' : ''), rewrite);
});
const staticRegexReplace$1 = unaryKernelFuncFromImpl(StaticRegexReplace, staticRegexReplaceImpl);
const staticRegexReplaceConfig$1 = {
kernelName: StaticRegexReplace,
backendName: 'cpu',
kernelFunc: staticRegexReplace$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stridedSliceImpl(outShape, xBuf, strides, begin) {
const outBuf = buffer(outShape, xBuf.dtype);
for (let i = 0; i < outBuf.size; i++) {
const loc = outBuf.indexToLoc(i);
const newLoc = new Array(loc.length);
for (let j = 0; j < newLoc.length; j++) {
newLoc[j] = loc[j] * strides[j] + begin[j];
}
outBuf.set(xBuf.get(...newLoc), ...loc);
}
return outBuf;
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* The StringNGramsOp class creates ngrams from ragged string data.
* The constructor contains all attributes related to the operation such as
* padding widths and strings, and the compute function can be used to
* compute the ngrams for different ragged tensor inputs.
*/
class StringNGramsOp {
constructor(separator, nGramWidths, leftPad, rightPad, padWidth, preserveShortSequences) {
this.separator = encodeString(separator);
this.nGramWidths = nGramWidths;
this.leftPad = encodeString(leftPad);
this.rightPad = encodeString(rightPad);
this.padWidth = padWidth;
this.preserveShort = preserveShortSequences;
}
getPadWidth(nGramWidth) {
// Ngrams can be padded with either a fixed pad width or a dynamic pad
// width depending on the 'padWidth' arg, but in no case should the padding
// ever be wider than 'nGramWidth' - 1.
return Math.min(this.padWidth < 0 ? nGramWidth - 1 : this.padWidth, nGramWidth - 1);
}
getNumNGrams(length, nGramWidth) {
const padWidth = this.getPadWidth(nGramWidth);
return Math.max(0, ((length + 2 * padWidth) - nGramWidth) + 1);
}
createNGrams(data, splitIndex, output, outputStartIndex, numNGrams, nGramWidth) {
for (let nGramIndex = 0; nGramIndex < numNGrams; ++nGramIndex) {
const padWidth = this.getPadWidth(nGramWidth);
const leftPadding = Math.max(0, padWidth - nGramIndex);
const rightPadding = Math.max(0, padWidth - (numNGrams - (nGramIndex + 1)));
const numTokens = nGramWidth - (leftPadding + rightPadding);
const dataStartIndex = splitIndex + (leftPadding > 0 ? 0 : nGramIndex - padWidth);
// Calculate the total expected size of the nGram so we can reserve the
// correct amount of space in the string.
let nGramSize = 0;
// Size of the left padding.
nGramSize += leftPadding * this.leftPad.length;
// Size of the tokens.
for (let n = 0; n < numTokens; ++n) {
nGramSize += data[dataStartIndex + n].length;
}
// Size of the right padding.
nGramSize += rightPadding * this.rightPad.length;
// Size of the separators.
const numSeparators = leftPadding + rightPadding + numTokens - 1;
nGramSize += numSeparators * this.separator.length;
// Build the nGram.
output[outputStartIndex + nGramIndex] = new Uint8Array(nGramSize);
const nGram = output[outputStartIndex + nGramIndex];
let nextNGramIndex = 0;
const appendToNGram = (str) => str.forEach((value) => nGram[nextNGramIndex++] = value);
for (let n = 0; n < leftPadding; ++n) {
appendToNGram(this.leftPad);
appendToNGram(this.separator);
}
// Only output first numTokens - 1 pairs of data and separator
for (let n = 0; n < numTokens - 1; ++n) {
appendToNGram(data[dataStartIndex + n]);
appendToNGram(this.separator);
}
// Handle case when there are no tokens or no right padding as these
// can result in consecutive separators.
if (numTokens > 0) {
// If we have tokens, then output last and then pair each separator
// with the right padding that follows, to ensure nGram ends either with
// the token or with the right pad.
appendToNGram(data[dataStartIndex + numTokens - 1]);
for (let n = 0; n < rightPadding; ++n) {
appendToNGram(this.separator);
appendToNGram(this.rightPad);
}
}
else {
// If we don't have tokens, then the last item inserted into the nGram
// has been the separator from the left padding loop above. Hence,
// output right pad and separator and make sure to finish with a
// padding, not a separator.
for (let n = 0; n < rightPadding - 1; ++n) {
appendToNGram(this.rightPad);
appendToNGram(this.separator);
}
appendToNGram(this.rightPad);
}
}
}
// Data and splits together form the definition of the ragged tensor,
// where data is 1 dimensional and contains the values of the tensor
// and splits denotes the indices at which each row starts.
compute(data, splits) {
// Validate that the splits are valid indices into data, only if there are
// splits specified.
const inputDataSize = data.length;
const splitsSize = splits.length;
if (splitsSize > 0) {
let prevSplit = splits[0];
if (prevSplit !== 0) {
throw new Error(`First split value must be 0, got ${prevSplit}`);
}
for (let i = 1; i < splitsSize; ++i) {
let validSplits = splits[i] >= prevSplit;
validSplits = validSplits && (splits[i] <= inputDataSize);
if (!validSplits) {
throw new Error(`Invalid split value ${splits[i]}, must be in [${prevSplit}, ${inputDataSize}]`);
}
prevSplit = splits[i];
}
if (prevSplit !== inputDataSize) {
throw new Error(`Last split value must be data size. Expected ${inputDataSize}, got ${prevSplit}`);
}
}
const numBatchItems = splitsSize - 1;
const nGramsSplits = getArrayFromDType('int32', splitsSize);
// If there is no data or size, return an empty ragged tensor.
if (inputDataSize === 0 || splitsSize === 0) {
const empty = new Array(inputDataSize);
for (let i = 0; i <= numBatchItems; ++i) {
nGramsSplits[i] = 0;
}
return [empty, nGramsSplits];
}
nGramsSplits[0] = 0;
for (let i = 1; i <= numBatchItems; ++i) {
const length = splits[i] - splits[i - 1];
let numNGrams = 0;
this.nGramWidths.forEach((nGramWidth) => {
numNGrams += this.getNumNGrams(length, nGramWidth);
});
if (this.preserveShort && length > 0 && numNGrams === 0) {
numNGrams = 1;
}
nGramsSplits[i] = nGramsSplits[i - 1] + numNGrams;
}
const nGrams = new Array(nGramsSplits[numBatchItems]);
for (let i = 0; i < numBatchItems; ++i) {
const splitIndex = splits[i];
let outputStartIdx = nGramsSplits[i];
this.nGramWidths.forEach((nGramWidth) => {
const length = splits[i + 1] - splits[i];
const numNGrams = this.getNumNGrams(length, nGramWidth);
this.createNGrams(data, splitIndex, nGrams, outputStartIdx, numNGrams, nGramWidth);
outputStartIdx += numNGrams;
});
// If we're preserving short sequences, check to see if no sequence was
// generated by comparing the current output start idx to the original
// one (nGramSplitsdata). If no ngrams were generated, then they will
// be equal (since we increment outputStartIdx by numNGrams every
// time we create a set of ngrams.)
if (this.preserveShort && outputStartIdx === nGramsSplits[i]) {
const dataLength = splits[i + 1] - splits[i];
// One legitimate reason to not have any ngrams when this.preserveShort
// is true is if the sequence itself is empty. In that case, move on.
if (dataLength === 0) {
continue;
}
// We don't have to worry about dynamic padding sizes here: if padding
// was dynamic, every sequence would have had sufficient padding to
// generate at least one nGram.
const nGramWidth = dataLength + 2 * this.padWidth;
const numNGrams = 1;
this.createNGrams(data, splitIndex, nGrams, outputStartIdx, numNGrams, nGramWidth);
}
}
return [nGrams, nGramsSplits];
}
}
function stringNGramsImpl(data, dataSplits, separator, nGramWidths, leftPad, rightPad, padWidth, preserveShortSequences) {
return new StringNGramsOp(separator, nGramWidths, leftPad, rightPad, padWidth, preserveShortSequences)
.compute(data, dataSplits);
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function split(str, delimiters, skipEmpty, result) {
if (!str.length) {
return;
}
// When the delimiter is empty, the input is split into individual characters.
if (delimiters.length === 0) {
for (let i = 0; i < str.length; ++i) {
result.push(str.subarray(i, i + 1));
}
return;
}
// When there is one delimiter, the input is split only at that delimiter.
if (delimiters.length === 1) {
const delimiter = delimiters[0];
let f = str.indexOf(delimiter);
while (f !== -1) {
const token = str.subarray(0, f);
if (!skipEmpty || token.length !== 0) {
result.push(token);
}
str = str.subarray(f + 1);
f = str.indexOf(delimiter);
}
if (!skipEmpty || str.length !== 0) {
result.push(str);
}
return;
}
// When there are multiple delimiters, the input is split at every instance
// one of the delimiters appears.
let tokenStart = 0;
for (let i = 0; i < str.length + 1; i++) {
if ((i === str.length) || (delimiters.indexOf(str[i]) !== -1)) {
const token = str.subarray(tokenStart, i);
if (!skipEmpty || token.length !== 0) {
result.push(token);
}
tokenStart = i + 1;
}
}
}
function stringSplitImpl(input, delimiter, skipEmpty) {
const batchSize = input.length;
// Empty delimiter means split the input character by character.
const tokens = [];
let outputSize = 0;
let maxNumEntries = 0;
const numIndices = new Array(batchSize);
for (let i = 0; i < batchSize; ++i) {
const prevTokensLength = tokens.length;
split(input[i], delimiter, skipEmpty, tokens);
const nEntries = tokens.length - prevTokensLength;
numIndices[i] = nEntries;
outputSize += nEntries;
maxNumEntries = Math.max(maxNumEntries, nEntries);
}
const indices = getArrayFromDType('int32', outputSize * 2);
const values = new Array(outputSize);
const shape = [batchSize, maxNumEntries];
let c = 0;
for (let i = 0; i < batchSize; ++i) {
for (let j = 0; j < numIndices[i]; ++j) {
// indices is a 2d tensor with shape of [outputSize, 2]
indices[c * 2] = i;
indices[c * 2 + 1] = j;
values[c] = tokens[c];
++c;
}
}
return [indices, values, shape];
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stringToHashBucketFastImpl(input, numBuckets) {
const output = getArrayFromDType('int32', input.length);
for (let i = 0; i < input.length; ++i) {
output[i] =
fingerPrint64(input[i]).modulo(numBuckets).getLowBitsUnsigned();
}
return output;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const subImpl = createSimpleBinaryKernelImpl(((aValue, bValue) => aValue - bValue));
const subComplexImpl = createComplexBinaryKernelImpl(((aReal, aImag, bReal, bImag) => {
return { real: aReal - bReal, imag: aImag - bImag };
}));
const sub$1 = binaryKernelFunc$1(Sub, subImpl, subComplexImpl);
const subConfig$1 = {
kernelName: Sub,
backendName: 'cpu',
kernelFunc: sub$1
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* An implementation of the tile kernel shared between webgl and cpu for string
* tensors only.
*/
function tileImpl(xBuf, reps) {
const newShape = new Array(xBuf.rank);
for (let i = 0; i < newShape.length; i++) {
newShape[i] = xBuf.shape[i] * reps[i];
}
const result = buffer(newShape, xBuf.dtype);
for (let i = 0; i < result.values.length; ++i) {
const newLoc = result.indexToLoc(i);
const originalLoc = new Array(xBuf.rank);
for (let j = 0; j < originalLoc.length; j++) {
originalLoc[j] = newLoc[j] % xBuf.shape[j];
}
const originalIndex = xBuf.locToIndex(originalLoc);
result.values[i] = xBuf.values[originalIndex];
}
return result;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** An implementation of the TopK kernel shared between webgl and cpu. */
const comparePair = (a, b) => {
const valueDiff = b.value - a.value;
return valueDiff === 0 ? a.index - b.index : valueDiff;
};
/**
* Partitions array where all elements smaller than the (k+1) smallest element
* are found to the left of it, and all larger to the right of it.
* Based on the Floyd-Rivest Algorithm, ref:
* https://en.wikipedia.org/wiki/Floyd%E2%80%93Rivest_algorithm
* @param array: Array to partition
* @param left: Left index for the interval
* @param right: Right index for the interval
* @param k: Desired index value, where array[k] is the (k+1)th smallest element
* when left = 0
*/
function select$2(array, k, left = 0, right = array.length - 1) {
while (right > left) {
// Use select recursively to sample a smaller set of size s
// the arbitrary constants 600 and 0.5 are used in the original
// version to minimize execution time.
if (right - left > 600) {
const n = right - left + 1;
const i = k - left + 1;
const z = Math.log(n);
const s = 0.5 * Math.exp(2 * z / 3);
const sd = 0.5 * Math.sqrt(z * s * (n - s) / n) * Math.sign(i - n / 2);
const newLeft = Math.max(left, Math.floor(k - i * s / n + sd));
const newRight = Math.min(right, Math.floor(k + (n - i) * s / n + sd));
select$2(array, k, newLeft, newRight);
}
// partition the elements between left and right around t
const t = array[k];
let i = left;
let j = right;
swap(array, left, k);
if (comparePair(array[right], t) > 0) {
swap(array, left, right);
}
while (i < j) {
swap(array, i, j);
i++;
j--;
while (comparePair(array[i], t) < 0) {
i = i + 1;
}
while (comparePair(array[j], t) > 0) {
j = j - 1;
}
}
if (comparePair(array[left], t) === 0) {
swap(array, left, j);
}
else {
j = j + 1;
swap(array, j, right);
}
// Adjust left and right towards the boundaries of the subset
// containing the (k - left + 1)th smallest element.
if (j <= k) {
left = j + 1;
}
if (k <= j) {
right = j - 1;
}
}
}
function topKImpl(x, xShape, xDtype, k, sorted) {
// Reshape into a 2d tensor [batch, lastDim] and compute topk along lastDim.
const lastDim = xShape[xShape.length - 1];
const [batch, size] = [x.length / lastDim, lastDim];
const allTopKVals = getTypedArrayFromDType(xDtype, batch * k);
const allTopKIndices = getTypedArrayFromDType('int32', batch * k);
for (let b = 0; b < batch; b++) {
const offset = b * size;
const vals = x.subarray(offset, offset + size);
let valAndInd = new Array(vals.length);
vals.forEach((value, index) => valAndInd[index] = { value, index });
if (k < valAndInd.length) {
select$2(valAndInd, k);
valAndInd = valAndInd.slice(0, k);
}
if (sorted) {
valAndInd.sort(comparePair);
}
const outOffset = b * k;
const topKVals = allTopKVals.subarray(outOffset, outOffset + k);
const topKIndices = allTopKIndices.subarray(outOffset, outOffset + k);
for (let i = 0; i < k; i++) {
topKVals[i] = valAndInd[i].value;
topKIndices[i] = valAndInd[i].index;
}
}
// Reshape back to the original input shape, except that the last
// dimension is k.
const outputShape = xShape.slice();
outputShape[outputShape.length - 1] = k;
return [
buffer(outputShape, xDtype, allTopKVals),
buffer(outputShape, 'int32', allTopKIndices)
];
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function uniqueImpl(values, axis, shape, dtype) {
// Normalize and validate axis.
const $axis = parseAxisParam(axis, shape)[0];
// Calculate the new shape that is suitable for extracting data along the
// given axis.
//
// The rank is 3.
// The size of the 1st dimension is the size of all the axes < the given axis.
// The size of the 2nd dimension is the same as the size of the given axis.
// The size of the 3rd dimension is the size of all the axes > the given axis.
//
// For example, for a 4D tensor with shape=[2, 3, 5, 4] and axis=2, the
// newShape would be: [2*3, 5, 4].
//
// Note that this is not the final output shape. This will be the shape for an
// intermediate TensorBuffer (see inputBuffer below) to allow us to extract
// values along the given axis. To demonstrate how it works, consider the
// following example:
//
// Input: a 3D tensor, with shape [1, 2, 3]
// [
// [
// [1,2,3],
// [4,5,6]
// ]
// ]
// Axis: 2 (the last axis).
// Along axis 2, we expect to extract 3 tensors: [1,4], [2,5], [3,6].
//
// For this example, newShape would be: [2, 3, 1], where 2 is calculated from
// 1*2. The re-shaped data would look like:
//
// [
// [
// [1], [2], [3]
// ],
// [
// [4], [5], [6]
// ]
// ]
//
// Then, we can construct a 3-level nested loop by the following dimension
// order to extract the values along the axis (dimension1):
// i: dimension1 // 0,1,2 (newShape[1])
// m: dimension0 // 0,1 (newShape[0])
// n: dimension2 // 0 (newShape[2])
//
// m, i, n
// ---------
// Iteration 0: data at [0, 0, 0] => "1"
// Iteration 1: data at [1, 0, 0] => "4"
// We got [1,4].
// Iteration 2: data at [0, 1, 0] => "2"
// Iteration 3: data at [1, 1, 0] => "5"
// We got [2,5].
// Iteration 4: data at [0, 2, 0] => "3"
// Iteration 5: data at [1, 2, 0] => "6"
// We got [3,6].
const newShape = [1, shape[0], 1];
for (let i = 0; i < $axis; i++) {
newShape[0] *= shape[i];
}
newShape[1] = shape[$axis];
for (let i = $axis + 1; i < shape.length; i++) {
newShape[2] *= shape[i];
}
// A map from unique elements (their string representations) to their values
// in "indices" (below).
const uniqueElements = new Map();
// The indices of each unique element in the original tensor along the given
// axis. It is 1D and has the same size as the given axis.
const indices = new Int32Array(shape[$axis]);
// Create a buffer so we can easily extract value at a given location.
const inputBuffer = new TensorBuffer(newShape, dtype, values);
// The indices along the given axis that have unique elements. This is a
// de-duped version of "indices" above.
const uniqueIndices = [];
const is1DTensor = newShape[0] === 1 && newShape[2] === 1;
for (let i = 0; i < shape[$axis]; i++) {
// Extract values along the axis.
let element;
if (is1DTensor) {
// Fast path for 1D tensor input.
element = values[i].toString();
}
else {
const axisValues = [];
for (let m = 0; m < newShape[0]; m++) {
for (let n = 0; n < newShape[2]; n++) {
axisValues.push(inputBuffer.get(m, i, n));
}
}
element = axisValues.join(',');
}
// Dedup and update various indices.
const existingIndex = uniqueElements.get(element);
if (existingIndex != null) {
indices[i] = existingIndex;
}
else {
const uniqueIndex = uniqueElements.size;
uniqueElements.set(element, uniqueIndex);
indices[i] = uniqueIndex;
uniqueIndices.push(i);
}
}
// Now we know where each of the unique elements are located along the axis
// (uniqueIndices). Extract them from input buffer and store them in the
// output buffer.
const outputTmpShape = newShape.slice();
outputTmpShape[1] = uniqueElements.size;
const outputBuffer = new TensorBuffer(outputTmpShape, dtype);
uniqueIndices.forEach((uniqueElementIndex, i) => {
for (let m = 0; m < newShape[0]; m++) {
for (let n = 0; n < newShape[2]; n++) {
outputBuffer.set(inputBuffer.get(m, uniqueElementIndex, n), m, i, n);
}
}
});
// The output shape can be calculated from the input shape with the size of
// the given axis replaced by the number of unique elements along that axis.
const outputShape = shape.slice();
outputShape[$axis] = outputTmpShape[1];
return {
outputValues: outputBuffer.values,
outputShape,
indices,
};
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Shared functionality among backends.
var shared = /*#__PURE__*/Object.freeze({
__proto__: null,
addImpl: addImpl,
bincountImpl: bincountImpl,
bincountReduceImpl: bincountReduceImpl,
bitwiseAndImpl: bitwiseAndImpl,
castImpl: castImpl,
ceilImpl: ceilImpl,
concatImpl: concatImpl$1,
equalImpl: equalImpl,
expImpl: expImpl,
expm1Impl: expm1Impl,
floorDivImpl: floorDivImpl,
floorImpl: floorImpl,
gatherNdImpl: gatherNdImpl,
gatherV2Impl: gatherV2Impl,
greaterEqualImpl: greaterEqualImpl,
greaterImpl: greaterImpl,
lessEqualImpl: lessEqualImpl,
lessImpl: lessImpl,
linSpaceImpl: linSpaceImpl,
logImpl: logImpl,
maxImpl: maxImpl$1,
maximumImpl: maximumImpl,
minimumImpl: minimumImpl,
multiplyImpl: multiplyImpl,
negImpl: negImpl,
notEqualImpl: notEqualImpl,
prodImpl: prodImpl,
raggedGatherImpl: raggedGatherImpl,
raggedRangeImpl: raggedRangeImpl,
raggedTensorToTensorImpl: raggedTensorToTensorImpl,
rangeImpl: rangeImpl,
rsqrtImpl: rsqrtImpl,
scatterImpl: scatterImpl,
sigmoidImpl: sigmoidImpl,
simpleAbsImpl: simpleAbsImpl,
sliceImpl: sliceImpl,
sparseFillEmptyRowsImpl: sparseFillEmptyRowsImpl,
sparseReshapeImpl: sparseReshapeImpl,
sparseSegmentReductionImpl: sparseSegmentReductionImpl,
sqrtImpl: sqrtImpl,
squaredDifferenceImpl: squaredDifferenceImpl,
staticRegexReplaceImpl: staticRegexReplaceImpl,
stridedSliceImpl: stridedSliceImpl,
stringNGramsImpl: stringNGramsImpl,
stringSplitImpl: stringSplitImpl,
stringToHashBucketFastImpl: stringToHashBucketFastImpl,
subImpl: subImpl,
tileImpl: tileImpl,
topKImpl: topKImpl,
transposeImpl: transposeImpl$1,
uniqueImpl: uniqueImpl
});
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Import shared functionality from tfjs-backend-cpu without triggering
// side effects.
// tslint:disable-next-line: no-imports-from-dist
const { addImpl: addImplCPU, bincountImpl: bincountImplCPU, bincountReduceImpl: bincountReduceImplCPU, bitwiseAndImpl: bitwiseAndImplCPU, castImpl: castImplCPU, ceilImpl: ceilImplCPU, concatImpl: concatImplCPU, equalImpl: equalImplCPU, expImpl: expImplCPU, expm1Impl: expm1ImplCPU, floorImpl: floorImplCPU, gatherNdImpl: gatherNdImplCPU, gatherV2Impl: gatherV2ImplCPU, greaterImpl: greaterImplCPU, greaterEqualImpl: greaterEqualImplCPU, lessImpl: lessImplCPU, lessEqualImpl: lessEqualImplCPU, linSpaceImpl: linSpaceImplCPU, logImpl: logImplCPU, maxImpl: maxImplCPU, maximumImpl: maximumImplCPU, minimumImpl: minimumImplCPU, multiplyImpl: multiplyImplCPU, negImpl: negImplCPU, notEqualImpl: notEqualImplCPU, prodImpl: prodImplCPU, raggedGatherImpl: raggedGatherImplCPU, raggedRangeImpl: raggedRangeImplCPU, raggedTensorToTensorImpl: raggedTensorToTensorImplCPU, rangeImpl: rangeImplCPU, rsqrtImpl: rsqrtImplCPU, scatterImpl: scatterImplCPU, sigmoidImpl: sigmoidImplCPU, simpleAbsImpl: simpleAbsImplCPU, sliceImpl: sliceImplCPU, sparseFillEmptyRowsImpl: sparseFillEmptyRowsImplCPU, sparseReshapeImpl: sparseReshapeImplCPU, sparseSegmentReductionImpl: sparseSegmentReductionImplCPU, sqrtImpl: sqrtImplCPU, staticRegexReplaceImpl: staticRegexReplaceImplCPU, stridedSliceImpl: stridedSliceImplCPU, stringNGramsImpl: stringNGramsImplCPU, stringSplitImpl: stringSplitImplCPU, stringToHashBucketFastImpl: stringToHashBucketFastImplCPU, subImpl: subImplCPU, tileImpl: tileImplCPU, topKImpl: topKImplCPU, transposeImpl: transposeImplCPU, uniqueImpl: uniqueImplCPU, } = shared;
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function getVecChannels(name, rank) {
return ['x', 'y', 'z', 'w', 'u', 'v'].slice(0, rank).map(d => `${name}.${d}`);
}
function getChannels(name, rank) {
if (rank === 1) {
return [name];
}
return getVecChannels(name, rank);
}
function getSourceCoords$2(rank, dims) {
if (rank === 1) {
return 'rc';
}
let coords = '';
for (let i = 0; i < rank; i++) {
coords += dims[i];
if (i < rank - 1) {
coords += ',';
}
}
return coords;
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class PackProgram {
constructor(outputShape) {
this.variableNames = ['A'];
this.packedInputs = false;
this.packedOutput = true;
// Only input / output 3D tensors.
this.outputShape = outputShape;
this.rank = outputShape.length;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
if (this.rank === 0) {
this.userCode = `
void main() {
setOutput(vec4(getA(), 0., 0., 0.));
}
`;
}
else {
const channels = getChannels('rc', this.rank);
const dtype = getCoordsDataType(this.rank);
const outOfBoundsCondition = this.getOutOfBoundsCondition(channels);
const setup = this.getSetup(channels);
const output = this.getOutput(channels);
this.userCode = `
void main() {
${dtype} rc = getOutputCoords();
if(${outOfBoundsCondition}) {
setOutput(vec4(0));
} else {
${setup}
setOutput(vec4(${output}));
}
}
`;
}
}
getSourceCoordsArr(dims) {
const coords = [];
for (let row = 0; row <= 1; row++) {
for (let col = 0; col <= 1; col++) {
let coord = `${row === 0 ? 'r' : 'rp1'}, ${col === 0 ? 'c' : 'cp1'}`;
for (let d = 2; d < this.rank; d++) {
coord = `${dims[dims.length - 1 - d]},` + coord;
}
coords.push(coord);
}
}
return coords;
}
getOutOfBoundsCondition(dims) {
if (this.rank === 1) {
return `rc > ${this.enableShapeUniforms ? 'outShape' : this.outputShape[0]}`;
}
let cond = '';
for (let i = this.rank - 2; i < this.rank; i++) {
cond += `${dims[i]} >= ${this.enableShapeUniforms ? `outShape[${i}]` : this.outputShape[i]}`;
if (i < this.rank - 1) {
cond += '||';
}
}
return cond;
}
getSetup(dims) {
if (this.rank === 1) {
return '';
}
const innerDims = dims.slice(-2);
const col = this.enableShapeUniforms ? `outShape[${this.rank} - 1]` :
this.outputShape[this.rank - 1];
const row = this.enableShapeUniforms ? `outShape[${this.rank} - 2]` :
this.outputShape[this.rank - 2];
return `
int r = ${innerDims[0]};
int c = ${innerDims[1]};
int rp1 = r + 1;
int cp1 = c + 1;
bool cEdge = cp1 >= ${col};
bool rEdge = rp1 >= ${row};
`;
}
getOutput(dims) {
const sourceCoords = this.getSourceCoordsArr(dims);
if (this.rank === 1) {
const outShape = this.enableShapeUniforms ? 'outShape' : this.outputShape[0];
return `getA(rc), (rc + 1 >= ${outShape} ? 0. : getA(rc + 1)), 0, 0`;
}
return `getA(${sourceCoords[0]}),
cEdge ? 0. : getA(${sourceCoords[1]}),
rEdge ? 0. : getA(${sourceCoords[2]}),
rEdge || cEdge ? 0. : getA(${sourceCoords[3]})`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ReshapePackedProgram {
constructor(outputShape, inputShape) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
this.customUniforms = [{ name: 'inputShape', type: 'ivec3' }];
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
let mainLoop = ``;
for (let i = 0; i < 4; i++) {
let thisRC = `thisRC = rc;`;
if (i % 2 === 1) {
thisRC += `thisRC.z += 1;`;
}
if (i > 1) {
thisRC += `thisRC.y += 1;`;
}
mainLoop += `
${thisRC}
${i > 0 ? `if(thisRC.y < rows && thisRC.z < cols){` : ''}
int flatIndex = getFlatIndex(thisRC);
ivec3 inputRC = inputCoordsFromReshapedOutCoords(flatIndex);
vec2 inputRCInnerDims = vec2(float(inputRC.y),float(inputRC.z));
result[${i}] =
getChannel(getA(inputRC.x, inputRC.y, inputRC.z), inputRCInnerDims);
${i > 0 ? '}' : ''}
`;
}
this.userCode = `
${getReshapedInputCoords(inputShape, this.enableShapeUniforms)}
${this.enableShapeUniforms ? getFlatIndexFrom3DOutput() :
getFlatIndexFrom3D(outputShape)}
void main() {
ivec3 rc = getOutputCoords();
vec4 result = vec4(0.);
ivec3 thisRC;
int rows = ${this.enableShapeUniforms ? 'outShape[1]' : outputShape[1]};
int cols = ${this.enableShapeUniforms ? 'outShape[2]' : outputShape[2]};
${mainLoop}
setOutput(result);
}
`;
}
}
function getReshapedInputCoords(shape, enableShapeUniforms) {
const coordsFromIndexSnippet = enableShapeUniforms ?
getLogicalCoordinatesFromFlatIndexByUniform(['r', 'c', 'd'], 'inputShape') :
getLogicalCoordinatesFromFlatIndex(['r', 'c', 'd'], shape);
return `
ivec3 inputCoordsFromReshapedOutCoords(int index) {
${coordsFromIndexSnippet}
return ivec3(r, c, d);
}
`;
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class TextureManager {
constructor(gpgpu) {
this.gpgpu = gpgpu;
this.numUsedTextures = 0;
this.numFreeTextures = 0;
this._numBytesAllocated = 0;
// Number of bytes that have been allocated and available for reuse.
this._numBytesFree = 0;
this.freeTextures = {};
this.usedTextures = {};
this.logEnabled = false;
}
acquireTexture(shapeRC, usage, isPacked) {
const physicalTexType = getPhysicalFromLogicalTextureType(usage, isPacked);
const shapeKey = getKeyFromTextureShape(shapeRC, physicalTexType, isPacked);
if (!(shapeKey in this.freeTextures)) {
this.freeTextures[shapeKey] = [];
}
if (!(shapeKey in this.usedTextures)) {
this.usedTextures[shapeKey] = [];
}
const texBytes = computeBytes(shapeRC, physicalTexType, this.gpgpu.gl, this.gpgpu.textureConfig, isPacked);
if (this.freeTextures[shapeKey].length > 0) {
this.numFreeTextures--;
this.numUsedTextures++;
this._numBytesFree -= texBytes;
this.log();
const newTexture = this.freeTextures[shapeKey].pop();
this.usedTextures[shapeKey].push(newTexture);
return newTexture;
}
let newTexture;
if (physicalTexType === PhysicalTextureType.PACKED_2X2_FLOAT32) {
newTexture = this.gpgpu.createPackedMatrixTexture(shapeRC[0], shapeRC[1]);
}
else if (physicalTexType === PhysicalTextureType.PACKED_2X2_FLOAT16) {
newTexture =
this.gpgpu.createFloat16PackedMatrixTexture(shapeRC[0], shapeRC[1]);
}
else if (physicalTexType === PhysicalTextureType.UNPACKED_FLOAT32) {
newTexture =
this.gpgpu.createFloat32MatrixTexture(shapeRC[0], shapeRC[1]);
}
else if (physicalTexType === PhysicalTextureType.UNPACKED_FLOAT16) {
newTexture =
this.gpgpu.createFloat16MatrixTexture(shapeRC[0], shapeRC[1]);
}
else if (physicalTexType === PhysicalTextureType.PACKED_4X1_UNSIGNED_BYTE) {
newTexture =
this.gpgpu.createUnsignedBytesMatrixTexture(shapeRC[0], shapeRC[1]);
}
this.usedTextures[shapeKey].push(newTexture);
this.numUsedTextures++;
this._numBytesAllocated += texBytes;
this.log();
return newTexture;
}
releaseTexture(texture, shape, logicalTexType, isPacked) {
if (this.freeTextures == null) {
// Already disposed.
return;
}
const physicalTexType = getPhysicalFromLogicalTextureType(logicalTexType, isPacked);
const shapeKey = getKeyFromTextureShape(shape, physicalTexType, isPacked);
if (!(shapeKey in this.freeTextures)) {
this.freeTextures[shapeKey] = [];
}
const texBytes = computeBytes(shape, physicalTexType, this.gpgpu.gl, this.gpgpu.textureConfig, isPacked);
const deleteTexThreshold = env()
.getNumber('WEBGL_DELETE_TEXTURE_THRESHOLD');
if (deleteTexThreshold !== -1 &&
this._numBytesAllocated > deleteTexThreshold) {
this.gpgpu.deleteMatrixTexture(texture.texture);
this._numBytesAllocated -= texBytes;
}
else {
this.freeTextures[shapeKey].push(texture);
this.numFreeTextures++;
this._numBytesFree += texBytes;
}
this.numUsedTextures--;
const texList = this.usedTextures[shapeKey];
const texIndex = texList && texList.indexOf(texture);
if (texIndex == null || texIndex < 0) {
throw new Error('Cannot release a texture that was never provided by this ' +
'texture manager');
}
texList[texIndex] = texList[texList.length - 1];
texList.pop();
this.log();
}
log() {
if (!this.logEnabled) {
return;
}
const total = this.numFreeTextures + this.numUsedTextures;
console.log('Free/Used', `${this.numFreeTextures} / ${this.numUsedTextures}`, `(${total})`);
const freeRatio = this._numBytesFree / this._numBytesAllocated;
console.log(`Bytes allocated: ${this._numBytesAllocated}`);
console.log(`Bytes unused: ${this._numBytesFree} (${Math.round(100 * freeRatio)}%)`);
}
get numBytesAllocated() {
return this._numBytesAllocated;
}
get numBytesFree() {
return this._numBytesFree;
}
getNumUsedTextures() {
return this.numUsedTextures;
}
getNumFreeTextures() {
return this.numFreeTextures;
}
dispose() {
if (this.freeTextures == null) {
// Already disposed.
return;
}
for (const texShape in this.freeTextures) {
this.freeTextures[texShape].forEach(tex => {
this.gpgpu.deleteMatrixTexture(tex.texture);
});
}
for (const texShape in this.usedTextures) {
this.usedTextures[texShape].forEach(tex => {
this.gpgpu.deleteMatrixTexture(tex.texture);
});
}
// TODO: Assign non-null value (empty object) to textures after disposed.
this.freeTextures = null;
this.usedTextures = null;
this.numUsedTextures = 0;
this.numFreeTextures = 0;
this._numBytesAllocated = 0;
this._numBytesFree = 0;
}
}
function numBytesForInternalFormat(gl, internalFormat) {
// tslint:disable-next-line:no-any
const glany = gl;
if (internalFormat === glany.R32F) {
return 4;
}
else if (internalFormat === glany.R16F) {
return 2;
}
else if (internalFormat === glany.RGBA32F) {
return 16;
}
else if (internalFormat === gl.RGBA) {
return 16;
}
else if (internalFormat === glany.RGBA16F) {
return 8;
}
else if (internalFormat === glany.RGBA8) {
return 4;
}
throw new Error(`Unknown internal format ${internalFormat}`);
}
function computeBytes(shape, physicalTexType, gl, textureConfig, isPacked) {
// It is not possible to infer packed status from the texture type because
// depending on the textureConfig, different texture types may resolve to the
// same internal format (e.g. in WebGL1, the internal format for
// UNPACKED_FLOAT16 textures is gl.RGBA). Therefore we pass in `isPacked`
// explicitly.
const internalFormat = internalFormatForPhysicalTexType(physicalTexType, textureConfig);
let numElements;
if (isPacked) {
const [packedWidth, packedHeight] = getPackedMatrixTextureShapeWidthHeight(shape[0], shape[1]);
numElements = packedWidth * packedHeight;
}
else {
const [width, height] = getUnpackedMatrixTextureShapeWidthHeight(shape[0], shape[1]);
numElements = width * height;
}
const bytesPerElement = numBytesForInternalFormat(gl, internalFormat);
return numElements * bytesPerElement;
}
function internalFormatForPhysicalTexType(physicalTexType, textureConfig) {
switch (physicalTexType) {
case PhysicalTextureType.PACKED_2X2_FLOAT32:
return getInternalFormatForPackedMatrixTexture(textureConfig);
case PhysicalTextureType.PACKED_2X2_FLOAT16:
return getInternalFormatForFloat16PackedMatrixTexture(textureConfig);
case PhysicalTextureType.UNPACKED_FLOAT32:
return getInternalFormatForFloat32MatrixTexture(textureConfig);
case PhysicalTextureType.UNPACKED_FLOAT16:
return getInternalFormatForFloat16MatrixTexture(textureConfig);
case PhysicalTextureType.PACKED_4X1_UNSIGNED_BYTE:
return getInternalFormatForUnsignedBytesMatrixTexture(textureConfig);
default:
throw new Error(`Unknown physical texture type ${physicalTexType}`);
}
}
function getPhysicalTextureForRendering(isPacked) {
if (env().getBool('WEBGL_RENDER_FLOAT32_ENABLED')) {
if (isPacked) {
return PhysicalTextureType.PACKED_2X2_FLOAT32;
}
return PhysicalTextureType.UNPACKED_FLOAT32;
}
if (isPacked) {
return PhysicalTextureType.PACKED_2X2_FLOAT16;
}
return PhysicalTextureType.UNPACKED_FLOAT16;
}
function getPhysicalFromLogicalTextureType(logicalTexType, isPacked) {
if (logicalTexType === TextureUsage.UPLOAD) {
return PhysicalTextureType.PACKED_2X2_FLOAT32;
}
else if (logicalTexType === TextureUsage.RENDER || logicalTexType == null) {
return getPhysicalTextureForRendering(isPacked);
}
else if (logicalTexType === TextureUsage.DOWNLOAD ||
logicalTexType === TextureUsage.PIXELS) {
return PhysicalTextureType.PACKED_4X1_UNSIGNED_BYTE;
}
throw new Error(`Unknown logical texture type ${logicalTexType}`);
}
function getKeyFromTextureShape(shapeRowsCol, physicalTexType, isPacked) {
return `${shapeRowsCol[0]}_${shapeRowsCol[1]}_${physicalTexType}_${isPacked}`;
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class UnaryOpProgram {
constructor(aShape, opSnippet) {
this.variableNames = ['A'];
this.outputShape = aShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
this.userCode = `
float unaryOperation(float x) {
${opSnippet}
}
void main() {
float x = getAAtOutCoords();
float y = unaryOperation(x);
setOutput(y);
}
`;
}
}
const CHECK_NAN_SNIPPET$1 = `if (isnan(x)) return x;`;
const LINEAR$1 = `return x;`;
const ABS$1 = `return abs(x);`;
const ELU$2 = `return (x >= 0.0) ? x : (exp(x) - 1.0);`;
const RELU$2 = CHECK_NAN_SNIPPET$1 + `
return (x < 0.0) ? 0.0 : x;
`;
const RELU6$2 = CHECK_NAN_SNIPPET$1 + `
return (x < 0.0) ? 0.0 : min(6.0, x);
`;
const CLONE = 'return x;';
const SIGMOID$2 = `return 1.0 / (1.0 + exp(-1.0 * x));`;
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LINEAR = `return x;`;
const ELU$1 = `
vec4 result;
result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);
return result;
`;
const RELU$1 = `
vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : result.r;
result.g = isNaN.g ? x.g : result.g;
result.b = isNaN.b ? x.b : result.b;
result.a = isNaN.a ? x.a : result.a;
return result;
`;
const RELU6$1 = `
vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : result.r;
result.g = isNaN.g ? x.g : result.g;
result.b = isNaN.b ? x.b : result.b;
result.a = isNaN.a ? x.a : result.a;
return result;
`;
const SIGMOID$1 = `return 1.0 / (1.0 + exp(-1.0 * x));`;
class UnaryOpPackedProgram {
constructor(aShape, opSnippet) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = aShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
this.userCode = `
vec4 unaryOperation(vec4 x) {
${opSnippet}
}
void main() {
vec4 x = getAAtOutCoords();
vec4 y = unaryOperation(x);
setOutput(y);
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class UnpackProgram {
constructor(outputShape) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = false;
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
const rank = outputShape.length;
const channels = getChannels('rc', rank);
const dtype = getCoordsDataType(rank);
const sourceCoords = getSourceCoords$2(rank, channels);
const innerDims = channels.slice(-2);
const coords = rank <= 1 ? 'rc' : `vec2(${innerDims.join(',')})`;
this.userCode = `
void main() {
${dtype} rc = getOutputCoords();
vec4 packedInput = getA(${sourceCoords});
setOutput(getChannel(packedInput, ${coords}));
}
`;
}
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Import webgl flags.
const whereImpl$1 = whereImpl$2;
const EPSILON_FLOAT32 = 1e-7;
const EPSILON_FLOAT16 = 1e-4;
const binaryCaches = {};
function getBinaryCache(webGLVersion) {
if (webGLVersion in binaryCaches) {
return binaryCaches[webGLVersion];
}
binaryCaches[webGLVersion] = {};
return binaryCaches[webGLVersion];
}
// Empirically determined constant used to determine size threshold for handing
// off execution to the CPU.
const CPU_HANDOFF_SIZE_THRESHOLD = env().getNumber('CPU_HANDOFF_SIZE_THRESHOLD');
// Empirically determined constant used to decide the number of MB on GPU
// before we warn about high memory use. The MB are this constant * screen area
// * dpi / 1024 / 1024.
const BEFORE_PAGING_CONSTANT = 600;
function numMBBeforeWarning() {
if (env().global.screen == null) {
return 1024; // 1 GB.
}
return (env().global.screen.height * env().global.screen.width *
window.devicePixelRatio) *
BEFORE_PAGING_CONSTANT / 1024 / 1024;
}
class MathBackendWebGL extends KernelBackend {
nextDataId() {
return MathBackendWebGL.nextDataId++;
}
constructor(gpuResource) {
super();
// Maps data ids that have a pending read operation, to list of subscribers.
this.pendingRead = new WeakMap();
// List of data ids that are scheduled for disposal, but are waiting on a
// pending read operation.
this.pendingDisposal = new WeakSet();
// Used to count the number of 'shallow' sliced tensors that point to the
// same data id.
this.dataRefCount = new WeakMap();
this.numBytesInGPU = 0;
// Accumulated time spent (including blocking) in uploading data to webgl.
this.uploadWaitMs = 0;
// Accumulated time spent (including blocking in downloading data from webgl.
this.downloadWaitMs = 0;
// record the last manual GL Flush time.
this.lastGlFlushTime = 0;
this.warnedAboutMemory = false;
this.pendingDeletes = 0;
this.disposed = false;
if (!env().getBool('HAS_WEBGL')) {
throw new Error('WebGL is not supported on this device');
}
let newGPGPU;
if (gpuResource != null) {
if (gpuResource instanceof GPGPUContext) {
newGPGPU = gpuResource;
}
else {
const gl = getWebGLContext(env().getNumber('WEBGL_VERSION'), gpuResource);
newGPGPU = new GPGPUContext(gl);
}
this.binaryCache = {};
this.gpgpuCreatedLocally = false;
}
else {
const gl = getWebGLContext(env().getNumber('WEBGL_VERSION'));
newGPGPU = new GPGPUContext(gl);
this.binaryCache = getBinaryCache(env().getNumber('WEBGL_VERSION'));
this.gpgpuCreatedLocally = true;
}
this.gpgpu = newGPGPU;
this.canvas = this.gpgpu.gl.canvas;
this.textureManager = new TextureManager(this.gpgpu);
this.numMBBeforeWarning = numMBBeforeWarning();
this.texData = new DataStorage(this, engine());
}
numDataIds() {
return this.texData.numDataIds() - this.pendingDeletes;
}
// Writes a new entry to the data store with a WebGL texture, and registers it
// to the texture manager.
writeTexture(texture, shape, dtype, texHeight, texWidth, channels) {
// Temporarily create an tensor info to make the texture compatible with
// the runWebGLProgram's input.
const input = this.makeTensorInfo(shape, dtype);
const inData = this.texData.get(input.dataId);
// Even though the input texture could be unpacked or dense packed, it is
// always considered as unpacked for EncodeMatrixProgram.
inData.isPacked = false;
// Bind texture to the input tensor.
inData.texture = { texture, texShape: [texHeight, texWidth] };
inData.texShape = [texHeight, texWidth];
const shapeAs3D = getShapeAs3D(shape);
const program = new EncodeMatrixProgram(shapeAs3D, false /* isByteArray */, channels);
const output = this.runWebGLProgram(program, [input], dtype, [[texHeight, texWidth]]);
output.shape = shape;
// Unbind the texture from the input tensor to avoid the texture being
// released.
inData.texture = null;
this.disposeIntermediateTensorInfo(input);
return output.dataId;
}
write(values, shape, dtype) {
if (env().getBool('WEBGL_CHECK_NUMERICAL_PROBLEMS') ||
env().getBool('DEBUG')) {
this.checkNumericalProblems(values);
}
if (dtype === 'complex64' && values != null) {
throw new Error(`Cannot write to a complex64 dtype. ` +
`Please use tf.complex(real, imag).`);
}
const dataId = { id: this.nextDataId() };
this.texData.set(dataId, { shape, dtype, values, usage: TextureUsage.UPLOAD, refCount: 1 });
return dataId;
}
/** Return refCount of a `TensorData`. */
refCount(dataId) {
if (this.texData.has(dataId)) {
const tensorData = this.texData.get(dataId);
return tensorData.refCount;
}
return 0;
}
/** Increase refCount of a `TextureData`. */
incRef(dataId) {
const texData = this.texData.get(dataId);
texData.refCount++;
}
/** Decrease refCount of a `TextureData`. */
decRef(dataId) {
if (this.texData.has(dataId)) {
const texData = this.texData.get(dataId);
texData.refCount--;
}
}
move(dataId, values, shape, dtype, refCount) {
if (env().getBool('DEBUG')) {
this.checkNumericalProblems(values);
}
if (dtype === 'complex64') {
throw new Error(`Cannot write to a complex64 dtype. ` +
`Please use tf.complex(real, imag).`);
}
this.texData.set(dataId, { shape, dtype, values, usage: TextureUsage.UPLOAD, refCount });
}
disposeIntermediateTensorInfo(tensorInfo) {
this.disposeData(tensorInfo.dataId);
}
readSync(dataId) {
const texData = this.texData.get(dataId);
const { values, dtype, complexTensorInfos, slice, shape, isPacked } = texData;
// The presence of `slice` indicates this tensor is a shallow slice of a
// different tensor, and is using that original tensor's texture. Run
// `clone` in order to copy that texture and read from it.
if (slice != null) {
let program;
if (isPacked) {
program = new UnaryOpPackedProgram(shape, CLONE);
}
else {
program = new UnaryOpProgram(shape, CLONE);
}
const res = this.runWebGLProgram(program, [{ dataId, shape, dtype }], dtype);
const data = this.readSync(res.dataId);
this.disposeIntermediateTensorInfo(res);
return data;
}
if (values != null) {
return this.convertAndCacheOnCPU(dataId);
}
if (dtype === 'string') {
return values;
}
const shouldTimeProgram = this.activeTimers != null;
let start;
if (shouldTimeProgram) {
start = now();
}
let result;
if (dtype === 'complex64') {
const realValues = this.readSync(complexTensorInfos.real.dataId);
const imagValues = this.readSync(complexTensorInfos.imag.dataId);
result = mergeRealAndImagArrays(realValues, imagValues);
}
else {
result = this.getValuesFromTexture(dataId);
}
if (shouldTimeProgram) {
this.downloadWaitMs += now() - start;
}
return this.convertAndCacheOnCPU(dataId, result);
}
async read(dataId) {
if (this.pendingRead.has(dataId)) {
const subscribers = this.pendingRead.get(dataId);
return new Promise(resolve => subscribers.push(resolve));
}
const texData = this.texData.get(dataId);
const { values, shape, slice, dtype, complexTensorInfos, isPacked } = texData;
// The presence of `slice` indicates this tensor is a shallow slice of a
// different tensor, and is using that original tensor's texture. Run
// `clone` in order to copy that texture and read from it.
if (slice != null) {
let program;
if (isPacked) {
program = new UnaryOpPackedProgram(shape, CLONE);
}
else {
program = new UnaryOpProgram(shape, CLONE);
}
const res = this.runWebGLProgram(program, [{ dataId, shape, dtype }], dtype);
const data = this.read(res.dataId);
this.disposeIntermediateTensorInfo(res);
return data;
}
if (values != null) {
return this.convertAndCacheOnCPU(dataId);
}
if (env().getBool('DEBUG')) {
// getBool('WEBGL_DOWNLOAD_FLOAT_ENABLED') caused a blocking GPU call.
// For performance reason, only check it for debugging. In production,
// it doesn't handle this use case anyway, so behavior is not changed.
if (!env().getBool('WEBGL_DOWNLOAD_FLOAT_ENABLED') &&
env().getNumber('WEBGL_VERSION') === 2) {
throw new Error(`tensor.data() with WEBGL_DOWNLOAD_FLOAT_ENABLED=false and ` +
`WEBGL_VERSION=2 not yet supported.`);
}
}
let buffer = null;
let tmpDownloadTarget;
if (dtype !== 'complex64' && env().get('WEBGL_BUFFER_SUPPORTED')) {
// Possibly copy the texture into a buffer before inserting a fence.
tmpDownloadTarget = this.decode(dataId);
const tmpData = this.texData.get(tmpDownloadTarget.dataId);
buffer = this.gpgpu.createBufferFromTexture(tmpData.texture.texture, ...getDenseTexShape(shape));
}
this.pendingRead.set(dataId, []);
if (dtype !== 'complex64') {
// Create a fence and wait for it to resolve.
await this.gpgpu.createAndWaitForFence();
}
// Download the values from the GPU.
let vals;
if (dtype === 'complex64') {
const ps = await Promise.all([
this.read(complexTensorInfos.real.dataId),
this.read(complexTensorInfos.imag.dataId)
]);
const realValues = ps[0];
const imagValues = ps[1];
vals = mergeRealAndImagArrays(realValues, imagValues);
}
else if (buffer == null) {
vals = this.getValuesFromTexture(dataId);
}
else {
const size = sizeFromShape(shape);
vals = this.gpgpu.downloadFloat32MatrixFromBuffer(buffer, size);
}
if (tmpDownloadTarget != null) {
this.disposeIntermediateTensorInfo(tmpDownloadTarget);
}
if (buffer != null) {
const gl = this.gpgpu.gl;
callAndCheck(gl, () => gl.deleteBuffer(buffer));
}
const dTypeVals = this.convertAndCacheOnCPU(dataId, vals);
const subscribers = this.pendingRead.get(dataId);
this.pendingRead.delete(dataId);
// Notify all pending reads.
subscribers.forEach(resolve => resolve(dTypeVals));
if (this.pendingDisposal.has(dataId)) {
this.pendingDisposal.delete(dataId);
if (this.disposeData(dataId)) {
engine().removeDataId(dataId, this);
}
this.pendingDeletes--;
}
return dTypeVals;
}
/**
* Read tensor to a new texture that is densely packed for ease of use.
* @param dataId The source tensor.
* @param options
* customTexShape: Optional. If set, will use the user defined texture
* shape to create the texture.
*/
readToGPU(dataId, options = {}) {
const texData = this.texData.get(dataId);
const { values, shape, slice, dtype, isPacked, texture } = texData;
if (dtype === 'complex64') {
throw new Error('Does not support reading texture for complex64 dtype.');
}
// The presence of `slice` indicates this tensor is a shallow slice of a
// different tensor, and is using that original tensor's texture. Run
// `clone` in order to copy that texture and read from it.
if (slice != null) {
let program;
if (isPacked) {
program = new UnaryOpPackedProgram(shape, CLONE);
}
else {
program = new UnaryOpProgram(shape, CLONE);
}
const res = this.runWebGLProgram(program, [{ dataId, shape, dtype }], dtype);
const gpuResouorce = this.readToGPU(res, options);
this.disposeIntermediateTensorInfo(res);
return gpuResouorce;
}
if (texture == null) {
if (values != null) {
throw new Error('Data is not on GPU but on CPU.');
}
else {
throw new Error('There is no data on GPU or CPU.');
}
}
// Decode the texture so that it is stored densely (using four channels).
const tmpTarget = this.decode(dataId, options.customTexShape);
// Make engine track this tensor, so that we can dispose it later.
const tensorRef = engine().makeTensorFromTensorInfo(tmpTarget);
const tmpData = this.texData.get(tmpTarget.dataId);
return Object.assign({ tensorRef }, tmpData.texture);
}
bufferSync(t) {
const data = this.readSync(t.dataId);
if (t.dtype === 'string') {
try {
// Decode the bytes into string.
const strings = data.map(d => decodeString(d));
return buffer(t.shape, t.dtype, strings);
}
catch (_a) {
throw new Error('Failed to decode encoded string bytes into utf-8');
}
}
return buffer(t.shape, t.dtype, data);
}
checkNumericalProblems(values) {
if (values == null) {
return;
}
for (let i = 0; i < values.length; i++) {
const num = values[i];
if (!canBeRepresented(num)) {
if (env().getBool('WEBGL_RENDER_FLOAT32_CAPABLE')) {
throw Error(`The value ${num} cannot be represented with your ` +
`current settings. Consider enabling float32 rendering: ` +
`'tf.env().set('WEBGL_RENDER_FLOAT32_ENABLED', true);'`);
}
throw Error(`The value ${num} cannot be represented on this device.`);
}
}
}
getValuesFromTexture(dataId) {
const { shape, dtype, isPacked } = this.texData.get(dataId);
const size = sizeFromShape(shape);
if (env().getBool('WEBGL_DOWNLOAD_FLOAT_ENABLED')) {
const tmpTarget = this.decode(dataId);
const tmpData = this.texData.get(tmpTarget.dataId);
const vals = this.gpgpu
.downloadMatrixFromPackedTexture(tmpData.texture.texture, ...getDenseTexShape(shape))
.subarray(0, size);
this.disposeIntermediateTensorInfo(tmpTarget);
return vals;
}
const shouldUsePackedProgram = env().getBool('WEBGL_PACK') && isPacked === true;
const outputShape = shouldUsePackedProgram ? getShapeAs3D(shape) : shape;
const program = shouldUsePackedProgram ?
new EncodeFloatPackedProgram(outputShape) :
new EncodeFloatProgram(outputShape);
const output = this.runWebGLProgram(program, [{ shape: outputShape, dtype, dataId }], 'float32');
const tmpData = this.texData.get(output.dataId);
const vals = this.gpgpu
.downloadByteEncodedFloatMatrixFromOutputTexture(tmpData.texture.texture, tmpData.texShape[0], tmpData.texShape[1])
.subarray(0, size);
this.disposeIntermediateTensorInfo(output);
return vals;
}
timerAvailable() {
return env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE') > 0;
}
time(f) {
const oldActiveTimers = this.activeTimers;
const newActiveTimers = [];
let outerMostTime = false;
if (this.programTimersStack == null) {
this.programTimersStack = newActiveTimers;
outerMostTime = true;
}
else {
this.activeTimers.push(newActiveTimers);
}
this.activeTimers = newActiveTimers;
f();
// needing to split these up because util.flatten only accepts certain types
const flattenedActiveTimerQueries = flatten$1(this.activeTimers.map((d) => d.query))
.filter(d => d != null);
const flattenedActiveTimerNames = flatten$1(this.activeTimers.map((d) => d.name))
.filter(d => d != null);
this.activeTimers = oldActiveTimers;
if (outerMostTime) {
this.programTimersStack = null;
}
const res = {
uploadWaitMs: this.uploadWaitMs,
downloadWaitMs: this.downloadWaitMs,
kernelMs: null,
wallMs: null // will be filled by the engine
};
return (async () => {
if (env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE') >
0) {
const kernelMs = await Promise.all(flattenedActiveTimerQueries);
res['kernelMs'] = sum$3(kernelMs);
res['getExtraProfileInfo'] = () => kernelMs
.map((d, i) => ({ name: flattenedActiveTimerNames[i], ms: d }))
.map(d => `${d.name}: ${d.ms}`)
.join(', ');
}
else {
res['kernelMs'] = {
error: 'WebGL query timers are not supported in this environment.'
};
}
this.uploadWaitMs = 0;
this.downloadWaitMs = 0;
return res;
})();
}
memory() {
return {
unreliable: false,
numBytesInGPU: this.numBytesInGPU,
numBytesInGPUAllocated: this.textureManager.numBytesAllocated,
numBytesInGPUFree: this.textureManager.numBytesFree
};
}
startTimer() {
if (env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE') > 0) {
return this.gpgpu.beginQuery();
}
return { startMs: now(), endMs: null };
}
endTimer(query) {
if (env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE') > 0) {
this.gpgpu.endQuery();
return query;
}
query.endMs = now();
return query;
}
async getQueryTime(query) {
if (env().getNumber('WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE') > 0) {
return this.gpgpu.waitForQueryAndGetTime(query);
}
const timerQuery = query;
return timerQuery.endMs - timerQuery.startMs;
}
/**
* Decrease the RefCount on the dataId and dispose the memory if the dataId
* has 0 refCount. If there are pending read on the data, the disposal would
* added to the pending delete queue. Return true if the dataId is removed
* from backend or the backend does not contain the dataId, false if the
* dataId is not removed. Memory may or may not be released even when dataId
* is removed, which also depends on dataRefCount, see `releaseGPU`.
* @param dataId
* @oaram force Optional, remove the data regardless of refCount
*/
disposeData(dataId, force = false) {
if (this.pendingDisposal.has(dataId)) {
return false;
}
// No-op if already disposed.
if (!this.texData.has(dataId)) {
return true;
}
// if force flag is set, change refCount to 0, this would ensure disposal
// when added to the pendingDisposal queue. Memory may or may not be
// released, which also depends on dataRefCount, see `releaseGPU`.
if (force) {
this.texData.get(dataId).refCount = 0;
}
else {
this.texData.get(dataId).refCount--;
}
if (!force && this.texData.get(dataId).refCount > 0) {
return false;
}
if (this.pendingRead.has(dataId)) {
this.pendingDisposal.add(dataId);
this.pendingDeletes++;
return false;
}
this.releaseGPUData(dataId);
const { complexTensorInfos } = this.texData.get(dataId);
if (complexTensorInfos != null) {
this.disposeData(complexTensorInfos.real.dataId, force);
this.disposeData(complexTensorInfos.imag.dataId, force);
}
this.texData.delete(dataId);
return true;
}
releaseGPUData(dataId) {
const { texture, dtype, texShape, usage, isPacked, slice } = this.texData.get(dataId);
const key = slice && slice.origDataId || dataId;
const refCount = this.dataRefCount.get(key);
if (refCount > 1) {
this.dataRefCount.set(key, refCount - 1);
}
else {
this.dataRefCount.delete(key);
if (texture != null) {
this.numBytesInGPU -= this.computeBytes(texShape, dtype);
this.textureManager.releaseTexture(texture, texShape, usage, isPacked);
}
}
const texData = this.texData.get(dataId);
texData.texture = null;
texData.texShape = null;
texData.isPacked = false;
texData.slice = null;
}
getTexture(dataId) {
this.uploadToGPU(dataId);
return this.texData.get(dataId).texture.texture;
}
/**
* Returns internal information for the specific data bucket. Used in unit
* tests.
*/
getDataInfo(dataId) {
return this.texData.get(dataId);
}
/*
Tests whether all the inputs to an op are small and on the CPU. This heuristic
determines when it would be faster to execute a kernel on the CPU. WebGL
kernels opt into running this check and forwarding when appropriate.
TODO(https://github.com/tensorflow/tfjs/issues/872): Develop a more
sustainable strategy for optimizing backend execution of ops.
*/
shouldExecuteOnCPU(inputs, sizeThreshold = CPU_HANDOFF_SIZE_THRESHOLD) {
return env().getBool('WEBGL_CPU_FORWARD') &&
inputs.every(input => this.texData.get(input.dataId).texture == null &&
sizeFromShape(input.shape) < sizeThreshold);
}
getGPGPUContext() {
return this.gpgpu;
}
where(condition) {
warn('tf.where() in webgl locks the UI thread. ' +
'Call tf.whereAsync() instead');
const condVals = condition.dataSync();
return whereImpl$1(condition.shape, condVals);
}
packedUnaryOp(x, op, dtype) {
const program = new UnaryOpPackedProgram(x.shape, op);
const outInfo = this.compileAndRun(program, [x], dtype);
return engine().makeTensorFromTensorInfo(outInfo);
}
// TODO(msoulanille) remove this once the backend has been modularized
// a copy is needed here to break a circular dependency.
// Also remove the op from unary_op.
abs(x) {
// TODO: handle cases when x is complex.
if (this.shouldExecuteOnCPU([x]) && x.dtype !== 'complex64') {
const outValues = simpleAbsImplCPU(this.texData.get(x.dataId).values);
return this.makeOutput(x.shape, x.dtype, outValues);
}
if (env().getBool('WEBGL_PACK_UNARY_OPERATIONS')) {
return this.packedUnaryOp(x, ABS$1, x.dtype);
}
const program = new UnaryOpProgram(x.shape, ABS$1);
const outInfo = this.compileAndRun(program, [x]);
return engine().makeTensorFromTensorInfo(outInfo);
}
makeTensorInfo(shape, dtype, values) {
let dataId;
if (dtype === 'string' && values != null && values.length > 0 &&
isString(values[0])) {
const encodedValues = values.map(d => encodeString(d));
dataId = this.write(encodedValues, shape, dtype);
}
else {
dataId = this.write(values, shape, dtype);
}
this.texData.get(dataId).usage = null;
return { dataId, shape, dtype };
}
makeOutput(shape, dtype, values) {
return engine().makeTensorFromTensorInfo(this.makeTensorInfo(shape, dtype, values), this);
}
unpackTensor(input) {
const program = new UnpackProgram(input.shape);
return this.runWebGLProgram(program, [input], input.dtype);
}
packTensor(input) {
const program = new PackProgram(input.shape);
const preventEagerUnpackingOutput = true;
return this.runWebGLProgram(program, [input], input.dtype, null /* customUniformValues */, preventEagerUnpackingOutput);
}
packedReshape(input, afterShape) {
const input3DShape = [
getBatchDim(input.shape),
...getRowsCols(input.shape)
];
const input3D = {
dtype: input.dtype,
shape: input3DShape,
dataId: input.dataId
};
const afterShapeAs3D = [
getBatchDim(afterShape), ...getRowsCols(afterShape)
];
const program = new ReshapePackedProgram(afterShapeAs3D, input3DShape);
const preventEagerUnpackingOfOutput = true;
const customValues = [input3DShape];
const output = this.runWebGLProgram(program, [input3D], input.dtype, customValues, preventEagerUnpackingOfOutput);
return { dataId: output.dataId, shape: afterShape, dtype: output.dtype };
}
decode(dataId, customTexShape) {
const texData = this.texData.get(dataId);
const { isPacked, shape, dtype } = texData;
if (customTexShape != null) {
const size = sizeFromShape(shape);
const texSize = customTexShape[0] * customTexShape[1] * 4;
assert$1(size <= texSize, () => 'customTexShape is too small. ' +
'Row * Column * 4 should be equal or larger than the ' +
'size of the tensor data.');
}
const shapeAs3D = getShapeAs3D(shape);
let program;
if (isPacked) {
program = new DecodeMatrixPackedProgram(shapeAs3D);
}
else {
program = new DecodeMatrixProgram(shapeAs3D);
}
const preventEagerUnpackingOfOutput = true;
const customValues = [customTexShape != null ? customTexShape :
getDenseTexShape(shapeAs3D)];
const out = this.runWebGLProgram(program, [{ shape: shapeAs3D, dtype, dataId }], dtype, customValues, preventEagerUnpackingOfOutput, customTexShape);
return { dtype, shape, dataId: out.dataId };
}
runWebGLProgram(program, inputs, outputDtype, customUniformValues, preventEagerUnpackingOfOutput = false, customTexShape) {
const output = this.makeTensorInfo(program.outputShape, outputDtype);
const outData = this.texData.get(output.dataId);
if (program.packedOutput) {
outData.isPacked = true;
}
if (program.outPackingScheme === PackingScheme.DENSE) {
const texelShape = customTexShape != null ?
customTexShape :
getDenseTexShape(program.outputShape);
// For a densely packed output, we explicitly set texShape
// so it doesn't get assigned later according to our typical packing
// scheme wherein a single texel can only contain values from adjacent
// rows/cols.
outData.texShape = texelShape.map(d => d * 2);
}
if (program.outTexUsage != null) {
outData.usage = program.outTexUsage;
}
if (sizeFromShape(output.shape) === 0) {
// Short-circuit the computation since the result is empty (has 0 in its
// shape).
outData.values =
getTypedArrayFromDType(output.dtype, 0);
return output;
}
const dataToDispose = [];
const inputsData = inputs.map(input => {
if (input.dtype === 'complex64') {
throw new Error(`GPGPUProgram does not support complex64 input. For complex64 ` +
`dtypes, please separate the program into real and imaginary ` +
`parts.`);
}
let texData = this.texData.get(input.dataId);
if (texData.texture == null) {
if (!program.packedInputs &&
sizeFromShape(input.shape) <=
env().getNumber('WEBGL_SIZE_UPLOAD_UNIFORM')) {
// Upload small tensors that live on the CPU as uniforms, not as
// textures. Do this only when the environment supports 32bit floats
// due to problems when comparing 16bit floats with 32bit floats.
// TODO(https://github.com/tensorflow/tfjs/issues/821): Make it
// possible for packed shaders to sample from uniforms.
return {
shape: input.shape,
texData: null,
isUniform: true,
uniformValues: texData.values
};
}
// This ensures that if a packed program's inputs have not yet been
// uploaded to the GPU, they get uploaded as packed right off the bat.
if (program.packedInputs) {
texData.isPacked = true;
texData.shape = input.shape;
}
}
this.uploadToGPU(input.dataId);
if (!!texData.isPacked !== !!program.packedInputs) {
input = texData.isPacked ? this.unpackTensor(input) :
this.packTensor(input);
dataToDispose.push(input);
texData = this.texData.get(input.dataId);
}
else if (texData.isPacked &&
!isReshapeFree(texData.shape, input.shape)) {
// This is a special case where a texture exists for a tensor
// but the shapes are incompatible (due to packing constraints) because
// the tensor did not have a chance to go through the packed reshape
// shader. This only happens when we reshape the *same* tensor to form
// *distinct* inputs to an op, e.g. dotting a vector with itself. This
// case will disappear once packed uploading is the default.
const savedInput = input;
const targetShape = input.shape;
input.shape = texData.shape;
input = this.packedReshape(input, targetShape);
dataToDispose.push(input);
texData = this.texData.get(input.dataId);
savedInput.shape = targetShape;
}
return { shape: input.shape, texData, isUniform: false };
});
this.uploadToGPU(output.dataId);
const outputData = { shape: output.shape, texData: outData, isUniform: false };
const key = makeShaderKey(program, inputsData, outputData);
const binary = this.getAndSaveBinary(key, () => {
return compileProgram(this.gpgpu, program, inputsData, outputData);
});
const shouldTimeProgram = this.activeTimers != null;
let query;
if (shouldTimeProgram) {
query = this.startTimer();
}
if (!env().get('ENGINE_COMPILE_ONLY')) {
runProgram(this.gpgpu, binary, inputsData, outputData, customUniformValues);
}
dataToDispose.forEach(info => this.disposeIntermediateTensorInfo(info));
if (shouldTimeProgram) {
query = this.endTimer(query);
this.activeTimers.push({ name: program.constructor.name, query: this.getQueryTime(query) });
}
const glFlushThreshold = env().getNumber('WEBGL_FLUSH_THRESHOLD');
// Manually GL flush requested
if (glFlushThreshold > 0) {
const time = now();
if ((time - this.lastGlFlushTime) > glFlushThreshold) {
this.gpgpu.gl.flush();
this.lastGlFlushTime = time;
}
}
if (!env().getBool('WEBGL_LAZILY_UNPACK') && outData.isPacked &&
preventEagerUnpackingOfOutput === false) {
const unpacked = this.unpackTensor(output);
this.disposeIntermediateTensorInfo(output);
return unpacked;
}
return output;
}
compileAndRun(program, inputs, outputDtype, customUniformValues, preventEagerUnpackingOfOutput = false) {
outputDtype = outputDtype || inputs[0].dtype;
const outInfo = this.runWebGLProgram(program, inputs, outputDtype, customUniformValues, preventEagerUnpackingOfOutput);
return outInfo;
}
getAndSaveBinary(key, getBinary) {
if (!(key in this.binaryCache)) {
this.binaryCache[key] = getBinary();
}
return this.binaryCache[key];
}
getTextureManager() {
return this.textureManager;
}
dispose() {
if (this.disposed) {
return;
}
// Avoid disposing the compiled webgl programs during unit testing because
// it slows down test execution.
if (!env().getBool('IS_TEST')) {
const allKeys = Object.keys(this.binaryCache);
allKeys.forEach(key => {
this.gpgpu.deleteProgram(this.binaryCache[key].webGLProgram);
delete this.binaryCache[key];
});
}
this.textureManager.dispose();
if (this.canvas != null &&
(typeof (HTMLCanvasElement) !== 'undefined' &&
this.canvas instanceof HTMLCanvasElement)) {
this.canvas.remove();
}
else {
this.canvas = null;
}
if (this.gpgpuCreatedLocally) {
this.gpgpu.program = null;
this.gpgpu.dispose();
}
this.disposed = true;
}
floatPrecision() {
if (this.floatPrecisionValue == null) {
this.floatPrecisionValue = tidy(() => {
if (!env().get('WEBGL_RENDER_FLOAT32_ENABLED')) {
// Momentarily switching DEBUG flag to false so we don't throw an
// error trying to upload a small value.
const debugFlag = env().getBool('DEBUG');
env().set('DEBUG', false);
const underflowCheckValue = this.abs(scalar(1e-8)).dataSync()[0];
env().set('DEBUG', debugFlag);
if (underflowCheckValue > 0) {
return 32;
}
}
return 16;
});
}
return this.floatPrecisionValue;
}
/** Returns the smallest representable number. */
epsilon() {
return this.floatPrecision() === 32 ? EPSILON_FLOAT32 : EPSILON_FLOAT16;
}
uploadToGPU(dataId) {
const texData = this.texData.get(dataId);
const { shape, dtype, values, texture, usage, isPacked } = texData;
if (texture != null) {
// Array is already on GPU. No-op.
return;
}
const shouldTimeProgram = this.activeTimers != null;
let start;
if (shouldTimeProgram) {
start = now();
}
let texShape = texData.texShape;
if (texShape == null) {
// This texShape may not be the final texture shape. For packed or dense
// textures, the texShape will be changed when textures are created.
texShape = getTextureShapeFromLogicalShape(shape, isPacked);
texData.texShape = texShape;
}
if (values != null) {
const shapeAs3D = getShapeAs3D(shape);
let program;
let width = texShape[1], height = texShape[0];
const isByteArray = values instanceof Uint8Array || values instanceof Uint8ClampedArray;
// texture for float array is PhysicalTextureType.PACKED_2X2_FLOAT32, we
// need to make sure the upload uses the same packed size
if (isPacked || !isByteArray) {
[width, height] = getPackedMatrixTextureShapeWidthHeight(texShape[0], texShape[1]);
}
if (isPacked) {
program = new EncodeMatrixPackedProgram(shapeAs3D, isByteArray);
}
else {
program = new EncodeMatrixProgram(shapeAs3D, isByteArray);
}
// TexShape for float array needs to be the original shape, which byte
// array needs to be packed size. This allow the data upload shape to be
// matched with texture creation logic.
const tempDenseInputTexShape = isByteArray ? [height, width] : texShape;
const tempDenseInputHandle = this.makeTensorInfo(tempDenseInputTexShape, dtype);
const tempDenseInputTexData = this.texData.get(tempDenseInputHandle.dataId);
if (isByteArray) {
tempDenseInputTexData.usage = TextureUsage.PIXELS;
}
else {
tempDenseInputTexData.usage = TextureUsage.UPLOAD;
}
tempDenseInputTexData.texShape = tempDenseInputTexShape;
this.gpgpu.uploadDenseMatrixToTexture(this.getTexture(tempDenseInputHandle.dataId), width, height, values);
const customValues = [[height, width]];
// We want the output to remain packed regardless of the value of
// WEBGL_PACK.
const preventEagerUnpacking = true;
const encodedOutputTarget = this.runWebGLProgram(program, [tempDenseInputHandle], dtype, customValues, preventEagerUnpacking);
// Have the original texture assume the identity of the encoded output.
const outputTexData = this.texData.get(encodedOutputTarget.dataId);
texData.texShape = outputTexData.texShape;
texData.isPacked = outputTexData.isPacked;
texData.usage = outputTexData.usage;
if (!env().get('ENGINE_COMPILE_ONLY')) {
texData.texture = outputTexData.texture;
// Once uploaded, don't store the values on cpu.
texData.values = null;
this.texData.delete(encodedOutputTarget.dataId);
}
else {
this.disposeData(encodedOutputTarget.dataId);
}
this.disposeIntermediateTensorInfo(tempDenseInputHandle);
if (shouldTimeProgram) {
this.uploadWaitMs += now() - start;
}
}
else {
const newTexture = this.acquireTexture(texShape, usage, dtype, isPacked);
texData.texture = newTexture;
}
}
convertAndCacheOnCPU(dataId, float32Values) {
const texData = this.texData.get(dataId);
const { dtype } = texData;
if (float32Values != null) {
texData.values = float32ToTypedArray(float32Values, dtype);
}
return texData.values;
}
acquireTexture(texShape, texType, dtype, isPacked) {
this.numBytesInGPU += this.computeBytes(texShape, dtype);
if (!this.warnedAboutMemory &&
this.numBytesInGPU > this.numMBBeforeWarning * 1024 * 1024) {
const mb = (this.numBytesInGPU / 1024 / 1024).toFixed(2);
this.warnedAboutMemory = true;
console.warn(`High memory usage in GPU: ${mb} MB, ` +
`most likely due to a memory leak`);
}
return this.textureManager.acquireTexture(texShape, texType, isPacked);
}
computeBytes(shape, dtype) {
return shape[0] * shape[1] * bytesPerElement(dtype);
}
checkCompileCompletion() {
for (const [, binary] of Object.entries(this.binaryCache)) {
this.checkCompletion_(binary);
}
}
async checkCompileCompletionAsync() {
const ps = [];
if (this.gpgpu.parallelCompilationExtension) {
for (const [, binary] of Object.entries(this.binaryCache)) {
ps.push(this.checkCompletionAsync_(binary));
}
return Promise.all(ps);
}
else {
for (const [, binary] of Object.entries(this.binaryCache)) {
const p = new Promise((resolve) => {
try {
this.checkCompletion_(binary);
resolve(true);
}
catch (error) {
throw error;
}
});
ps.push(p);
}
return Promise.all(ps);
}
}
async checkCompletionAsync_(binary) {
if (this.gpgpu.gl.getProgramParameter(binary.webGLProgram, this.gpgpu.parallelCompilationExtension.COMPLETION_STATUS_KHR)) {
return this.checkCompletion_(binary);
}
else {
await nextFrame();
return this.checkCompletionAsync_(binary);
}
}
checkCompletion_(binary) {
if (this.gpgpu.gl.getProgramParameter(binary.webGLProgram, this.gpgpu.gl.LINK_STATUS) === false) {
console.log(this.gpgpu.gl.getProgramInfoLog(binary.webGLProgram));
if (this.gpgpu.gl.getShaderParameter(binary.fragmentShader, this.gpgpu.gl.COMPILE_STATUS) === false) {
logShaderSourceAndInfoLog(binary.source, this.gpgpu.gl.getShaderInfoLog(binary.fragmentShader));
throw new Error('Failed to compile fragment shader.');
}
throw new Error('Failed to link vertex and fragment shaders.');
}
return true;
}
getUniformLocations() {
for (const binary of Object.values(this.binaryCache)) {
// TODO: Iterating through all binaries to build VAOs is supposed to be in
// a seperate function, like 'setVaos'. However, to avoid breaking changes
// for the users using parallel compile feature now, buildVao is silently
// added here.
this.gpgpu.buildVao(binary.webGLProgram);
const { variablesLocations, customUniformLocations, infLoc, nanLoc, outShapeLocation, outShapeStridesLocation, outTexShapeLocation } = getUniformLocations(this.gpgpu, binary.program, binary.webGLProgram);
binary.variablesLocations = variablesLocations;
binary.customUniformLocations = customUniformLocations;
binary.infLoc = infLoc;
binary.nanLoc = nanLoc;
binary.outShapeLocation = outShapeLocation;
binary.outShapeStridesLocation = outShapeStridesLocation;
binary.outTexShapeLocation = outTexShapeLocation;
}
}
/**
* Create a TF.js tensor out of an existing WebGL texture. A new texture will
* be created.
*/
createTensorFromGPUData(values, shape, dtype) {
values.channels = values.channels || 'RGBA';
const { texture, height, width, channels } = values;
const backend = engine().backend;
// Have to throw an error, otherwise WebGL just warns and returns wrong
// values.
if (!backend.gpgpu.gl.isTexture(texture)) {
throw new Error(`The texture is invalid. Also, please make sure the texture and ` +
`the TFJS WebGL backend are using the same canvas. If you want to ` +
`use your own custom canvas, you have to create and use the custom ` +
`TFJS WebGL backend created from the canvas through ` +
`'new tf.MathBackendWebGL(customCanvas)'.`);
}
const dataId = backend.writeTexture(texture, shape, dtype, height, width, channels);
return engine().makeTensorFromDataId(dataId, shape, dtype, backend);
}
}
MathBackendWebGL.nextDataId = 0;
function float32ToTypedArray(a, dtype) {
if (dtype === 'float32' || dtype === 'complex64') {
return a;
}
else if (dtype === 'int32' || dtype === 'bool') {
const result = (dtype === 'int32') ? new Int32Array(a.length) :
new Uint8Array(a.length);
for (let i = 0; i < result.length; ++i) {
result[i] = Math.round(a[i]);
}
return result;
}
else {
throw new Error(`Unknown dtype ${dtype}`);
}
}
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// base.ts is the webgl backend without auto kernel registration.
if (isBrowser()) {
registerBackend('webgl', () => new MathBackendWebGL(), 2 /* priority */);
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const CHECK_NAN_SNIPPET = `
if (isnan(a)) return a;
if (isnan(b)) return b;
`;
class BinaryOpProgram {
constructor(op, aShape, bShape) {
this.variableNames = ['A', 'B'];
this.outputShape = assertAndGetBroadcastShape(aShape, bShape);
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
this.userCode = `
float binaryOperation(float a, float b) {
${op}
}
void main() {
float a = getAAtOutCoords();
float b = getBAtOutCoords();
setOutput(binaryOperation(a, b));
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const CHECK_NAN_SNIPPET_PACKED = `
result.r = isNaN.r ? NAN : result.r;
result.g = isNaN.g ? NAN : result.g;
result.b = isNaN.b ? NAN : result.b;
result.a = isNaN.a ? NAN : result.a;
`;
class BinaryOpPackedProgram {
constructor(op, aShape, bShape, checkOutOfBounds = false) {
this.variableNames = ['A', 'B'];
this.supportsBroadcasting = true;
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = assertAndGetBroadcastShape(aShape, bShape);
const rank = this.outputShape.length;
this.enableShapeUniforms = useShapeUniforms(rank);
let checkOutOfBoundsString = '';
if (checkOutOfBounds) {
if (rank === 0 || sizeFromShape(this.outputShape) === 1) {
checkOutOfBoundsString = `
result.y = 0.;
result.z = 0.;
result.w = 0.;
`;
}
else {
const dtype = getCoordsDataType(rank);
checkOutOfBoundsString = `
${dtype} coords = getOutputCoords();
`;
if (rank === 1) {
if (this.enableShapeUniforms) {
checkOutOfBoundsString += `
result.y = (coords + 1) >= outShape ? 0. : result.y;
result.z = 0.;
result.w = 0.;
`;
}
else {
checkOutOfBoundsString += `
result.y = (coords + 1) >= ${this.outputShape[0]} ? 0. : result.y;
result.z = 0.;
result.w = 0.;
`;
}
}
else {
const channels = getChannels('coords', rank);
if (this.enableShapeUniforms) {
checkOutOfBoundsString += `
bool nextRowOutOfBounds =
(${channels[rank - 2]} + 1) >= outShape[${rank} - 2];
bool nextColOutOfBounds =
(${channels[rank - 1]} + 1) >= outShape[${rank} - 1];
result.y = nextColOutOfBounds ? 0. : result.y;
result.z = nextRowOutOfBounds ? 0. : result.z;
result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
`;
}
else {
checkOutOfBoundsString += `
bool nextRowOutOfBounds =
(${channels[rank - 2]} + 1) >= ${this.outputShape[rank - 2]};
bool nextColOutOfBounds =
(${channels[rank - 1]} + 1) >= ${this.outputShape[rank - 1]};
result.y = nextColOutOfBounds ? 0. : result.y;
result.z = nextRowOutOfBounds ? 0. : result.z;
result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
`;
}
}
}
}
this.userCode = `
vec4 binaryOperation(vec4 a, vec4 b) {
${op}
}
void main() {
vec4 a = getAAtOutCoords();
vec4 b = getBAtOutCoords();
vec4 result = binaryOperation(a, b);
${checkOutOfBoundsString}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function identity(args) {
const { inputs, backend } = args;
const { x } = inputs;
backend.incRef(x.dataId);
return { dataId: x.dataId, shape: x.shape, dtype: x.dtype };
}
const identityConfig = {
kernelName: Identity$1,
backendName: 'webgl',
kernelFunc: identity
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* In WebGL data is stored in GPU textures which can't be efficiently copied, so
* complex tensors share data with their real and imaginary components. Complex
* tensors' reference to the components is tracked by refCount on the individual
* component. The refCounts are increased by the identity call.
*
* When a complex tensor is disposed, it will reduce the refCount on the
* components by calling disposeData on each.
*/
function complex(args) {
const { inputs, backend } = args;
const { real, imag } = inputs;
const complexInfo = backend.makeTensorInfo(real.shape, 'complex64');
const complex = backend.texData.get(complexInfo.dataId);
const realTensorInfo = identity({ inputs: { x: real }, backend });
const imagTensorInfo = identity({ inputs: { x: imag }, backend });
complex.complexTensorInfos = { real: realTensorInfo, imag: imagTensorInfo };
return complexInfo;
}
const complexConfig = {
kernelName: Complex,
backendName: 'webgl',
kernelFunc: complex
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LEAKYRELU = `return (a < 0.) ? b * a : a;`;
const LEAKYRELU_PACKED = `
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
`;
function leakyRelu$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { alpha } = attrs;
const $alpha = backend.makeTensorInfo([], 'float32', createScalarValue(alpha, 'float32'));
const program = env().getBool('WEBGL_PACK_BINARY_OPERATIONS') ?
new BinaryOpPackedProgram(LEAKYRELU_PACKED, x.shape, $alpha.shape) :
new BinaryOpProgram(LEAKYRELU, x.shape, $alpha.shape);
const result = backend.runWebGLProgram(program, [x, $alpha], 'float32');
backend.disposeIntermediateTensorInfo($alpha);
return result;
}
const leakyReluConfig$1 = {
kernelName: LeakyRelu,
backendName: 'webgl',
kernelFunc: leakyRelu$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const PRELU = `return (a < 0.) ? b * a : a;`;
const PRELU_PACKED = `
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
`;
function prelu$1(args) {
const { inputs, backend } = args;
const { x, alpha } = inputs;
const program = env().getBool('WEBGL_PACK_BINARY_OPERATIONS') ?
new BinaryOpPackedProgram(PRELU_PACKED, x.shape, alpha.shape) :
new BinaryOpProgram(PRELU, x.shape, alpha.shape);
return backend.runWebGLProgram(program, [x, alpha], 'float32');
}
const preluConfig$1 = {
kernelName: Prelu,
backendName: 'webgl',
kernelFunc: prelu$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const CHECK_NAN_SNIPPET_UNARY = `if (isnan(x)) return x;`;
/**
* Template that creates a `KernelFunc` for unary ops.
* @param opSnippet Op snippet to create `UnaryOpProgram`.
* @param packedOpSnippet Op snippet to create `UnaryOpPackedProgram`.
* @param dtype Optional. If set, the result has this dtype. Otherwise, the
* result has the same dtype as the first input. This is mainly used in
* comparison kernels, such as Equal, Less, Greater, etc.
*/
function unaryKernelFunc({ opSnippet, packedOpSnippet, cpuKernelImpl, dtype }) {
return ({ inputs, backend }) => {
const { x } = inputs;
const webglBackend = backend;
const $dtype = dtype || x.dtype;
if (webglBackend.shouldExecuteOnCPU([x]) && cpuKernelImpl != null) {
const xData = webglBackend.texData.get(x.dataId);
const outValues = cpuKernelImpl(xData.values, $dtype);
return webglBackend.makeTensorInfo(x.shape, $dtype, outValues);
}
const shouldUsePackedProgram = env().getBool('WEBGL_PACK_UNARY_OPERATIONS') && packedOpSnippet != null;
let program;
if (shouldUsePackedProgram) {
program = new UnaryOpPackedProgram(x.shape, packedOpSnippet);
}
else {
program = new UnaryOpProgram(x.shape, opSnippet);
}
return webglBackend.runWebGLProgram(program, [x], $dtype);
};
}
/**
* Template that creates a `KernelFunc` for binary ops.
* @param opSnippet Op snippet to create `BinaryOpProgram`.
* @param packedOpSnippet Op snippet to create `BinaryOpPackedProgram`.
* @param checkOutOfBoundsForPackedProgram Whether to set checkOutOfBounds=true
* when creating BinaryOpPackedProgram.
* @param dtype Optional. If set, the result has this dtype. Otherwise, the
* result has the same dtype as the first input. This is mainly used in
* comparison kernels, such as Equal, Less, Greater, etc.
*/
function binaryKernelFunc({ opSnippet, packedOpSnippet, checkOutOfBounds = false, supportsComplex = false, cpuKernelImpl, dtype }) {
return ({ inputs, backend }) => {
const { a, b } = inputs;
const webglBackend = backend;
if (supportsComplex && a.dtype === 'complex64') {
const aData = webglBackend.texData.get(a.dataId);
const bData = webglBackend.texData.get(b.dataId);
const [real, imag] = [
[aData.complexTensorInfos.real, bData.complexTensorInfos.real],
[aData.complexTensorInfos.imag, bData.complexTensorInfos.imag]
].map(complexParts => {
const [aPart, bPart] = complexParts;
const aHandle = {
dataId: aPart.dataId,
dtype: aPart.dtype,
shape: a.shape
};
const bHandle = {
dataId: bPart.dataId,
dtype: bPart.dtype,
shape: b.shape
};
const program = new BinaryOpProgram(opSnippet, a.shape, b.shape);
return webglBackend.runWebGLProgram(program, [aHandle, bHandle], upcastType(aPart.dtype, bPart.dtype));
});
const complexOutput = complex({ inputs: { real, imag }, backend: webglBackend });
webglBackend.disposeIntermediateTensorInfo(real);
webglBackend.disposeIntermediateTensorInfo(imag);
// TODO(annxingyuan): Implement CPU forwarding for complex inputs.
return complexOutput;
}
const $dtype = dtype || upcastType(a.dtype, b.dtype);
if ((a.dtype === 'string' || b.dtype === 'string' ||
webglBackend.shouldExecuteOnCPU([a, b])) &&
cpuKernelImpl != null) {
const aVals = webglBackend.texData.get(a.dataId).values;
const bVals = webglBackend.texData.get(b.dataId).values;
const decodedAVals = a.dtype === 'string' ?
// tslint:disable-next-line: no-any
fromUint8ToStringArray(aVals) :
aVals;
const decodedBVals = a.dtype === 'string' ?
// tslint:disable-next-line: no-any
fromUint8ToStringArray(bVals) :
bVals;
const [outValues, outShape] = cpuKernelImpl(a.shape, b.shape, decodedAVals, decodedBVals, $dtype);
const out = webglBackend.makeTensorInfo(outShape, $dtype);
const outData = webglBackend.texData.get(out.dataId);
outData.values = outValues;
return out;
}
const shouldUsePackedProgram = env().getBool('WEBGL_PACK_BINARY_OPERATIONS') &&
packedOpSnippet != null;
let program;
if (shouldUsePackedProgram) {
program = new BinaryOpPackedProgram(packedOpSnippet, a.shape, b.shape, checkOutOfBounds);
}
else {
program = new BinaryOpProgram(opSnippet, a.shape, b.shape);
}
return webglBackend.runWebGLProgram(program, [a, b], $dtype);
};
}
function mapActivationToShaderProgram(activation, packed = false) {
if (activation === 'linear') {
if (packed) {
return LINEAR;
}
return LINEAR$1;
}
else if (activation === 'relu') {
if (packed) {
return RELU$1;
}
return RELU$2;
}
else if (activation === 'elu') {
if (packed) {
return ELU$1;
}
return ELU$2;
}
else if (activation === 'relu6') {
if (packed) {
return RELU6$1;
}
return RELU6$2;
}
else if (activation === 'prelu') {
if (packed) {
return PRELU_PACKED;
}
return PRELU;
}
else if (activation === 'leakyrelu') {
if (packed) {
return LEAKYRELU_PACKED;
}
return LEAKYRELU;
}
else if (activation === 'sigmoid') {
if (packed) {
return SIGMOID$1;
}
return SIGMOID$2;
}
throw new Error(`Activation ${activation} has not been implemented for the WebGL backend.`);
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class MatMulPackedProgram {
constructor(aShape, bShape, outputShape, transposeA = false, transposeB = false, addBias = false, activation = null, hasPreluActivation = false, hasLeakyreluActivation = false) {
this.variableNames = ['matrixA', 'matrixB'];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
const sharedDim = transposeA ? aShape[1] : aShape[2];
const sharedDimensionPacked = Math.ceil(sharedDim / 2);
const aSample = transposeA ? 'i * 2, rc.y' : 'rc.y, i * 2';
const bSample = transposeB ? 'rc.z, i * 2' : 'i * 2, rc.z';
const aSwizzle = transposeA ? ['a.xxyy', 'a.zzww'] : ['a.xxzz', 'a.yyww'];
const bSwizzle = transposeB ? ['b.xzxz', 'b.ywyw'] : ['b.xyxy', 'b.zwzw'];
let activationSnippet = '', applyActivationSnippet = '';
if (activation) {
if (hasPreluActivation) {
activationSnippet = `vec4 activation(vec4 a) {
vec4 b = getPreluActivationWeightsAtOutCoords();
${activation}
}`;
}
else if (hasLeakyreluActivation) {
activationSnippet = `vec4 activation(vec4 a) {
vec4 b = getLeakyreluAlphaAtOutCoords();
${activation}
}`;
}
else {
activationSnippet = `vec4 activation(vec4 x) {
${activation}
}`;
}
applyActivationSnippet = `result = activation(result);`;
}
const addBiasSnippet = addBias ? 'result += getBiasAtOutCoords();' : '';
if (addBias) {
this.variableNames.push('bias');
}
if (hasPreluActivation) {
this.variableNames.push('preluActivationWeights');
}
if (hasLeakyreluActivation) {
this.variableNames.push('leakyreluAlpha');
}
let batchASnippet = 'rc.x';
let batchBSnippet = 'rc.x';
if (aShape[0] < bShape[0]) {
batchASnippet = `imod(rc.x, ${aShape[0]})`;
}
else if (bShape[0] < aShape[0]) {
batchBSnippet = `imod(rc.x, ${bShape[0]})`;
}
this.userCode = `
${activationSnippet}
// Don't use uniform for sharedDimensionPacked for performance.
const float sharedDimension = ${sharedDimensionPacked}.0;
vec4 dot2x2ARowBCol(ivec3 rc) {
vec4 result = vec4(0);
int batchA = ${batchASnippet};
int batchB = ${batchBSnippet};
for (int i = 0; i < ${sharedDimensionPacked}; i++) {
vec4 a = getMatrixA(batchA, ${aSample});
vec4 b = getMatrixB(batchB, ${bSample});
// These swizzled products need to be separately added.
// See: https://github.com/tensorflow/tfjs/issues/1735
result += (${aSwizzle[0]} * ${bSwizzle[0]});
result += (${aSwizzle[1]} * ${bSwizzle[1]});
}
return result;
}
void main() {
ivec3 rc = getOutputCoords();
vec4 result = dot2x2ARowBCol(rc);
${addBiasSnippet}
${applyActivationSnippet}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// (Ar + Ai)(Br + Bi) =
// ArBr + ArBi + AiBr + AiBi = ArBr - AB + ArBi + AiBr
// Yr = ArBr - AB
// Yi = ArBi + AiBr
const COMPLEX_MULTIPLY = {
REAL: 'return areal * breal - aimag * bimag;',
IMAG: 'return areal * bimag + aimag * breal;'
};
class BinaryOpComplexProgram {
constructor(op, aShape, bShape) {
this.variableNames = ['AReal', 'AImag', 'BReal', 'BImag'];
this.outputShape = assertAndGetBroadcastShape(aShape, bShape);
this.userCode = `
float binaryOpComplex(
float areal, float aimag, float breal, float bimag) {
${op}
}
void main() {
float areal = getARealAtOutCoords();
float aimag = getAImagAtOutCoords();
float breal = getBRealAtOutCoords();
float bimag = getBImagAtOutCoords();
setOutput(binaryOpComplex(areal, aimag, breal, bimag));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const MUL = 'return a * b;';
function multiply(args) {
const { inputs, backend } = args;
const { a, b } = inputs;
const dtype = upcastType(a.dtype, b.dtype);
if (a.dtype === 'complex64') {
const aData = backend.texData.get(a.dataId);
const bData = backend.texData.get(b.dataId);
const realProgram = new BinaryOpComplexProgram(COMPLEX_MULTIPLY.REAL, a.shape, b.shape);
const imagProgram = new BinaryOpComplexProgram(COMPLEX_MULTIPLY.IMAG, a.shape, b.shape);
const inputs = [
{
dataId: aData.complexTensorInfos.real.dataId,
dtype: aData.complexTensorInfos.real.dtype,
shape: a.shape
},
{
dataId: aData.complexTensorInfos.imag.dataId,
dtype: aData.complexTensorInfos.imag.dtype,
shape: a.shape
},
{
dataId: bData.complexTensorInfos.real.dataId,
dtype: bData.complexTensorInfos.real.dtype,
shape: b.shape
},
{
dataId: bData.complexTensorInfos.imag.dataId,
dtype: bData.complexTensorInfos.imag.dtype,
shape: b.shape
}
];
const realPart = backend.runWebGLProgram(realProgram, inputs, 'float32');
const imagPart = backend.runWebGLProgram(imagProgram, inputs, 'float32');
const complexOutput = complex({ inputs: { real: realPart, imag: imagPart }, backend });
backend.disposeIntermediateTensorInfo(realPart);
backend.disposeIntermediateTensorInfo(imagPart);
// TODO(annxingyuan): CPU forwarding for complex inputs.
return complexOutput;
}
if (backend.shouldExecuteOnCPU([a, b])) {
const aData = backend.texData.get(a.dataId);
const bData = backend.texData.get(b.dataId);
const [outValues, outShape] = multiplyImplCPU(a.shape, b.shape, aData.values, bData.values, dtype);
const out = backend.makeTensorInfo(outShape, dtype);
const outData = backend.texData.get(out.dataId);
outData.values = outValues;
return out;
}
let program;
if (env().getBool('WEBGL_PACK_BINARY_OPERATIONS')) {
program = new BinaryOpPackedProgram(MUL, a.shape, b.shape);
}
else {
program = new BinaryOpProgram(MUL, a.shape, b.shape);
}
return backend.runWebGLProgram(program, [a, b], dtype);
}
const multiplyConfig = {
kernelName: Multiply,
backendName: 'webgl',
kernelFunc: multiply
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function packedReshape(input, afterShape, backend) {
const input3DShape = [getBatchDim(input.shape),
...getRowsCols(input.shape)];
const input3D = {
dtype: input.dtype,
shape: input3DShape,
dataId: input.dataId
};
const afterShapeAs3D = [getBatchDim(afterShape),
...getRowsCols(afterShape)];
const program = new ReshapePackedProgram(afterShapeAs3D, input3DShape);
const preventEagerUnpackingOfOutput = true;
const customValues = [input3DShape];
const output = backend.runWebGLProgram(program, [input3D], input.dtype, customValues, preventEagerUnpackingOfOutput);
return { dataId: output.dataId, shape: afterShape, dtype: output.dtype };
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function reshape$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { shape } = attrs;
const webglBackend = backend;
const xSize = sizeFromShape(x.shape);
const $shape = inferFromImplicitShape(shape, xSize);
const $xSize = sizeFromShape($shape);
assert$1(xSize === $xSize, () => `The new shape (${$shape}) has ${$xSize} elements and the old ` +
`shape (${x.shape}) has ${xSize} elements. The new shape and old ` +
`shape must have the same number of elements.`);
const xTexData = webglBackend.texData.get(x.dataId);
if (xTexData.isPacked && !isReshapeFree(x.shape, $shape) &&
!(xTexData.texture !== null && isReshapeFree(xTexData.shape, $shape))) {
return packedReshape(x, $shape, webglBackend);
}
webglBackend.incRef(x.dataId);
return { dataId: x.dataId, shape: $shape, dtype: x.dtype };
}
const reshapeConfig$1 = {
kernelName: Reshape$1,
backendName: 'webgl',
kernelFunc: reshape$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class MeanProgram {
constructor(reduceInfo, divisor) {
this.variableNames = ['x'];
const { windowSize, batchSize, inSize, outSize } = reduceInfo;
this.outputShape = [batchSize, outSize];
const windowSizeNearestVec4 = Math.floor(windowSize / 4) * 4;
const windowSizeVec4Remainder = windowSize % 4;
let updateSnippet = `sumValue += dot(values, ones);`;
if (divisor != null) {
const denominator = 1 / divisor;
updateSnippet = `sumValue += dot(values * ${isInt(denominator) ? denominator.toPrecision(2) :
denominator}, ones);`;
}
let checkOutOfBounds = '';
if (inSize % windowSize > 0) {
checkOutOfBounds = `
if (inIdx < 0 || inIdx >= ${inSize}) {
return 0.0;
}
`;
}
this.userCode = `
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
float getValue(int batch, int inIdx) {
${checkOutOfBounds}
return getX(batch, inIdx);
}
void main() {
ivec2 coords = getOutputCoords();
int batch = coords[0];
int outIdx = coords[1];
int inOffset = outIdx * ${windowSize};
float sumValue = 0.0;
for (int i = 0; i < ${windowSizeNearestVec4}; i += 4) {
int inIdx = inOffset + i;
vec4 values = vec4(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
getValue(batch, inIdx + 2),
getValue(batch, inIdx + 3)
);
${updateSnippet}
}
int inIdx = inOffset + ${windowSizeNearestVec4};
if (${windowSizeVec4Remainder === 1}) {
vec4 values = vec4(getValue(batch, inIdx), 0.0, 0.0, 0.0);
${updateSnippet}
} else if (${windowSizeVec4Remainder === 2}) {
vec4 values = vec4(
getValue(batch, inIdx),
getValue(batch, inIdx + 1), 0.0, 0.0);
${updateSnippet}
} else if (${windowSizeVec4Remainder === 3}) {
vec4 values = vec4(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
getValue(batch, inIdx + 2), 0.0);
${updateSnippet}
}
setOutput(sumValue);
}
`;
}
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ReduceProgram {
constructor(reduceInfo, reduceType) {
this.variableNames = ['x'];
const { windowSize, batchSize, inSize, outSize } = reduceInfo;
this.outputShape = [batchSize, outSize];
let initializationValue = '0.0';
let compareOp = ``;
if (reduceType === 'prod') {
initializationValue = '1.0';
}
else if (reduceType === 'min') {
// WebGL on Firefox Linux can't compile 1/0 so we do 1/eps.
initializationValue = '1.0 / 1e-20';
compareOp = `min`;
}
else if (reduceType === 'max') {
// WebGL on Firefox Linux can't compile 1/0 so we do 1/eps.
initializationValue = '-1.0 / 1e-20';
compareOp = `max`;
}
let returnValue = `${reduceType}(${reduceType}(${reduceType}(` +
'minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])';
if (reduceType === 'sum') {
returnValue = `sumValue`;
}
else if (reduceType === 'prod') {
returnValue = `prodValue`;
}
else if (reduceType === 'all') {
returnValue = `allValue`;
}
else if (reduceType === 'any') {
returnValue = `anyValue`;
}
const windowSizeNearestVec4 = Math.floor(windowSize / 4) * 4;
const windowSizeVec4Remainder = windowSize % 4;
let updateSnippet = `
if (${reduceType === 'sum'}) {
sumValue += dot(values, ones);
} else if (${reduceType === 'prod'}) {
vec2 tmp = vec2(values[0], values[1]) * vec2(values[2], values[3]);
prodValue *= tmp[0] * tmp[1];
} else {
minMaxValue = ${compareOp}(values, minMaxValue);
if (${reduceType === 'min'} || ${reduceType === 'max'}) {
minMaxValue = ${compareOp}(values, minMaxValue);
bvec4 isNaN = isnan(values);
if (isNaN.r || isNaN.g || isNaN.b || isNaN.a) {
minMaxValue = vec4(NAN);
}
}
}
`;
let vecType = `vec4`;
if (reduceType === 'all') {
initializationValue = '1.0';
updateSnippet = `
bool reducedAllValue = all(values);
float floatedReducedAllValue = float(reducedAllValue);
allValue = float(allValue >= 1.0 && floatedReducedAllValue >= 1.0);
`;
vecType = `bvec4`;
}
else if (reduceType === 'any') {
initializationValue = '0.0';
updateSnippet = `
bool reducedAnyValue = any(values);
float floatedReducedAnyValue = float(reducedAnyValue);
anyValue = float(anyValue >= 1.0 || floatedReducedAnyValue >= 1.0);
`;
vecType = `bvec4`;
}
let checkOutOfBounds = '';
if (inSize % windowSize > 0) {
checkOutOfBounds = `
if (inIdx < 0 || inIdx >= ${inSize}) {
return initializationValue;
}
`;
}
this.userCode = `
const float initializationValue = ${initializationValue};
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
float getValue(int batch, int inIdx) {
${checkOutOfBounds}
return getX(batch, inIdx);
}
void main() {
ivec2 coords = getOutputCoords();
int batch = coords[0];
int outIdx = coords[1];
int inOffset = outIdx * ${windowSize};
vec4 minMaxValue = vec4(${initializationValue});
float prodValue = 1.0;
float sumValue = 0.0;
float allValue = 1.0;
float anyValue = 0.0;
for (int i = 0; i < ${windowSizeNearestVec4}; i += 4) {
int inIdx = inOffset + i;
${vecType} values = ${vecType}(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
getValue(batch, inIdx + 2),
getValue(batch, inIdx + 3)
);
${updateSnippet}
}
int inIdx = inOffset + ${windowSizeNearestVec4};
if (${windowSizeVec4Remainder === 1}) {
${vecType} values = ${vecType}(
getValue(batch, inIdx),
initializationValue,
initializationValue,
initializationValue
);
${updateSnippet}
} else if (${windowSizeVec4Remainder === 2}) {
${vecType} values = ${vecType}(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
initializationValue,
initializationValue
);
${updateSnippet}
} else if (${windowSizeVec4Remainder === 3}) {
${vecType} values = ${vecType}(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
getValue(batch, inIdx + 2),
initializationValue
);
${updateSnippet}
}
setOutput(${returnValue});
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Returns an array of configuration objects that describe each stage of the
// reduction.
function getReductionStages(inShape) {
const stages = [];
while (stages.length === 0 || stages[stages.length - 1].outSize !== 1) {
const outSize = stages.length ? stages[stages.length - 1].outSize : inShape[1];
const windowSize = computeOptimalWindowSize(outSize);
stages.push({
inSize: outSize,
windowSize,
outSize: Math.ceil(outSize / windowSize)
});
}
return stages;
}
function reduce(x, dtype, reductionType, backend) {
const reductionStages = getReductionStages(x.shape);
let result = x;
for (let i = 0; i < reductionStages.length; i++) {
const { inSize, windowSize, outSize } = reductionStages[i];
let program;
let previousResult;
if (reductionType === 'mean') {
program = i === 0 ?
new MeanProgram({ windowSize, inSize, batchSize: x.shape[0], outSize }, inSize) :
new MeanProgram({ windowSize, inSize, batchSize: x.shape[0], outSize });
}
else {
program = new ReduceProgram({ windowSize, inSize, batchSize: x.shape[0], outSize }, reductionType);
}
previousResult = result;
result = backend.runWebGLProgram(program, [result], dtype);
if (previousResult.dataId !== x.dataId) {
backend.disposeIntermediateTensorInfo(previousResult);
}
}
return result;
}
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class TransposeProgram {
constructor(aShape, newDim) {
this.variableNames = ['A'];
const outputShape = new Array(aShape.length);
for (let i = 0; i < outputShape.length; i++) {
outputShape[i] = aShape[newDim[i]];
}
this.outputShape = outputShape;
this.rank = outputShape.length;
const dtype = getCoordsDataType(this.rank);
const switched = getSwitchedCoords(newDim);
this.userCode = `
void main() {
${dtype} resRC = getOutputCoords();
setOutput(getA(${switched}));
}
`;
}
}
function getSwitchedCoords(newDim) {
const rank = newDim.length;
if (rank > 6) {
throw Error(`Transpose for rank ${rank} is not yet supported`);
}
const originalOrder = ['resRC.x', 'resRC.y', 'resRC.z', 'resRC.w', 'resRC.u', 'resRC.v'];
const switchedCoords = new Array(rank);
for (let i = 0; i < newDim.length; i++) {
switchedCoords[newDim[i]] = originalOrder[i];
}
return switchedCoords.join();
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class TransposePackedProgram {
constructor(aShape, newDim) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
const outputShape = new Array(aShape.length);
for (let i = 0; i < outputShape.length; i++) {
outputShape[i] = aShape[newDim[i]];
}
this.outputShape = outputShape;
this.rank = outputShape.length;
if (this.rank > 6) {
throw Error(`Packed transpose for rank ${this.rank} is not yet supported.`);
}
const dtype = getCoordsDataType(this.rank);
const outputOrder = getVecChannels('rc', this.rank);
const switchedOrder = new Array(this.rank);
for (let i = 0; i < newDim.length; i++) {
switchedOrder[newDim[i]] = outputOrder[i];
}
const innerDims = `vec2(${switchedOrder.slice(-2).join()})`;
const nextColumn = `++${outputOrder[this.rank - 1]} < ${outputShape[this.rank - 1]}`;
const getc = `getChannel(getA(${switchedOrder.join()}), ${innerDims})`;
this.userCode = `
void main() {
${dtype} rc = getOutputCoords();
vec4 result = vec4(0.);
result[0] = ${getc};
if(${nextColumn}) {
result[1] = ${getc};
}
--${outputOrder[this.rank - 1]};
if(++${outputOrder[this.rank - 2]} < ${outputShape[this.rank - 2]}) {
result[2] = ${getc};
if(${nextColumn}) {
result[3] = ${getc};
}
}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function transposeImpl(x, perm, backend) {
const program = env().getBool('WEBGL_PACK_ARRAY_OPERATIONS') ?
new TransposePackedProgram(x.shape, perm) :
new TransposeProgram(x.shape, perm);
return backend.runWebGLProgram(program, [x], x.dtype);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sumImpl(x, axis, keepDims, backend) {
const reductionIndices = axis;
const xRank = x.shape.length;
const origAxes = parseAxisParam(reductionIndices, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
const sumInputIsTransposed = permutedAxes != null;
let sumInput = x;
if (sumInputIsTransposed) {
sumInput = transposeImpl(x, permutedAxes, backend);
axes = getInnerMostAxes(axes.length, xRank);
}
assertAxesAreInnerMostDims('sum', axes, xRank);
const [sumOutShape, reduceShape] = computeOutAndReduceShapes(sumInput.shape, axes);
let outShape = sumOutShape;
if (keepDims) {
// rather than reshape at the end, set the target shape here.
outShape = expandShapeToKeepDim(sumOutShape, origAxes);
}
const inSize = sizeFromShape(reduceShape);
const xSize = sizeFromShape(x.shape);
const batchSize = xSize / inSize;
const reshapedInput = reshape$1({ inputs: { x: sumInput }, attrs: { shape: [batchSize, inSize] }, backend });
const outType = sumOutType(x.dtype);
const reduced = reduce(reshapedInput, outType, 'sum', backend);
const out = reshape$1({ inputs: { x: reduced }, attrs: { shape: outShape }, backend });
backend.disposeIntermediateTensorInfo(reshapedInput);
backend.disposeIntermediateTensorInfo(reduced);
if (sumInputIsTransposed) {
backend.disposeIntermediateTensorInfo(sumInput);
}
return out;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sum$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
return sumImpl(x, axis, keepDims, backend);
}
const sumConfig$1 = {
kernelName: Sum,
backendName: 'webgl',
kernelFunc: sum$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function transpose(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { perm } = attrs;
const webglBackend = backend;
const xRank = x.shape.length;
const newShape = new Array(xRank);
for (let i = 0; i < newShape.length; i++) {
newShape[i] = x.shape[perm[i]];
}
let out;
if (webglBackend.shouldExecuteOnCPU([x])) {
const xTexData = webglBackend.texData.get(x.dataId);
const values = xTexData.values;
const outValues = transposeImplCPU(values, x.shape, x.dtype, perm, newShape);
out = webglBackend.makeTensorInfo(newShape, x.dtype);
const outData = webglBackend.texData.get(out.dataId);
outData.values = outValues;
}
else {
out = transposeImpl(x, perm, webglBackend);
}
return out;
}
const transposeConfig = {
kernelName: Transpose,
backendName: 'webgl',
kernelFunc: transpose
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Empirically determined minimal shared dimension in matmul before we forward
// to a.mul(b).sum() in order to take advantage of GPU parallelism. See
// https://github.com/tensorflow/tfjs-core/pull/1379 for benchmarks.
const MATMUL_SHARED_DIM_THRESHOLD = 1000;
function batchMatMulImpl({ a, b, transposeA, transposeB, backend, bias = null, preluActivationWeights = null, leakyreluAlpha = 0, activation = null }) {
const aRank = a.shape.length;
const bRank = b.shape.length;
const innerShapeA = transposeA ? a.shape[aRank - 2] : a.shape[aRank - 1];
const innerShapeB = transposeB ? b.shape[bRank - 1] : b.shape[bRank - 2];
const outerShapeA = transposeA ? a.shape[aRank - 1] : a.shape[aRank - 2];
const outerShapeB = transposeB ? b.shape[bRank - 2] : b.shape[bRank - 1];
const outerDimsA = a.shape.slice(0, -2);
const outerDimsB = b.shape.slice(0, -2);
const batchDimA = sizeFromShape(outerDimsA);
const batchDimB = sizeFromShape(outerDimsB);
const outShapeOuterDims = assertAndGetBroadcastShape(a.shape.slice(0, -2), b.shape.slice(0, -2));
const outShape = outShapeOuterDims.concat([outerShapeA, outerShapeB]);
assert$1(innerShapeA === innerShapeB, () => `Error in matMul: inner shapes (${innerShapeA}) and (` +
`${innerShapeB}) of Tensors with shapes ${a.shape} and ` +
`${b.shape} and transposeA=${transposeA}` +
` and transposeB=${transposeB} must match.`);
const a3dShape = transposeA ?
[batchDimA, innerShapeA, outerShapeA] :
[batchDimA, outerShapeA, innerShapeA];
const b3dShape = transposeB ?
[batchDimB, outerShapeB, innerShapeB] :
[batchDimB, innerShapeB, outerShapeB];
// The rest of the implementation is designed to operate on rank-3 tensors
const a3d = reshape$1({ inputs: { x: a }, backend, attrs: { shape: a3dShape } });
const b3d = reshape$1({ inputs: { x: b }, backend, attrs: { shape: b3dShape } });
const intermediates = [a3d, b3d];
const batchDim = Math.max(batchDimA, batchDimB);
const sharedDim = transposeA ? a3d.shape[1] : a3d.shape[2];
const hasBias = bias != null;
const hasPreluActivationWeights = preluActivationWeights != null;
const hasLeakyreluAlpha = activation === 'leakyrelu';
const fusedActivation = activation != null ?
mapActivationToShaderProgram(activation, true) :
null;
const containsFusedOps = hasBias || hasPreluActivationWeights ||
hasLeakyreluAlpha || fusedActivation != null;
let out;
// Since the matrices are vectors, it is faster to call mul().sum()
// because sum() is O(sqrt(N)) due to divide-and-conquer.
if ((outerShapeA === 1 || outerShapeB === 1) &&
sharedDim > MATMUL_SHARED_DIM_THRESHOLD && containsFusedOps === false) {
let aVec = a3d;
let bVec = b3d;
if (transposeA) {
aVec = transpose({ inputs: { x: a3d }, backend, attrs: { perm: [0, 2, 1] } });
intermediates.push(aVec);
}
if (transposeB) {
bVec = transpose({ inputs: { x: b3d }, backend, attrs: { perm: [0, 2, 1] } });
intermediates.push(bVec);
}
const shouldReshapeA = outerShapeB !== 1;
const shouldReshapeB = outerShapeB === 1;
let aVec3d = aVec;
if (shouldReshapeA) {
aVec3d = reshape$1({
inputs: { x: aVec },
backend,
attrs: { shape: [batchDim, sharedDim, 1] }
});
intermediates.push(aVec3d);
}
const axis = outerShapeB === 1 ? 2 : 1;
let bVec3d = bVec;
if (shouldReshapeB) {
bVec3d = reshape$1({
inputs: { x: bVec },
backend,
attrs: { shape: [batchDim, 1, sharedDim] }
});
intermediates.push(bVec3d);
}
const product = multiply({ inputs: { a: aVec3d, b: bVec3d }, backend });
out = sum$1({ inputs: { x: product }, backend, attrs: { axis, keepDims: true } });
intermediates.push(product);
}
else {
const dtype = upcastType(a.dtype, b.dtype);
const program = new MatMulPackedProgram(a3dShape, b3dShape, [batchDim, outerShapeA, outerShapeB], transposeA, transposeB, hasBias, fusedActivation, hasPreluActivationWeights, hasLeakyreluAlpha);
const inputs = [a3d, b3d];
if (bias != null) {
inputs.push(bias);
}
if (hasPreluActivationWeights) {
inputs.push(preluActivationWeights);
}
if (hasLeakyreluAlpha) {
const $leakyreluAlpha = backend.makeTensorInfo([], 'float32', createScalarValue(leakyreluAlpha, 'float32'));
inputs.push($leakyreluAlpha);
intermediates.push($leakyreluAlpha);
}
out = backend.runWebGLProgram(program, inputs, dtype);
}
const outReshaped = reshape$1({ inputs: { x: out }, backend, attrs: { shape: outShape } });
intermediates.push(out);
for (const i of intermediates) {
backend.disposeIntermediateTensorInfo(i);
}
return outReshaped;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function _fusedMatMul$1(args) {
const { inputs, backend, attrs } = args;
const { a, b, bias, preluActivationWeights } = inputs;
const { transposeA, transposeB, activation, leakyreluAlpha } = attrs;
return batchMatMulImpl({
a,
b,
transposeA,
transposeB,
backend,
bias,
preluActivationWeights,
leakyreluAlpha,
activation
});
}
const _fusedMatMulConfig$1 = {
kernelName: _FusedMatMul,
backendName: 'webgl',
kernelFunc: _fusedMatMul$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ABS = `return abs(x);`;
function abs(args) {
const { inputs, backend } = args;
const { x } = inputs;
// TODO: handle cases when x is complex. Once the cpu implementation
// can handle complex values, refactor to use unaryKernelFunc.
if (backend.shouldExecuteOnCPU([x]) && x.dtype !== 'complex64') {
const xData = backend.texData.get(x.dataId);
const outValues = simpleAbsImplCPU(xData.values);
return backend.makeTensorInfo(x.shape, x.dtype, outValues);
}
let program;
if (env().getBool('WEBGL_PACK_UNARY_OPERATIONS')) {
program = new UnaryOpPackedProgram(x.shape, ABS);
}
else {
program = new UnaryOpProgram(x.shape, ABS);
}
return backend.runWebGLProgram(program, [x], x.dtype);
}
const absConfig = {
kernelName: Abs,
backendName: 'webgl',
kernelFunc: abs
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ACOS = CHECK_NAN_SNIPPET$1 + `
if (abs(x) > 1.) {
return NAN;
}
return acos(x);
`;
const acos$1 = unaryKernelFunc({ opSnippet: ACOS });
const acosConfig$1 = {
kernelName: Acos,
backendName: 'webgl',
kernelFunc: acos$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ACOSH = CHECK_NAN_SNIPPET$1 + `
if (x < 1.0) return NAN;
return log(x + sqrt(x * x - 1.0));`;
const acosh$1 = unaryKernelFunc({ opSnippet: ACOSH });
const acoshConfig$1 = {
kernelName: Acosh,
backendName: 'webgl',
kernelFunc: acosh$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ADD = 'return a + b;';
const addKernelFunc = binaryKernelFunc({
opSnippet: ADD,
packedOpSnippet: ADD,
supportsComplex: true,
cpuKernelImpl: addImplCPU
});
const addConfig = {
kernelName: Add,
backendName: 'webgl',
kernelFunc: addKernelFunc
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class AddNProgram {
constructor(outputShape, shapes) {
this.outputShape = [];
this.outputShape = outputShape;
this.variableNames = shapes.map((_, i) => `T${i}`);
const snippets = [];
// Get target elements from every input tensor.
this.variableNames.forEach(variable => {
snippets.push(`float v${variable} = get${variable}AtOutCoords();`);
});
// Calculate the sum of all elements.
const operation = this.variableNames
.map(variable => {
return `v${variable}`;
})
.join(' + ');
this.userCode = `
void main() {
${snippets.join('\n ')}
float result = ${operation};
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class AddNPackedProgram {
constructor(outputShape, shapes) {
this.outputShape = [];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = outputShape;
this.variableNames = shapes.map((_, i) => `T${i}`);
const snippets = [];
// Get target elements from every input tensor.
this.variableNames.forEach(variable => {
snippets.push(`vec4 v${variable} = get${variable}AtOutCoords();`);
});
// Calculate the sum of all elements.
const operation = this.variableNames
.map(variable => {
return `v${variable}`;
})
.join(' + ');
this.userCode = `
void main() {
${snippets.join('\n ')}
vec4 result = ${operation};
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function addN$1(args) {
const { inputs, backend } = args;
const tensors = inputs;
if (tensors.length === 1) {
return identity({ inputs: { x: tensors[0] }, backend });
}
// Limit the number of uploaded textures for optimization.
if (tensors.length > env().getNumber('WEBGL_MAX_TEXTURES_IN_SHADER')) {
const midIndex = Math.floor(tensors.length / 2);
const leftSide = addN$1({ inputs: tensors.slice(0, midIndex), backend });
const rightSide = addN$1({ inputs: tensors.slice(midIndex), backend });
return addN$1({ inputs: [leftSide, rightSide], backend });
}
const dtype = tensors.map(t => t.dtype).reduce((d1, d2) => upcastType(d1, d2));
const shapes = tensors.map(t => t.shape);
// We can make sure shapes are identical in op level.
const usePackedOp = env().getBool('WEBGL_PACK');
const program = usePackedOp ?
new AddNPackedProgram(tensors[0].shape, shapes) :
new AddNProgram(tensors[0].shape, shapes);
return backend.runWebGLProgram(program, tensors, dtype);
}
const addNConfig$1 = {
kernelName: AddN,
backendName: 'webgl',
kernelFunc: addN$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function all$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
const xRank = x.shape.length;
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
let permutedX = x;
if (permutedAxes != null) {
permutedX = transpose({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
axes = getInnerMostAxes(axes.length, xRank);
}
assertAxesAreInnerMostDims('all', axes, xRank);
const [outShape, reduceShape] = computeOutAndReduceShapes(permutedX.shape, axes);
const inSize = sizeFromShape(reduceShape);
const a2D = reshape$1({ inputs: { x: permutedX }, backend, attrs: { shape: [-1, inSize] } });
const reduced = reduce(a2D, a2D.dtype, 'all', backend);
let res;
if (keepDims) {
const newShape = expandShapeToKeepDim(outShape, origAxes);
res = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: newShape } });
}
else {
res = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: outShape } });
}
backend.disposeIntermediateTensorInfo(a2D);
backend.disposeIntermediateTensorInfo(reduced);
if (permutedAxes != null) {
backend.disposeIntermediateTensorInfo(permutedX);
}
return res;
}
const allConfig$1 = {
kernelName: All,
backendName: 'webgl',
kernelFunc: all$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function any$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
const xRank = x.shape.length;
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
let permutedX = x;
if (permutedAxes != null) {
permutedX = transpose({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
axes = getInnerMostAxes(axes.length, xRank);
}
assertAxesAreInnerMostDims('any', axes, xRank);
const [outShape, reduceShape] = computeOutAndReduceShapes(permutedX.shape, axes);
const inSize = sizeFromShape(reduceShape);
const a2D = reshape$1({ inputs: { x: permutedX }, backend, attrs: { shape: [-1, inSize] } });
const reduced = reduce(a2D, a2D.dtype, 'any', backend);
let res;
if (keepDims) {
const newShape = expandShapeToKeepDim(outShape, origAxes);
res = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: newShape } });
}
else {
res = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: outShape } });
}
backend.disposeIntermediateTensorInfo(a2D);
backend.disposeIntermediateTensorInfo(reduced);
if (permutedAxes != null) {
backend.disposeIntermediateTensorInfo(permutedX);
}
return res;
}
const anyConfig$1 = {
kernelName: Any,
backendName: 'webgl',
kernelFunc: any$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ArgMinMaxProgram {
constructor(reduceInfo, op, firstPass) {
this.variableNames = ['A'];
const { windowSize, batchSize, outSize } = reduceInfo;
if (!firstPass) {
this.variableNames.push('bestIndicesA');
}
this.outputShape = [batchSize, outSize];
const compOp = (op === 'max') ? '>' : '<';
const indexSnippet = firstPass ?
'inOffset + i;' :
'round(getBestIndicesA(batch, inOffset + i));';
this.userCode = `
void main() {
ivec2 coords = getOutputCoords();
int batch = coords[0];
int outIdx = coords[1];
int inOffset = outIdx * ${windowSize};
int bestIndex = inOffset;
float bestValue = getA(batch, bestIndex);
for (int i = 0; i < ${windowSize}; i++) {
int inIdx = ${indexSnippet};
float candidate = getA(batch, inIdx);
if (candidate ${compOp} bestValue) {
bestValue = candidate;
bestIndex = inIdx;
}
}
setOutput(float(bestIndex));
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ArgMinMaxPackedProgram {
constructor(shape, windowSize, op, firstPass) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
assert$1(shape.length > 2, () => `Packed arg${op.charAt(0).toUpperCase() +
op.slice(1)} supports only inputs with rank above 2.`);
const inSize = shape[shape.length - 1];
const outSize = Math.ceil(inSize / windowSize);
this.outputShape = shape.slice(0, -1);
if (outSize > 1) {
this.outputShape.push(outSize);
}
if (!firstPass) {
this.variableNames.push('bestIndicesA');
}
const outShape = this.outputShape;
const rank = outShape.length;
const dtype = getCoordsDataType(rank);
const coords = getChannels('coords', rank);
let sourceLocSetup;
let sourceRank;
if (outSize === 1) {
sourceRank = rank + 1;
const sourceLocDType = getCoordsDataType(sourceRank);
sourceLocSetup = `
${sourceLocDType} sourceLocR = ${sourceLocDType}(${coords.join()}, 0);
++${coords[rank - 1]};
${sourceLocDType} sourceLocG = ${sourceLocDType}(${coords.join()}, 0);
++${coords[rank - 2]};
${sourceLocDType} sourceLocA = ${sourceLocDType}(${coords.join()}, 0);
--${coords[rank - 1]};
${sourceLocDType} sourceLocB = ${sourceLocDType}(${coords.join()}, 0);
--${coords[rank - 2]};`;
}
else {
sourceRank = rank;
sourceLocSetup = `
${dtype} sourceLocR = coords;
++${coords[rank - 1]};
${dtype} sourceLocG = coords;
++${coords[rank - 2]};
${dtype} sourceLocA = coords;
--${coords[rank - 1]};
${dtype} sourceLocB = coords;
--${coords[rank - 2]};`;
}
const channels = ['x', 'y', 'z', 'w', 'u', 'v'].slice(0, sourceRank);
const inChannel = '.' + channels[sourceRank - 1]; // e.g. ".b" for rank 3.
const intChannels = channels.map(x => 'int ' + x);
const srcRCoords = getChannels('sourceLocR', sourceRank - 1).concat('inIdx.r');
const srcGCoords = getChannels('sourceLocG', sourceRank - 1).concat('inIdx.g');
const srcBCoords = getChannels('sourceLocB', sourceRank - 1).concat('inIdx.b');
const srcACoords = getChannels('sourceLocA', sourceRank - 1).concat('inIdx.a');
const compOp = (op === 'max') ? 'greaterThan' : 'lessThan';
const fetchCandidateIdx = firstPass ? '' : `
inIdx = round(vec4(getBestIndicesAChannel(${srcRCoords.join()}),
getBestIndicesAChannel(${srcGCoords.join()}),
getBestIndicesAChannel(${srcBCoords.join()}),
getBestIndicesAChannel(${srcACoords.join()})));`;
const fetchValue = `vec4(
getAChannel(${srcRCoords.join()}),
hasNextCol ? getAChannel(${srcGCoords.join()}) : 0.,
hasNextRow ? getAChannel(${srcBCoords.join()}) : 0.,
hasNextRow && hasNextCol ? getAChannel(${srcACoords.join()}) : 0.)`;
const getBestIndicesAChannelSnippet = firstPass ? '' : `
float getBestIndicesAChannel(${intChannels.join()}) {
return getChannel(getBestIndicesA(${channels.join()}),
vec2(${channels.slice(-2).join()}));
}`;
this.userCode = `
float getAChannel(${intChannels.join()}) {
return getChannel(getA(${channels.join()}),
vec2(${channels.slice(-2).join()}));
}
${getBestIndicesAChannelSnippet}
void main() {
${dtype} coords = getOutputCoords();
bool hasNextCol = ${coords[rank - 1]} < ${outShape[rank - 1] - 1};
bool hasNextRow = ${coords[rank - 2]} < ${outShape[rank - 2] - 1};
${sourceLocSetup}
ivec4 srcIdx = ivec4(sourceLocR${inChannel}, sourceLocG${inChannel},
sourceLocB${inChannel}, sourceLocA${inChannel}) * ${windowSize};
ivec4 inIdx = srcIdx;
vec4 bestIndex = vec4(inIdx);
vec4 bestValue = ${fetchValue};
for (int i = 0; i < ${windowSize}; i++) {
inIdx = srcIdx;
${fetchCandidateIdx}
vec4 candidate = ${fetchValue};
bvec4 nan = isnan(candidate);
bvec4 replace = bvec4(
vec4(${compOp}(candidate, bestValue)) * (vec4(1.0) - vec4(nan)));
bestValue = vec4(replace.x ? candidate.x : bestValue.x,
replace.y ? candidate.y : bestValue.y,
replace.z ? candidate.z : bestValue.z,
replace.w ? candidate.w : bestValue.w);
bestIndex = mix(bestIndex, vec4(inIdx), vec4(replace));
srcIdx++;
}
setOutput(bestIndex);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function argReduce(backend, x, reduceType, bestIndicesA = null) {
let batchSize = x.shape[0];
let inSize = x.shape[1];
if (bestIndicesA != null) {
batchSize = bestIndicesA.shape[0];
inSize = bestIndicesA.shape[1];
}
const windowSize = computeOptimalWindowSize(inSize);
const reduceInfo = { windowSize, inSize, batchSize, outSize: Math.ceil(inSize / windowSize) };
const program = new ArgMinMaxProgram(reduceInfo, reduceType, bestIndicesA == null);
const inputs = [x];
if (bestIndicesA != null) {
inputs.push(bestIndicesA);
}
const output = backend.runWebGLProgram(program, inputs, 'int32');
// No need to run another GPGPU program.
if (output.shape[1] === 1) {
return output;
}
const result = argReduce(backend, x, reduceType, output);
backend.disposeIntermediateTensorInfo(output);
return result;
}
function argReducePacked(backend, x, reduceType, bestIndicesA = null) {
const inShape = bestIndicesA != null ? bestIndicesA.shape : x.shape;
const inSize = inShape[inShape.length - 1];
const windowSize = computeOptimalWindowSize(inSize);
const program = new ArgMinMaxPackedProgram(inShape, windowSize, reduceType, bestIndicesA == null);
const inputs = bestIndicesA == null ? [x] : [x, bestIndicesA];
const output = backend.runWebGLProgram(program, inputs, 'int32');
if (output.shape.length === x.shape.length) {
const result = argReducePacked(backend, x, reduceType, output);
backend.disposeIntermediateTensorInfo(output);
return result;
}
return output;
}
function argMinMaxReduce(backend, x, axis, reduceType) {
const axes = [axis];
assertAxesAreInnerMostDims('arg' + reduceType.charAt(0).toUpperCase() + reduceType.slice(1), axes, x.shape.length);
if (!env().getBool('WEBGL_PACK_REDUCE') || x.shape.length <= 2) {
const intermediateTensorInfos = [];
// Eagerly unpack x input since it is passed in to all the shaders which
// require unpacked inputs.
const xtexData = backend.texData.get(x.dataId);
const xIsPacked = xtexData !== null && xtexData.isPacked;
let xUnPacked = x;
if (xIsPacked) {
xUnPacked = backend.unpackTensor(x);
intermediateTensorInfos.push(xUnPacked);
}
const [outShape, reduceShape] = computeOutAndReduceShapes(xUnPacked.shape, axes);
const inSize = sizeFromShape(reduceShape);
const a2D = reshape$1({ inputs: { x: xUnPacked }, backend, attrs: { shape: [-1, inSize] } });
intermediateTensorInfos.push(a2D);
const reduced = argReduce(backend, a2D, reduceType);
intermediateTensorInfos.push(reduced);
const reshaped = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: outShape } });
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return reshaped;
}
return argReducePacked(backend, x, reduceType);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function argMax$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis } = attrs;
let axes = parseAxisParam(axis, x.shape);
const permutedAxes = getAxesPermutation(axes, x.shape.length);
let $x = x;
const intermediateTensorInfos = [];
if (permutedAxes != null) {
$x = transpose({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
intermediateTensorInfos.push($x);
axes = getInnerMostAxes(axes.length, $x.shape.length);
}
assertAxesAreInnerMostDims('argMax', [axes[0]], $x.shape.length);
const out = argMinMaxReduce(backend, $x, axes[0], 'max');
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return out;
}
const argMaxConfig$1 = {
kernelName: ArgMax,
backendName: 'webgl',
kernelFunc: argMax$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function argMin$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis } = attrs;
let axes = parseAxisParam(axis, x.shape);
const permutedAxes = getAxesPermutation(axes, x.shape.length);
let $x = x;
const intermediateTensorInfos = [];
if (permutedAxes != null) {
$x = transpose({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
intermediateTensorInfos.push($x);
axes = getInnerMostAxes(axes.length, $x.shape.length);
}
assertAxesAreInnerMostDims('argMin', [axes[0]], $x.shape.length);
const out = argMinMaxReduce(backend, $x, axes[0], 'min');
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return out;
}
const argMinConfig$1 = {
kernelName: ArgMin,
backendName: 'webgl',
kernelFunc: argMin$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ASIN = CHECK_NAN_SNIPPET$1 + `
if (abs(x) > 1.) {
return NAN;
}
return asin(x);
`;
const asin$1 = unaryKernelFunc({ opSnippet: ASIN });
const asinConfig$1 = {
kernelName: Asin,
backendName: 'webgl',
kernelFunc: asin$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ASINH = CHECK_NAN_SNIPPET$1 + `return log(x + sqrt(x * x + 1.0));`;
const asinh$1 = unaryKernelFunc({ opSnippet: ASINH });
const asinhConfig$1 = {
kernelName: Asinh,
backendName: 'webgl',
kernelFunc: asinh$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ATAN = CHECK_NAN_SNIPPET$1 + `
return atan(x);
`;
const atan$1 = unaryKernelFunc({ opSnippet: ATAN });
const atanConfig$1 = {
kernelName: Atan,
backendName: 'webgl',
kernelFunc: atan$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ATAN2 = CHECK_NAN_SNIPPET + `
return atan(a, b);
`;
const ATAN2_PACKED = `
vec4 result = atan(a, b);
bvec4 isNaNA = isnan(a);
bvec4 isNaNB = isnan(b);
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
` +
CHECK_NAN_SNIPPET_PACKED + `
return result;
`;
const atan2$1 = binaryKernelFunc({ opSnippet: ATAN2, packedOpSnippet: ATAN2_PACKED });
const atan2Config$1 = {
kernelName: Atan2,
backendName: 'webgl',
kernelFunc: atan2$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ATANH = CHECK_NAN_SNIPPET$1 + `
if ((x < -1.0) || (x > 1.0)) return NAN;
return (log(1.0 + x) - log(1.0 - x)) / 2.0;`;
const atanh$1 = unaryKernelFunc({ opSnippet: ATANH });
const atanhConfig$1 = {
kernelName: Atanh,
backendName: 'webgl',
kernelFunc: atanh$1,
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Pool2DProgram {
constructor(convInfo, poolType, computePositions, flattenPositions = false, includeBatchInIndex = false) {
this.variableNames = ['x'];
if (poolType === 'avg' && computePositions) {
throw new Error('Cannot compute positions for average pool.');
}
const filterWidth = convInfo.filterWidth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
this.outputShape = convInfo.outShape;
const isAvgPool = poolType === 'avg';
const batchFlattenPositionStr = `((batch * ${convInfo.inHeight} + xR) * ${convInfo.inWidth} + xC) * ${convInfo.inChannels} + d`;
const flattenPositionStr = `(xR * ${convInfo.inWidth} + xC) * ${convInfo.inChannels} + d`;
let initializationValue = '0.0';
if (!isAvgPool) {
// WebGL on Firefox Linux can't compile 1/0 so we do 1/eps.
initializationValue = '-1.0 / 1e-20';
}
if (computePositions) {
const compareOp = '>=';
this.userCode = `
const ivec2 strides = ivec2(${strideHeight}, ${strideWidth});
const ivec2 pads = ivec2(${padTop}, ${padLeft});
void main() {
ivec4 coords = getOutputCoords();
int batch = coords[0];
int d = coords[3];
ivec2 xRCCorner = coords.yz * strides - pads;
int xRCorner = xRCCorner.x;
int xCCorner = xRCCorner.y;
// max/min x(?, ?, d) to get y(yR, yC, d).
// ? = to be determined
float minMaxValue = 0.0;
float minMaxValueFound = 0.0;
int minMaxPosition = 0;
float avgValue = 0.0;
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
int xR = xRCorner + wR;
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int wC = 0; wC < ${effectiveFilterWidth};
wC += ${dilationWidth}) {
int xC = xCCorner + wC;
if (xC < 0 || xC >= ${convInfo.inWidth}) {
continue;
}
float value = getX(batch, xR, xC, d);
// If a min / max value has already been found, use it. If not,
// use the current value.
float currMinMaxValue = mix(
value, minMaxValue, minMaxValueFound);
if (value ${compareOp} currMinMaxValue) {
minMaxValue = value;
minMaxValueFound = 1.0;
minMaxPosition = ${flattenPositions ? (includeBatchInIndex ? batchFlattenPositionStr :
flattenPositionStr) :
`wR * ${effectiveFilterWidth} + wC`};
}
}
}
setOutput(float(minMaxPosition));
}
`;
return;
}
const compareOp = 'max';
let returnValue = `${poolType}(${poolType}(${poolType}(` +
'minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])';
if (poolType === 'avg') {
returnValue = `avgValue / max(count, 1.0)`;
}
const filterWidthNearestVec4 = Math.floor(filterWidth / 4) * 4;
const filterWidthVec4Remainder = filterWidth % 4;
const updateSnippet = `
if (${isAvgPool}) {
avgValue += dot(values, ones);
} else {
minMaxValue = ${compareOp}(values, minMaxValue);
}
`;
this.userCode = `
const ivec2 strides = ivec2(${strideHeight}, ${strideWidth});
const ivec2 pads = ivec2(${padTop}, ${padLeft});
const float initializationValue = ${initializationValue};
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
float count = 0.0;
float getValue(int batch, int xR, int xC, int d) {
if (xC < 0 || xC >= ${convInfo.inWidth}) {
return initializationValue;
}
count += 1.0;
return getX(batch, xR, xC, d);
}
void main() {
ivec4 coords = getOutputCoords();
int batch = coords[0];
int d = coords[3];
ivec2 xRCCorner = coords.yz * strides - pads;
int xRCorner = xRCCorner.x;
int xCCorner = xRCCorner.y;
// max/min x(?, ?, d) to get y(yR, yC, d).
// ? = to be determined
vec4 minMaxValue = vec4(${initializationValue});
float avgValue = 0.0;
count = 0.0;
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
int xR = xRCorner + wR;
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int wC = 0; wC < ${filterWidthNearestVec4}; wC += 4) {
int xC = xCCorner + wC * ${dilationWidth};
vec4 values = vec4(
getValue(batch, xR, xC, d),
getValue(batch, xR, xC + ${dilationWidth}, d),
getValue(batch, xR, xC + 2 * ${dilationWidth}, d),
getValue(batch, xR, xC + 3 * ${dilationWidth}, d)
);
${updateSnippet}
}
int xC = xCCorner + ${filterWidthNearestVec4};
if (${filterWidthVec4Remainder === 1}) {
vec4 values = vec4(
getValue(batch, xR, xC, d),
initializationValue,
initializationValue,
initializationValue
);
${updateSnippet}
} else if (${filterWidthVec4Remainder === 2}) {
vec4 values = vec4(
getValue(batch, xR, xC, d),
getValue(batch, xR, xC + ${dilationWidth}, d),
initializationValue,
initializationValue
);
${updateSnippet}
} else if (${filterWidthVec4Remainder === 3}) {
vec4 values = vec4(
getValue(batch, xR, xC, d),
getValue(batch, xR, xC + ${dilationWidth}, d),
getValue(batch, xR, xC + 2 * ${dilationWidth}, d),
initializationValue
);
${updateSnippet}
}
}
setOutput(${returnValue});
}
`;
}
}
class Pool3DProgram {
constructor(convInfo, poolType, computePositions, flattenPositions = false, includeBatchInIndex = false) {
this.variableNames = ['x'];
if (poolType === 'avg' && computePositions) {
throw new Error('Cannot compute positions for average pool.');
}
const filterWidth = convInfo.filterWidth;
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterDepth = convInfo.effectiveFilterDepth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padFront = convInfo.padInfo.front;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
this.outputShape = convInfo.outShape;
const isAvgPool = poolType === 'avg';
let initializationValue = '0.0';
if (!isAvgPool) {
// WebGL on Firefox Linux can't compile 1/0 so we do 1/eps.
initializationValue = '-1.0 / 1e-20';
}
if (computePositions) {
const compareOp = '>=';
this.userCode = `
const ivec3 strides =
ivec3(${strideDepth}, ${strideHeight}, ${strideWidth});
const ivec3 pads = ivec3(${padFront}, ${padTop}, ${padLeft});
void main() {
ivec5 coords = getOutputCoords();
int batch = coords.x;
int ch = coords.u;
ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
int xDCorner = xCorner.x;
int xRCorner = xCorner.y;
int xCCorner = xCorner.z;
// max/min x(?, ?, ?, ch) to get y(yD, yR, yC, ch).
// ? = to be determined
float minMaxValue = 0.0;
float minMaxValueFound = 0.0;
int minMaxPosition = 0;
for (int wD = 0; wD < ${effectiveFilterDepth};
wD += ${dilationDepth}) {
int xD = xDCorner + wD;
if (xD < 0 || xD >= ${convInfo.inDepth}) {
continue;
}
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
int xR = xRCorner + wR;
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int wC = 0; wC < ${effectiveFilterWidth};
wC += ${dilationWidth}) {
int xC = xCCorner + wC;
if (xC < 0 || xC >= ${convInfo.inWidth}) {
continue;
}
float value = getX(batch, xD, xR, xC, ch);
// If a min / max value has already been found, use it. If not,
// use the current value.
float currMinMaxValue = mix(
value, minMaxValue, minMaxValueFound);
if (value ${compareOp} currMinMaxValue) {
minMaxValue = value;
minMaxValueFound = 1.0;
minMaxPosition = ${flattenPositions ?
(includeBatchInIndex ?
`(((batch * ${convInfo.inDepth} + xD) * ${convInfo.inHeight} + xR) * ${convInfo.inWidth} + xC) * ${convInfo.inChannels} + ch` :
`((xD * ${convInfo.inHeight} + xR) * ${convInfo.inWidth} + xC) * ${convInfo.inChannels} + ch`) :
`wD * ${effectiveFilterHeight} * ${effectiveFilterWidth} +
wR * ${effectiveFilterWidth} + wC`};
}
}
}
}
setOutput(float(minMaxPosition));
}
`;
return;
}
const compareOp = 'max';
let returnValue = `${poolType}(${poolType}(${poolType}(` +
'minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])';
if (poolType === 'avg') {
// Use `max(count, 1.0)` instead of `count` in case count === 0.0.
// If count === 0.0, `avgValue` is always 0.0 and we change `count`'s
// value to avoid dividing zero.
returnValue = `avgValue / max(count, 1.0)`;
}
const filterWidthNearestVec4 = Math.floor(filterWidth / 4) * 4;
const filterWidthVec4Remainder = filterWidth % 4;
const updateSnippet = `
if (${isAvgPool}) {
avgValue += dot(values, ones);
} else {
minMaxValue = ${compareOp}(values, minMaxValue);
}
`;
this.userCode = `
const ivec3 strides =
ivec3(${strideDepth}, ${strideHeight}, ${strideWidth});
const ivec3 pads = ivec3(${padFront}, ${padTop}, ${padLeft});
const float initializationValue = ${initializationValue};
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
float count = 0.0;
float getValue(int batch, int xD, int xR, int xC, int ch) {
if (xC < 0 || xC >= ${convInfo.inWidth}) {
return initializationValue;
}
count += 1.0;
return getX(batch, xD, xR, xC, ch);
}
void main() {
ivec5 coords = getOutputCoords();
int batch = coords.x;
int ch = coords.u;
ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
int xDCorner = xCorner.x;
int xRCorner = xCorner.y;
int xCCorner = xCorner.z;
// max/min x(?, ?, ?, d) to get y(yD, yR, yC, ch).
// ? = to be determined
vec4 minMaxValue = vec4(${initializationValue});
float avgValue = 0.0;
count = 0.0;
for (int wD = 0; wD < ${effectiveFilterDepth};
wD += ${dilationDepth}) {
int xD = xDCorner + wD;
if (xD < 0 || xD >= ${convInfo.inDepth}) {
continue;
}
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
int xR = xRCorner + wR;
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int wC = 0; wC < ${filterWidthNearestVec4}; wC += 4) {
int xC = xCCorner + wC * ${dilationWidth};
vec4 values = vec4(
getValue(batch, xD, xR, xC, ch),
getValue(batch, xD, xR, xC + ${dilationWidth}, ch),
getValue(batch, xD, xR, xC + 2 * ${dilationWidth}, ch),
getValue(batch, xD, xR, xC + 3 * ${dilationWidth}, ch)
);
${updateSnippet}
}
int xC = xCCorner + ${filterWidthNearestVec4};
if (${filterWidthVec4Remainder === 1}) {
vec4 values = vec4(
getValue(batch, xD, xR, xC, ch),
initializationValue,
initializationValue,
initializationValue
);
${updateSnippet}
} else if (${filterWidthVec4Remainder === 2}) {
vec4 values = vec4(
getValue(batch, xD, xR, xC, ch),
getValue(batch, xD, xR, xC + ${dilationWidth}, ch),
initializationValue,
initializationValue
);
${updateSnippet}
} else if (${filterWidthVec4Remainder === 3}) {
vec4 values = vec4(
getValue(batch, xD, xR, xC, ch),
getValue(batch, xD, xR, xC + ${dilationWidth}, ch),
getValue(batch, xD, xR, xC + 2 * ${dilationWidth}, ch),
initializationValue
);
${updateSnippet}
}
}
}
setOutput(${returnValue});
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPool$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
assertNotComplex$1(x, 'avgPool');
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const dilations = 1;
assert$1(eitherStridesOrDilationsAreOne(strides, dilations), () => 'Error in avgPool: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
const convInfo = computePool2DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode);
if (convInfo.filterWidth === 1 && convInfo.filterHeight === 1 &&
arraysEqual(convInfo.inShape, convInfo.outShape)) {
return identity({ inputs: { x }, backend });
}
const avgPoolProgram = new Pool2DProgram(convInfo, 'avg', false);
return backend.runWebGLProgram(avgPoolProgram, [x], 'float32');
}
const avgPoolConfig$1 = {
kernelName: AvgPool,
backendName: 'webgl',
kernelFunc: avgPool$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPool3D$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { filterSize, strides, pad, dimRoundingMode, dataFormat } = attrs;
const dilations = [1, 1, 1];
const convInfo = computePool3DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode, dataFormat);
const avgPoolProgram = new Pool3DProgram(convInfo, 'avg', false);
return backend.runWebGLProgram(avgPoolProgram, [x], 'float32');
}
const avgPool3DConfig$1 = {
kernelName: AvgPool3D,
backendName: 'webgl',
kernelFunc: avgPool3D$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class AvgPool2DBackpropProgram {
constructor(convInfo) {
this.variableNames = ['dy'];
this.outputShape = convInfo.inShape;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const avgMultiplier = 1 / (filterHeight * filterWidth);
this.userCode = `
const ivec2 pads = ivec2(${padTop}, ${padLeft});
const float avgMultiplier = float(${avgMultiplier});
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
ivec2 dyRCCorner = coords.yz - pads;
int dyRCorner = dyRCCorner.x;
int dyCCorner = dyRCCorner.y;
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
float dyR = float(dyRCorner + wR) / ${strideHeight}.0;
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 || fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
for (int wC = 0; wC < ${effectiveFilterWidth};
wC+= ${dilationWidth}) {
float dyC = float(dyCCorner + wC) / ${strideWidth}.0;
if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
fract(dyC) > 0.0) {
continue;
}
int idyC = int(dyC);
float dyValue = getDy(b, idyR, idyC, d);
dotProd += dyValue * avgMultiplier;
}
}
setOutput(dotProd);
}
`;
}
}
class AvgPool3DBackpropProgram {
constructor(convInfo) {
this.variableNames = ['dy'];
this.outputShape = convInfo.inShape;
const filterDepth = convInfo.filterDepth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterDepth = convInfo.effectiveFilterDepth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padFront = effectiveFilterDepth - 1 - convInfo.padInfo.front;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const avgMultiplier = 1 / (filterDepth * filterHeight * filterWidth);
this.userCode = `
const ivec3 pads = ivec3(${padFront}, ${padTop}, ${padLeft});
const float avgMultiplier = float(${avgMultiplier});
void main() {
ivec5 coords = getOutputCoords();
int batch = coords.x;
int ch = coords.u;
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
int dyDCorner = dyCorner.x;
int dyRCorner = dyCorner.y;
int dyCCorner = dyCorner.z;
// Convolve dy(?, ?, ?, d) with pos mask(:, :, :, ch) to get
// dx(xD, xR, xC, ch).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
for (int wD = 0; wD < ${effectiveFilterDepth};
wD += ${dilationDepth}) {
float dyD = float(dyDCorner + wD) / ${strideDepth}.0;
if (dyD < 0.0 || dyD >= ${convInfo.outDepth}.0 || fract(dyD) > 0.0) {
continue;
}
int idyD = int(dyD);
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
float dyR = float(dyRCorner + wR) / ${strideHeight}.0;
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 ||
fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
for (int wC = 0; wC < ${effectiveFilterWidth};
wC += ${dilationWidth}) {
float dyC = float(dyCCorner + wC) / ${strideWidth}.0;
if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
fract(dyC) > 0.0) {
continue;
}
int idyC = int(dyC);
float dyValue = getDy(batch, idyD, idyR, idyC, ch);
dotProd += dyValue * avgMultiplier;
}
}
}
setOutput(dotProd);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPool3DGrad$1(args) {
const { inputs, backend, attrs } = args;
const { dy, input } = inputs;
const x = input;
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const dilations = [1, 1, 1];
const convInfo = computePool3DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode);
const avgPoolBackpropProgram = new AvgPool3DBackpropProgram(convInfo);
return backend.runWebGLProgram(avgPoolBackpropProgram, [dy], x.dtype);
}
const avgPool3DGradConfig$2 = {
kernelName: AvgPool3DGrad,
backendName: 'webgl',
kernelFunc: avgPool3DGrad$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPoolGrad$2(args) {
const { inputs, backend, attrs } = args;
const { dy, input } = inputs;
const x = input;
assertNotComplex$1([dy, input], 'avgPoolGrad');
const { filterSize, strides, pad } = attrs;
const convInfo = computePool2DInfo(x.shape, filterSize, strides, 1 /* dilations */, pad);
const avgPoolBackpropProgram = new AvgPool2DBackpropProgram(convInfo);
return backend.runWebGLProgram(avgPoolBackpropProgram, [dy], x.dtype);
}
const avgPoolGradConfig$2 = {
kernelName: AvgPoolGrad,
backendName: 'webgl',
kernelFunc: avgPoolGrad$2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function batchMatMul$1(args) {
const { inputs, backend, attrs } = args;
const { a, b } = inputs;
const { transposeA, transposeB } = attrs;
return batchMatMulImpl({ a, b, transposeA, transposeB, backend });
}
const batchMatMulConfig$1 = {
kernelName: BatchMatMul,
backendName: 'webgl',
kernelFunc: batchMatMul$1,
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class BatchNormProgram {
constructor(xShape, meanShape, varianceShape, offsetShape, scaleShape, varianceEpsilon) {
this.outputShape = [];
this.variableNames = ['x', 'mean', 'variance'];
assertAndGetBroadcastShape(xShape, meanShape);
assertAndGetBroadcastShape(xShape, varianceShape);
let offsetSnippet = '0.0';
if (offsetShape != null) {
assertAndGetBroadcastShape(xShape, offsetShape);
this.variableNames.push('offset');
offsetSnippet = 'getOffsetAtOutCoords()';
}
let scaleSnippet = '1.0';
if (scaleShape != null) {
assertAndGetBroadcastShape(xShape, scaleShape);
this.variableNames.push('scale');
scaleSnippet = 'getScaleAtOutCoords()';
}
this.outputShape = xShape;
this.userCode = `
void main() {
float x = getXAtOutCoords();
float mean = getMeanAtOutCoords();
float variance = getVarianceAtOutCoords();
float offset = ${offsetSnippet};
float scale = ${scaleSnippet};
float inv = scale * inversesqrt(variance + float(${varianceEpsilon}));
setOutput(dot(vec3(x, -mean, offset), vec3(inv, inv, 1)));
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class BatchNormPackedProgram {
constructor(xShape, meanShape, varianceShape, offsetShape, scaleShape, varianceEpsilon) {
this.packedInputs = true;
this.packedOutput = true;
this.variableNames = ['x', 'mean', 'variance'];
assertAndGetBroadcastShape(xShape, meanShape);
assertAndGetBroadcastShape(xShape, varianceShape);
let offsetSnippet = 'vec4(0.0)';
if (offsetShape != null) {
assertAndGetBroadcastShape(xShape, offsetShape);
this.variableNames.push('offset');
offsetSnippet = 'getOffsetAtOutCoords()';
}
let scaleSnippet = 'vec4(1.0)';
if (scaleShape != null) {
assertAndGetBroadcastShape(xShape, scaleShape);
this.variableNames.push('scale');
scaleSnippet = 'getScaleAtOutCoords()';
}
this.outputShape = xShape;
this.userCode = `
void main() {
vec4 offset = ${offsetSnippet};
vec4 scale = ${scaleSnippet};
vec4 x = getXAtOutCoords();
vec4 mean = getMeanAtOutCoords();
vec4 variance = getVarianceAtOutCoords();
vec4 inv = scale * inversesqrt(variance + vec4(${varianceEpsilon}));
setOutput((x - mean) * inv + offset);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const batchNorm$1 = ({ inputs, backend, attrs }) => {
const { x, mean, variance, offset, scale } = inputs;
assert$1(mean.shape.length === variance.shape.length, () => 'Batch normalization gradient requires mean and variance to have ' +
'equal ranks.');
assert$1(offset == null || mean.shape.length === offset.shape.length, () => 'Batch normalization gradient requires mean and offset to have ' +
'equal ranks.');
assert$1(scale == null || mean.shape.length === scale.shape.length, () => 'Batch normalization gradient requires mean and scale to have ' +
'equal ranks.');
let { varianceEpsilon } = attrs;
if (varianceEpsilon == null) {
varianceEpsilon = 0.001;
}
const finalInputs = [x, mean, variance];
let offsetShape = null;
if (offset != null) {
offsetShape = offset.shape;
finalInputs.push(offset);
}
let scaleShape = null;
if (scale != null) {
scaleShape = scale.shape;
finalInputs.push(scale);
}
const program = env().getBool('WEBGL_PACK_NORMALIZATION') ?
new BatchNormPackedProgram(x.shape, mean.shape, variance.shape, offsetShape, scaleShape, varianceEpsilon) :
new BatchNormProgram(x.shape, mean.shape, variance.shape, offsetShape, scaleShape, varianceEpsilon);
const output = backend.runWebGLProgram(program, finalInputs, finalInputs[0].dtype);
return output;
};
const batchNormConfig$1 = {
kernelName: FusedBatchNorm,
backendName: 'webgl',
kernelFunc: batchNorm$1,
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class SliceProgram {
constructor(destSize) {
this.variableNames = ['source'];
this.outputShape = destSize;
this.rank = destSize.length;
const dtype = getCoordsDataType(this.rank);
this.customUniforms = [{ name: 'start', arrayIndex: this.rank, type: 'int' }];
const sourceCoords = getCoords$1(this.rank);
let body;
const coordSum = destSize.map((_, i) => {
return `sourceLoc.${coords[i]} = start[${i}] + coords.${coords[i]};`;
});
body = `
${dtype} sourceLoc;
${dtype} coords = getOutputCoords();
${coordSum.join('\n')}
`;
this.userCode = `
void main() {
${body}
setOutput(getSource(${sourceCoords}));
}
`;
}
}
const coords = ['x', 'y', 'z', 'w', 'u', 'v'];
function getCoords$1(rank) {
if (rank === 1) {
return 'sourceLoc';
}
else if (rank <= 6) {
return coords.slice(0, rank).map(x => 'sourceLoc.' + x).join(',');
}
else {
throw Error(`Slicing for rank ${rank} is not yet supported`);
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class SlicePackedProgram {
constructor(destSize) {
this.variableNames = ['source'];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = destSize;
this.rank = destSize.length;
this.customUniforms = [{ name: 'start', arrayIndex: this.rank, type: 'int' }];
const dtype = getCoordsDataType(this.rank);
const coords = getChannels('coords', this.rank);
const sourceLoc = getChannels('sourceLoc', this.rank);
const innerDims = this.rank === 1 ? 'sourceLoc' : `vec2(${sourceLoc.slice(-2).join()})`;
const getChannel = `getChannel(getSource(${sourceLoc.join()}), ${innerDims})`;
const upperRow = `
result.x = ${getChannel};
if (++${coords[this.rank - 1]} < ${destSize[this.rank - 1]}) {
++${sourceLoc[this.rank - 1]};
result.y = ${getChannel};
--${sourceLoc[this.rank - 1]};
}
`;
const lowerRow = this.rank === 1 ? '' : `
--${coords[this.rank - 1]};
if (++${coords[this.rank - 2]} < ${destSize[this.rank - 2]}) {
++${sourceLoc[this.rank - 2]};
result.z = ${getChannel};
if (++${coords[this.rank - 1]} < ${destSize[this.rank - 1]}) {
++${sourceLoc[this.rank - 1]};
result.w = ${getChannel};
}
}
`;
const sourceLocSetup = this.rank <= 4 ?
`sourceLoc = coords +
${dtype}(${destSize.map((_, i) => `start[${i}]`).join()});` :
destSize.map((_, i) => `${sourceLoc[i]} = ${coords[i]} + start[${i}];`)
.join('\n');
this.userCode = `
void main() {
${dtype} coords = getOutputCoords();
${dtype} sourceLoc;
${sourceLocSetup}
vec4 result = vec4(0.);
${upperRow}
${lowerRow}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function shallowSlice(x, begin, size, backend) {
const xTexData = backend.texData.get(x.dataId);
const t = backend.makeTensorInfo(size, x.dtype);
const newTexData = backend.texData.get(t.dataId);
// Copy texture data from the original tensor.
Object.assign(newTexData, xTexData);
newTexData.refCount = 1;
newTexData.shape = size;
newTexData.dtype = x.dtype;
let flatOffset = computeFlatOffset(begin, computeStrides(x.shape));
if (xTexData.slice) {
// We are slicing an already sliced tensor, so we have to accumulate
// the offset.
flatOffset += xTexData.slice.flatOffset;
}
newTexData.slice = {
flatOffset,
// Point to the original dataId, which is used to do ref counting.
origDataId: xTexData.slice && xTexData.slice.origDataId || x.dataId
};
// Increase the ref count for that data bucket.
const refCount = backend.dataRefCount.get(newTexData.slice.origDataId) || 1;
backend.dataRefCount.set(newTexData.slice.origDataId, refCount + 1);
return t;
}
function slice(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { begin, size } = attrs;
const [$begin, $size] = parseSliceParams(x, begin, size);
assertParamsValid(x, $begin, $size);
if (sizeFromShape($size) === 0) {
return backend.makeTensorInfo($size, x.dtype, []);
}
// Run on cpu if dtype is string. For string, the backend represents it
// as Uint8Array[], where each Uint8Array is a character. Given that the
// computation is only on the outer array, uploading the whole data onto
// gpu is wasteful. Also, currently webgl doesn't have a design to
// upload and retrieve Uint8Array[] between cpu and gpu. Therefore, we
// just run the kernel on cpu if dtype is string.
if (backend.shouldExecuteOnCPU([x]) || x.dtype === 'string') {
const xTexData = backend.texData.get(x.dataId);
const outValues = sliceImplCPU(xTexData.values, $begin, $size, x.shape, x.dtype);
return backend.makeTensorInfo($size, x.dtype, outValues);
}
const { isPacked } = backend.texData.get(x.dataId);
const isContinous = isSliceContinous(x.shape, $begin, $size);
if (isPacked || !isContinous) {
const program = env().getBool('WEBGL_PACK_ARRAY_OPERATIONS') ?
new SlicePackedProgram($size) :
new SliceProgram($size);
const customValues = [$begin];
return backend.runWebGLProgram(program, [x], x.dtype, customValues);
}
backend.uploadToGPU(x.dataId);
return shallowSlice(x, $begin, $size, backend);
}
const sliceConfig = {
kernelName: Slice,
backendName: 'webgl',
kernelFunc: slice
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const batchToSpaceND$1 = (args) => {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { blockShape, crops } = attrs;
assert$1(x.shape.length <= 4, () => 'batchToSpaceND for rank > 4 with a WebGL backend not ' +
'implemented yet');
const prod = blockShape.reduce((a, b) => a * b);
const reshaped = getReshaped(x.shape, blockShape, prod);
const permuted = getPermuted(reshaped.length, blockShape.length);
const reshapedPermuted = getReshapedPermuted(x.shape, blockShape, prod);
const sliceBeginCoords = getSliceBeginCoords(crops, blockShape.length);
const sliceSize = getSliceSize(reshapedPermuted, crops, blockShape.length);
const toDispose = [];
const reshapedIntermediate = reshape$1({ inputs: { x }, backend, attrs: { shape: reshaped } });
const transposedIntermediate = transpose({ inputs: { x: reshapedIntermediate }, backend, attrs: { perm: permuted } });
const reshapedIntermediate2 = reshape$1({
inputs: { x: transposedIntermediate },
backend,
attrs: { shape: reshapedPermuted }
});
const sliced = slice({
inputs: { x: reshapedIntermediate2 },
backend,
attrs: { begin: sliceBeginCoords, size: sliceSize }
});
toDispose.push(reshapedIntermediate);
toDispose.push(transposedIntermediate);
toDispose.push(reshapedIntermediate2);
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return sliced;
};
const batchToSpaceNDConfig$1 = {
kernelName: BatchToSpaceND,
backendName: 'webgl',
kernelFunc: batchToSpaceND$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function bincount$1(args) {
const { inputs, backend, attrs } = args;
const { x, weights } = inputs;
const { size } = attrs;
const xVals = backend.readSync(x.dataId);
const weightsVals = backend.readSync(weights.dataId);
const outVals = bincountImplCPU(xVals, weightsVals, weights.dtype, weights.shape, size);
return backend.makeTensorInfo([size], weights.dtype, outVals);
}
const bincountConfig$1 = {
kernelName: Bincount,
backendName: 'webgl',
kernelFunc: bincount$1
};
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const BITWISEAND = `
int r = int(a.r) & int(b.r);
int g = int(a.g) & int(b.g);
int rb = int(a.b) & int(b.b);
int ra = int(a.a) & int(b.a);
return vec4(r, g, rb, ra);
`;
const BITWISEAND_UNPACKED = `
return float(int(a.r) & int(b.r));
`;
function bitwiseAnd(args) {
const { inputs, backend } = args;
const { a, b } = inputs;
const shouldUsePackedProgram = env().getBool('WEBGL_PACK_BINARY_OPERATIONS');
const versionNumber = env().getNumber('WEBGL_VERSION');
// The type of a and b are ensured to be `int32` in core, therefore no need to
// consider other type situations.
if ((backend.shouldExecuteOnCPU([a, b])) || versionNumber === 1) {
const aVals = backend.texData.get(a.dataId).values;
const bVals = backend.texData.get(b.dataId).values;
const [outValues, outShape] = bitwiseAndImplCPU(a.shape, b.shape, aVals, bVals, a.dtype);
const out = backend.makeTensorInfo(outShape, a.dtype);
const outData = backend.texData.get(out.dataId);
outData.values = outValues;
return out;
}
let program;
if (shouldUsePackedProgram) {
program = new BinaryOpPackedProgram(BITWISEAND, a.shape, b.shape, false);
}
else {
program = new BinaryOpProgram(BITWISEAND_UNPACKED, a.shape, b.shape);
}
return backend.runWebGLProgram(program, [a, b], a.dtype);
}
const bitwiseAndConfig = {
kernelName: BitwiseAnd,
backendName: 'webgl',
kernelFunc: bitwiseAnd
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function broadcastArgs$1(args) {
const { inputs, backend } = args;
const { s0, s1 } = inputs;
const s0Vals = backend.readSync(s0.dataId);
const s1Vals = backend.readSync(s1.dataId);
const broadcastShape = assertAndGetBroadcastShape(Array.from(s0Vals), Array.from(s1Vals));
return backend.makeTensorInfo([broadcastShape.length], 'int32', Int32Array.from(broadcastShape));
}
const broadcastArgsConfig$1 = {
kernelName: BroadcastArgs,
backendName: 'webgl',
kernelFunc: broadcastArgs$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const NOT_EQUAL = `return float(a != b);`;
const notEqual = binaryKernelFunc({ opSnippet: NOT_EQUAL, cpuKernelImpl: notEqualImplCPU, dtype: 'bool' });
const notEqualConfig = {
kernelName: NotEqual,
backendName: 'webgl',
kernelFunc: notEqual,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function real(args) {
const { inputs, backend } = args;
const { input } = inputs;
const inputData = backend.texData.get(input.dataId);
return identity({ inputs: { x: inputData.complexTensorInfos.real }, backend });
}
const realConfig = {
kernelName: Real,
backendName: 'webgl',
kernelFunc: real
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const TO_INT = `return float(int(x));`;
function int(input, backend) {
const program = new UnaryOpProgram(input.shape, TO_INT);
const output = backend.runWebGLProgram(program, [input], 'int32');
return { dataId: output.dataId, shape: output.shape, dtype: output.dtype };
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function cast$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { dtype } = attrs;
// Casting to complex64.
if (dtype === 'complex64') {
if (x.dtype === 'complex64') {
return identity({ inputs: { x }, backend });
}
// TODO(annxingyuan): Import kernel function once zeros is modularized.
const zerosTensor = zeros$1(x.shape);
const floatX = cast$1({ inputs: { x }, backend, attrs: { dtype: 'float32' } });
const result = complex({ inputs: { real: floatX, imag: zerosTensor }, backend });
zerosTensor.dispose();
backend.disposeIntermediateTensorInfo(floatX);
return result;
}
// Casting from complex64
if (x.dtype === 'complex64') {
const realPart = real({ inputs: { input: x }, backend });
const result = cast$1({ inputs: { x: realPart }, backend, attrs: { dtype } });
backend.disposeIntermediateTensorInfo(realPart);
return result;
}
if (!hasEncodingLoss(x.dtype, dtype)) {
// We don't change the underlying data, since we cast to higher
// precision.
const result = identity({ inputs: { x }, backend });
return { dataId: result.dataId, shape: result.shape, dtype };
}
if (backend.shouldExecuteOnCPU([x])) {
const values = backend.texData.get(x.dataId).values;
const [resultShape, resultType, resultData] = castImplCPU(values, x.shape, x.dtype, dtype);
return backend.makeTensorInfo(resultShape, resultType, resultData);
}
if (dtype === 'int32') {
return int(x, backend);
}
if (dtype === 'bool') {
const zerosTensorInfo = backend.makeTensorInfo([], 'bool', getTypedArrayFromDType('bool', 1));
const binaryInputs = { a: x, b: zerosTensorInfo };
const result = notEqual({ inputs: binaryInputs, backend });
backend.disposeIntermediateTensorInfo(zerosTensorInfo);
return result;
}
throw new Error(`Error in Cast: failed to cast ${x.dtype} to ${dtype}`);
}
const castConfig = {
kernelName: Cast,
backendName: 'webgl',
kernelFunc: cast$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const CEIL = `return ceil(x);`;
const ceil = unaryKernelFunc({ opSnippet: CEIL, packedOpSnippet: CEIL, cpuKernelImpl: ceilImplCPU });
const ceilConfig = {
kernelName: Ceil,
backendName: 'webgl',
kernelFunc: ceil
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ClipProgram {
constructor(aShape) {
this.variableNames = ['A'];
this.customUniforms = [
{ name: 'minVal', type: 'float' },
{ name: 'maxVal', type: 'float' }
];
this.outputShape = aShape;
this.userCode = `
void main() {
float value = getAAtOutCoords();
if (isnan(value)) {
setOutput(value);
return;
}
setOutput(clamp(value, minVal, maxVal));
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ClipPackedProgram {
constructor(aShape) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
this.customUniforms = [
{ name: 'minVal', type: 'float' },
{ name: 'maxVal', type: 'float' }
];
this.outputShape = aShape;
this.userCode = `
void main() {
vec4 value = getAAtOutCoords();
if (any(isnan(value))) {
setOutput(value);
return;
}
setOutput(clamp(value, vec4(minVal), vec4(maxVal)));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function clipByValue$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { clipValueMin, clipValueMax } = attrs;
let program;
if (env().getBool('WEBGL_PACK_CLIP')) {
program = new ClipPackedProgram(x.shape);
}
else {
program = new ClipProgram(x.shape);
}
const customValues = [[clipValueMin], [clipValueMax]];
return backend.runWebGLProgram(program, [x], x.dtype, customValues);
}
const clipByValueConfig$1 = {
kernelName: ClipByValue,
backendName: 'webgl',
kernelFunc: clipByValue$1
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ComplexAbsProgram {
constructor(shape) {
this.variableNames = ['real', 'imag'];
this.outputShape = shape;
this.userCode = `
void main() {
float re = abs(getRealAtOutCoords());
float im = abs(getImagAtOutCoords());
float mx = max(re, im);
// sadly the length function in glsl is not underflow-safe
// (at least not on Intel GPUs). So the safe solution is
// to ensure underflow-safety in all cases.
setOutput(
mx == 0.0 ? 0.0 : mx * length(vec2(1, min(re, im)/mx))
);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Returns a TensorInfo with the complex shape and the dataId of the
// underlying part. We need to do this because a reshaped complex tensor is
// not reflected in its parts.
function makeComplexComponentTensorInfo(complexTensor, complexPart) {
return {
dataId: complexPart.dataId,
dtype: complexPart.dtype,
shape: complexTensor.shape
};
}
function complexAbs$1(args) {
const { inputs, backend } = args;
const { x } = inputs;
const xData = backend.texData.get(x.dataId);
const program = new ComplexAbsProgram(x.shape);
const programInputs = [
makeComplexComponentTensorInfo(x, xData.complexTensorInfos.real),
makeComplexComponentTensorInfo(x, xData.complexTensorInfos.imag),
];
return backend.runWebGLProgram(program, programInputs, programInputs[0].dtype);
}
const complexAbsConfig$1 = {
kernelName: ComplexAbs,
backendName: 'webgl',
kernelFunc: complexAbs$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ConcatProgram {
// Concats 2d tensors along axis=1. See comments in MathBackendWebGL.concat().
constructor(shapes) {
this.outputShape = [];
this.outputShape = computeOutShape$1(shapes, 1 /* axis */);
this.variableNames = shapes.map((_, i) => `T${i}`);
const offsets = new Array(shapes.length - 1);
offsets[0] = shapes[0][1];
for (let i = 1; i < offsets.length; i++) {
offsets[i] = offsets[i - 1] + shapes[i][1];
}
const snippets = [`if (yC < ${offsets[0]}) setOutput(getT0(yR, yC));`];
for (let i = 1; i < offsets.length; i++) {
const shift = offsets[i - 1];
snippets.push(`else if (yC < ${offsets[i]}) ` +
`setOutput(getT${i}(yR, yC-${shift}));`);
}
const lastIndex = offsets.length;
const lastShift = offsets[offsets.length - 1];
snippets.push(`else setOutput(getT${lastIndex}(yR, yC-${lastShift}));`);
this.userCode = `
void main() {
ivec2 coords = getOutputCoords();
int yR = coords.x;
int yC = coords.y;
${snippets.join('\n ')}
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ConcatPackedProgram {
constructor(shapes, axis) {
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = [];
this.outputShape = computeOutShape$1(shapes, axis);
const shape = this.outputShape;
const rank = shape.length;
const dtype = getCoordsDataType(rank);
const coords = getChannels('coords', rank);
const channels = ['x', 'y', 'z', 'w', 'u', 'v'].slice(0, rank);
this.variableNames = shapes.map((_, i) => `T${i}`);
const offsets = new Array(shapes.length - 1);
offsets[0] = shapes[0][axis];
for (let i = 1; i < offsets.length; i++) {
offsets[i] = offsets[i - 1] + shapes[i][axis];
}
const channel = channels[axis];
const lastChannels = channels.slice(-2);
const allChannels = channels.join();
let getValueSnippet = `if (${channel} < ${offsets[0]}) {
return getChannel(
getT0(${allChannels}), vec2(${lastChannels.join()}));
}`;
for (let i = 1; i < offsets.length; i++) {
const shift = offsets[i - 1];
// Note: the >= comparison below may seem unnecessary given the check
// above but is needed to workaround branch execution issues on some
// devices. It makes all the conditions exclusive without relying on
// execution order.
getValueSnippet += `
if (${channel} < ${offsets[i]} && ${channel} >= ${offsets[i - 1]}) {
return getChannel(
getT${i}(${shiftedChannels(channels, channel, shift)}),
vec2(${shiftedChannels(lastChannels, channel, shift)}));
}`;
}
const lastIndex = offsets.length;
const shift = offsets[offsets.length - 1];
getValueSnippet += `
return getChannel(
getT${lastIndex}(${shiftedChannels(channels, channel, shift)}),
vec2(${shiftedChannels(lastChannels, channel, shift)}));`;
this.userCode = `
float getValue(${channels.map(x => 'int ' + x)}) {
${getValueSnippet}
}
void main() {
${dtype} coords = getOutputCoords();
vec4 result = vec4(getValue(${coords}), 0., 0., 0.);
${coords[rank - 1]} = ${coords[rank - 1]} + 1;
if (${coords[rank - 1]} < ${shape[rank - 1]}) {
result.g = getValue(${coords});
}
${coords[rank - 2]} = ${coords[rank - 2]} + 1;
if (${coords[rank - 2]} < ${shape[rank - 2]}) {
result.a = getValue(${coords});
}
${coords[rank - 1]} = ${coords[rank - 1]} - 1;
if (${coords[rank - 2]} < ${shape[rank - 2]} &&
${coords[rank - 1]} < ${shape[rank - 1]}) {
result.b = getValue(${coords});
}
setOutput(result);
}
`;
}
}
/**
* Return an expression for coordinates into a vector where a given channel
* will be offset by [shift].
*
* @param channels the channels to consider
* @param channel the channel we want shifted
* @param shift the amount to subtract from the channel.
*
* @returns a string of the form 'x, y-[shift], z' where any one channel can
* have the shift applied.
*/
function shiftedChannels(channels, channel, shift) {
const channelIdx = channels.indexOf(channel);
const res = channels.map((c, idx) => {
if (idx === channelIdx) {
return `${c} - ${shift}`;
}
else {
return c;
}
});
return res.join();
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function imag$1(args) {
const { inputs, backend } = args;
const { input } = inputs;
const inputData = backend.texData.get(input.dataId);
return identity({ inputs: { x: inputData.complexTensorInfos.imag }, backend });
}
const imagConfig$1 = {
kernelName: Imag,
backendName: 'webgl',
kernelFunc: imag$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function concatImpl(inputs, axis, backend) {
const dtype = inputs[0].dtype;
if (dtype === 'complex64') {
const reals = inputs.map((t) => real({ inputs: { input: t }, backend }));
const imags = inputs.map((t) => imag$1({ inputs: { input: t }, backend }));
const realConcated = concatImpl(reals, axis, backend);
const imagConcated = concatImpl(imags, axis, backend);
const result = complex({ inputs: { real: realConcated, imag: imagConcated }, backend });
reals.forEach(r => backend.disposeIntermediateTensorInfo(r));
imags.forEach(i => backend.disposeIntermediateTensorInfo(i));
backend.disposeIntermediateTensorInfo(realConcated);
backend.disposeIntermediateTensorInfo(imagConcated);
return result;
}
let runOnCpu = backend.shouldExecuteOnCPU(inputs);
// Run on cpu if dtype is string. For string, the backend represents it
// as Uint8Array[], where each Uint8Array is a character. Given that the
// computation is only on the outer array, uploading the whole data onto
// gpu is wasteful. Also, currently webgl doesn't have a design to
// upload and retrieve Uint8Array[] between cpu and gpu. Therefore, we
// just run the kernel on cpu if dtype is string.
if (dtype === 'string') {
runOnCpu = true;
}
if (runOnCpu) {
// Any concat of n-dimensional tensors across any axis can be reduced to
// a concatenation of two-dimensional tensors across the axis 1 by first
// partitioning the axes of the original tensors into those less than the
// axis to be concatenated and the rest. Then reshape the tensors
// into a two-dimensional tensor by collapsing these two sets of axes and
// concatenate the resulting matrices across the axis 1, finally reshaping
// the result to have the proper shape.
const tensors2D = inputs.map(t => {
const innerSize = sizeFromShape(t.shape.slice(axis));
const shape = [-1, innerSize];
return reshape$1({ inputs: { x: t }, backend, attrs: { shape } });
});
const inputsValShapes = tensors2D.map(t => {
return { vals: backend.readSync(t.dataId), shape: t.shape };
});
// Concats 2d tensors along axis=1.
const outShape = computeOutShape$1(tensors2D.map(t => t.shape), 1 /* axis */);
const simplyConcat = tensors2D[0].shape[0] === 1;
const outVals = concatImplCPU(inputsValShapes, outShape, dtype, simplyConcat);
const finalOutShape = computeOutShape$1(inputs.map(t => t.shape), axis);
const outInfo = backend.makeTensorInfo(finalOutShape, dtype, outVals);
tensors2D.forEach(t => backend.disposeIntermediateTensorInfo(t));
return outInfo;
}
// Keep only non-empty tensors (ignore tensors with 0 in their shape).
const $inputs = inputs.filter(t => sizeFromShape(t.shape) > 0);
const shouldPack = env().getBool('WEBGL_PACK_ARRAY_OPERATIONS') &&
$inputs[0].shape.length > 1;
if ($inputs.length === 1) {
// Clone tensor.
const program = shouldPack ?
new UnaryOpProgram(inputs[0].shape, CLONE) :
new UnaryOpPackedProgram(inputs[0].shape, CLONE);
return backend.runWebGLProgram(program, inputs, dtype);
}
const maxTexturesInShader = env().getNumber('WEBGL_MAX_TEXTURES_IN_SHADER');
if ($inputs.length > maxTexturesInShader) {
const reducedInputs = [];
for (let i = 0; i < $inputs.length; i += maxTexturesInShader) {
const subArray = $inputs.slice(i, i + maxTexturesInShader);
reducedInputs.push(concatImpl(subArray, axis, backend));
}
const result = concatImpl(reducedInputs, axis, backend);
for (const i of reducedInputs) {
backend.disposeIntermediateTensorInfo(i);
}
return result;
}
if (shouldPack) {
const program = new ConcatPackedProgram($inputs.map(t => t.shape), axis);
return backend.runWebGLProgram(program, $inputs, dtype);
}
const { tensors2D, outShape } = computeTensors2D($inputs, axis, backend);
const program = new ConcatProgram(tensors2D.map(t => t.shape));
const result = backend.runWebGLProgram(program, tensors2D, dtype);
tensors2D.forEach(r => backend.disposeIntermediateTensorInfo(r));
const reshapedResult = reshape$1({ inputs: { x: result }, attrs: { shape: outShape }, backend });
backend.disposeIntermediateTensorInfo(result);
return reshapedResult;
}
function computeTensors2D(inputs, axis, backend) {
// Any concat of n-dimensional tensors across any axis can be reduced to
// a concatenation of two-dimensional tensors across the axis 1 by first
// partitioning the axes of the original tensors into those less than the
// axis to be concatenated and the rest. Then reshape the tensors
// into a two-dimensional tensor by collapsing these two sets of axes and
// concatenate the resulting matrices across the axis 1, finally reshaping
// the result to have the proper shape.
const outShape = computeOutShape$1(inputs.map(t => t.shape), axis);
const tensors2D = inputs.map(x => reshape$1({
inputs: { x },
attrs: { shape: [-1, sizeFromShape(x.shape.slice(axis))] },
backend
}));
return { tensors2D, outShape };
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function concat$1(args) {
const { inputs, backend, attrs } = args;
const { axis } = attrs;
const $axis = parseAxisParam(axis, inputs[0].shape)[0];
const shapes = inputs.map(t => t.shape);
assertParamsConsistent(shapes, $axis);
const outShape = computeOutShape$1(inputs.map(t => t.shape), $axis);
if (sizeFromShape(outShape) === 0) {
return backend.makeTensorInfo(outShape, inputs[0].dtype, []);
}
// Keep only non-empty tensors (ignore tensors with 0 in their shape).
const $inputs = inputs.filter(t => sizeFromShape(t.shape) > 0);
if ($inputs.length === 1) {
return identity({ inputs: { x: $inputs[0] }, backend });
}
return concatImpl($inputs, $axis, backend);
}
const concatConfig$1 = {
kernelName: Concat,
backendName: 'webgl',
kernelFunc: concat$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Conv2DProgram {
constructor(convInfo, addBias = false, activation = null, hasPreluActivationWeights = false, hasLeakyreluAlpha = false) {
this.variableNames = ['x', 'W'];
this.outputShape = convInfo.outShape;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const inputDepthNearestVec4 = Math.floor(convInfo.inChannels / 4) * 4;
const inputDepthVec4Remainder = convInfo.inChannels % 4;
const isChannelsLast = convInfo.dataFormat === 'channelsLast';
const rowDim = isChannelsLast ? 1 : 2;
const colDim = isChannelsLast ? 2 : 3;
const channelDim = isChannelsLast ? 3 : 1;
let activationSnippet = '', applyActivationSnippet = '';
if (activation) {
if (hasPreluActivationWeights) {
activationSnippet = `float activation(float a) {
float b = getPreluActivationWeightsAtOutCoords();
${activation}
}`;
}
else if (hasLeakyreluAlpha) {
activationSnippet = `float activation(float a) {
float b = getLeakyreluAlphaAtOutCoords();
${activation}
}`;
}
else {
activationSnippet = `
float activation(float x) {
${activation}
}
`;
}
applyActivationSnippet = `result = activation(result);`;
}
const addBiasSnippet = addBias ? 'result += getBiasAtOutCoords();' : '';
if (addBias) {
this.variableNames.push('bias');
}
if (hasPreluActivationWeights) {
this.variableNames.push('preluActivationWeights');
}
if (hasLeakyreluAlpha) {
this.variableNames.push('leakyreluAlpha');
}
this.userCode = `
${activationSnippet}
const ivec2 strides = ivec2(${strideHeight}, ${strideWidth});
const ivec2 pads = ivec2(${padTop}, ${padLeft});
void main() {
ivec4 coords = getOutputCoords();
int batch = coords[0];
int d2 = coords[${channelDim}];
ivec2 xRCCorner =
ivec2(coords[${rowDim}], coords[${colDim}]) * strides - pads;
int xRCorner = xRCCorner.x;
int xCCorner = xRCCorner.y;
// Convolve x(?, ?, d1) with w(:, :, d1, d2) to get y(yR, yC, d2).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
for (int wR = 0; wR < ${filterHeight}; wR++) {
int xR = xRCorner + wR * ${dilationHeight};
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int wC = 0; wC < ${filterWidth}; wC++) {
int xC = xCCorner + wC * ${dilationWidth};
if (xC < 0 || xC >= ${convInfo.inWidth}) {
continue;
}
for (int d1 = 0; d1 < ${inputDepthNearestVec4}; d1 += 4) {
vec4 wValues = vec4(
getW(wR, wC, d1, d2),
getW(wR, wC, d1 + 1, d2),
getW(wR, wC, d1 + 2, d2),
getW(wR, wC, d1 + 3, d2)
);
if (${isChannelsLast}) {
vec4 xValues = vec4(
getX(batch, xR, xC, d1),
getX(batch, xR, xC, d1 + 1),
getX(batch, xR, xC, d1 + 2),
getX(batch, xR, xC, d1 + 3)
);
dotProd += dot(xValues, wValues);
} else {
vec4 xValues = vec4(
getX(batch, d1, xR, xC),
getX(batch, d1 + 1, xR, xC),
getX(batch, d1 + 2, xR, xC),
getX(batch, d1 + 3, xR, xC)
);
dotProd += dot(xValues, wValues);
}
}
if (${inputDepthVec4Remainder === 1}) {
if (${isChannelsLast}) {
dotProd +=
getX(batch, xR, xC, ${inputDepthNearestVec4}) *
getW(wR, wC, ${inputDepthNearestVec4}, d2);
} else {
dotProd +=
getX(batch, ${inputDepthNearestVec4}, xR, xC) *
getW(wR, wC, ${inputDepthNearestVec4}, d2);
}
} else if (${inputDepthVec4Remainder === 2}) {
vec2 wValues = vec2(
getW(wR, wC, ${inputDepthNearestVec4}, d2),
getW(wR, wC, ${inputDepthNearestVec4} + 1, d2)
);
if (${isChannelsLast}) {
vec2 xValues = vec2(
getX(batch, xR, xC, ${inputDepthNearestVec4}),
getX(batch, xR, xC, ${inputDepthNearestVec4} + 1)
);
dotProd += dot(xValues, wValues);
} else {
vec2 xValues = vec2(
getX(batch, ${inputDepthNearestVec4}, xR, xC),
getX(batch, ${inputDepthNearestVec4} + 1, xR, xC)
);
dotProd += dot(xValues, wValues);
}
} else if (${inputDepthVec4Remainder === 3}) {
vec3 wValues = vec3(
getW(wR, wC, ${inputDepthNearestVec4}, d2),
getW(wR, wC, ${inputDepthNearestVec4} + 1, d2),
getW(wR, wC, ${inputDepthNearestVec4} + 2, d2)
);
if (${isChannelsLast}) {
vec3 xValues = vec3(
getX(batch, xR, xC, ${inputDepthNearestVec4}),
getX(batch, xR, xC, ${inputDepthNearestVec4} + 1),
getX(batch, xR, xC, ${inputDepthNearestVec4} + 2)
);
dotProd += dot(xValues, wValues);
} else {
vec3 xValues = vec3(
getX(batch, ${inputDepthNearestVec4}, xR, xC),
getX(batch, ${inputDepthNearestVec4} + 1, xR, xC),
getX(batch, ${inputDepthNearestVec4} + 2, xR, xC)
);
dotProd += dot(xValues, wValues);
}
}
}
}
float result = dotProd;
${addBiasSnippet}
${applyActivationSnippet}
setOutput(result);
}
`;
}
}
class Conv3DProgram {
constructor(convInfo) {
this.variableNames = ['x', 'W'];
this.outputShape = convInfo.outShape;
const padFront = convInfo.padInfo.front;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const filterDepth = convInfo.filterDepth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const inputDepthNearestVec4 = Math.floor(convInfo.inChannels / 4) * 4;
const inputDepthVec4Remainder = convInfo.inChannels % 4;
this.userCode = `
const ivec3 strides = ivec3(${strideDepth}, ${strideHeight}, ${strideWidth});
const ivec3 pads = ivec3(${padFront}, ${padTop}, ${padLeft});
void main() {
ivec5 coords = getOutputCoords();
int batch = coords.x;
int d2 = coords.u;
ivec3 xFRCCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
int xFCorner = xFRCCorner.x;
int xRCorner = xFRCCorner.y;
int xCCorner = xFRCCorner.z;
// Convolve x(?, ?, ?, d1) with w(:, :, :, d1, d2) to get
// y(yF, yR, yC, d2). ? = to be determined. : = across all
// values in that axis.
float dotProd = 0.0;
for (int wF = 0; wF < ${filterDepth}; wF++) {
int xF = xFCorner + wF * ${dilationDepth};
if (xF < 0 || xF >= ${convInfo.inDepth}) {
continue;
}
for (int wR = 0; wR < ${filterHeight}; wR++) {
int xR = xRCorner + wR * ${dilationHeight};
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int wC = 0; wC < ${filterWidth}; wC++) {
int xC = xCCorner + wC * ${dilationWidth};
if (xC < 0 || xC >= ${convInfo.inWidth}) {
continue;
}
for (int d1 = 0; d1 < ${inputDepthNearestVec4}; d1 += 4) {
vec4 xValues = vec4(
getX(batch, xF, xR, xC, d1),
getX(batch, xF, xR, xC, d1 + 1),
getX(batch, xF, xR, xC, d1 + 2),
getX(batch, xF, xR, xC, d1 + 3)
);
vec4 wValues = vec4(
getW(wF, wR, wC, d1, d2),
getW(wF, wR, wC, d1 + 1, d2),
getW(wF, wR, wC, d1 + 2, d2),
getW(wF, wR, wC, d1 + 3, d2)
);
dotProd += dot(xValues, wValues);
}
if (${inputDepthVec4Remainder === 1}) {
dotProd +=
getX(batch, xF, xR, xC, ${inputDepthNearestVec4}) *
getW(wF, wR, wC, ${inputDepthNearestVec4}, d2);
} else if (${inputDepthVec4Remainder === 2}) {
vec2 xValues = vec2(
getX(batch, xF, xR, xC, ${inputDepthNearestVec4}),
getX(batch, xF, xR, xC, ${inputDepthNearestVec4} + 1)
);
vec2 wValues = vec2(
getW(wF, wR, wC, ${inputDepthNearestVec4}, d2),
getW(wF, wR, wC, ${inputDepthNearestVec4} + 1, d2)
);
dotProd += dot(xValues, wValues);
} else if (${inputDepthVec4Remainder === 3}) {
vec3 xValues = vec3(
getX(batch, xF, xR, xC, ${inputDepthNearestVec4}),
getX(batch, xF, xR, xC, ${inputDepthNearestVec4} + 1),
getX(batch, xF, xR, xC, ${inputDepthNearestVec4} + 2)
);
vec3 wValues = vec3(
getW(wF, wR, wC, ${inputDepthNearestVec4}, d2),
getW(wF, wR, wC, ${inputDepthNearestVec4} + 1, d2),
getW(wF, wR, wC, ${inputDepthNearestVec4} + 2, d2)
);
dotProd += dot(xValues, wValues);
}
}
}
}
setOutput(dotProd);
}
`;
}
}
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Conv2DPackedProgram {
constructor(convInfo, addBias = false, activation = null, hasPreluActivation = false, hasLeakyReluAlpha = false) {
this.variableNames = ['x', 'W'];
this.packedInputs = true;
this.packedOutput = true;
this.customUniforms = [
{ name: 'pads', type: 'ivec2' },
{ name: 'strides', type: 'ivec2' },
{ name: 'dilations', type: 'ivec2' },
{ name: 'inDims', type: 'ivec2' },
];
this.outputShape = convInfo.outShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
const padLeft = convInfo.padInfo.left;
const strideWidth = convInfo.strideWidth;
const dilationWidth = convInfo.dilationWidth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const texelsAcross = filterWidth;
let mainLoop = `
int xR; int xC; int xCOffset;
vec4 wTexel; vec4 previous; vec4 final;`;
for (let c = 0; c < filterWidth; c++) {
mainLoop += `
vec4 xTexelC${c * 2};
int xTexelC${c * 2}Ready;
vec4 xTexelC${c * 2 + 1};
int xTexelC${c * 2 + 1}Ready;
vec4 xC${c};`;
}
/**
* This vectorized implementation works by gathering the values needed for
* each output channel's dot product into vec4's and then multiplying them
* all together (this happens in the final double for-loop below). Most of
* the main loop consists of constructing these vec4's with the minimum
* number of texture2D calls, which means making use of all four returned
* values from a texture2D call at once.
*/
mainLoop += `
for (int r = 0; r < ${filterHeight}; r++) {
for (int d1 = 0; d1 < ${convInfo.inChannels}; d1 += 2) {
`;
for (let c = 0; c < filterWidth; c++) {
mainLoop += `
xTexelC${c * 2} = vec4(0.0);
xTexelC${c * 2}Ready = 0;
xTexelC${c * 2 + 1} = vec4(0.0);
xTexelC${c * 2 + 1}Ready = 0;
xC${c} = vec4(0.0);`;
}
mainLoop += `
xR = xRCorner + r * dilations[0];
if (xR >=0 && xR < inDims[0]) {
`;
for (let texelC = 0; texelC < (texelsAcross + 1) / 2; texelC++) {
const colIndex = texelC * 2;
mainLoop += `
xC = xCCorner + ${colIndex * dilationWidth};
`;
if (strideWidth === 1) {
if (colIndex < filterWidth) {
// If padding is odd, the outer texels have to be composed.
if (padLeft % 2 === 1) {
// TODO: Ensure vec4 previous does not result in redundant sample,
// and avoid setting xTexelRC's that exceed the boundary in the
// first place rather than resetting them to vec4(0)).
// To compute xCOffset:
// - If padding is odd, we must add 1 to ensure we ask for an
// even-numbered row.
// - We subtract 2 to access the previous texel.
mainLoop += `
xCOffset = xC + 1;
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
`;
// This texel has been read in previous iteration if the dilation
// is 1.
if (dilationWidth === 1 && colIndex > 0) {
mainLoop += `
xC${colIndex} = vec4(xTexelC${colIndex - 2}.zw, xTexelC${colIndex}.xy);
`;
}
else {
mainLoop += `
xCOffset = xC + 1 - 2;
if (xCOffset >= 0 && xCOffset < inDims[1]) {
previous = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
previous.zw = vec2(0.0);
}
xC${colIndex} = vec4(previous.zw, xTexelC${colIndex}.xy);
} else {
xC${colIndex} = vec4(0.0, 0.0, xTexelC${colIndex}.xy);
}
`;
}
}
else {
// Padding is even, so xRC corresponds to a single texel.
mainLoop += `
if (xC >= 0 && xC < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xC, d1);
if (xC + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
xC${colIndex} = xTexelC${colIndex};
`;
}
if (colIndex + 1 < filterWidth) {
// If dilation is even, the second entry should match the first
// (either both are composed or both are single samples). But if
// dilation is odd, then the second entry should be the opposite
// of the first (if the first is composed, the second is a single
// sample, and vice versa.)
const nextTexelOffset = padLeft % 2 === 0 ?
nearestLargerEven(dilationWidth) :
dilationWidth;
if ((dilationWidth % 2 === 0 && padLeft % 2 === 1) ||
(dilationWidth % 2 !== 0 && padLeft % 2 !== 1)) {
mainLoop += `
xCOffset = xC + imod(pads[1], 2) + ${nextTexelOffset};
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.0);
}
xTexelC${colIndex + 1}Ready = 1;
}
`;
// If dilation > 1 then the xRC's will not be able to share any
// values, so each xRC will require two unique calls to getX.
if (dilationWidth > 1) {
mainLoop += `
xCOffset -= 2;
if (xCOffset >= 0 && xCOffset < inDims[1]) {
previous = getX(batch, xR, xCOffset, d1);
xC${colIndex + 1} = vec4(previous.zw, xTexelC${colIndex + 1}.xy);
} else {
xC${colIndex + 1} = vec4(0.0, 0.0, xTexelC${colIndex + 1}.xy);
}
`;
}
else {
mainLoop += `
xC${colIndex + 1} = vec4(xTexelC${colIndex}.zw, xTexelC${colIndex + 1}.xy);
`;
}
}
else {
// If dilation is 1 and padding is odd, we have already read the
// texel when constructing the previous x value. Here we can
// simply skip the texture read.
if (nextTexelOffset === 1) {
mainLoop += `
xC${colIndex + 1} = xTexelC${colIndex};
`;
}
else {
mainLoop += `
xCOffset = xC + ${nextTexelOffset};
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xCOffset, d1);
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.0);
}
xTexelC${colIndex + 1}Ready = 1;
}
xC${colIndex + 1} = xTexelC${colIndex + 1};
`;
}
}
}
}
}
else { // stride === 2
if (colIndex < filterWidth) {
// Depending on whether padLeft is even or odd, we want either the
// xy or zw channels from X texels for xC${colIndex}. If padLeft is
// even, xC${colIndex +1} is simply the zw channels of texels we've
// already sampled. But if padLeft is odd, xC{$c + 1}.zw will
// need to come from the xy channels of a new texel, hence the `
// vec4
// final` initialized below.
if (padLeft % 2 === 1) {
mainLoop += `
xCOffset = xC + 1 - strides[1];
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xC + 1, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xC + 2 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.0);
}
xTexelC${colIndex + 1}Ready = 1;
}
xC${colIndex} = vec4(xTexelC${colIndex}.zw, xTexelC${colIndex + 1}.zw);
`;
if (colIndex + 1 < filterWidth) {
mainLoop += `
final = vec4(0.0);
xCOffset = xC + 1 + strides[1];
if(xCOffset >= 0 && xCOffset < inDims[1]) {
final = getX(batch, xR, xCOffset, d1);
}
xC${colIndex + 1} = vec4(xTexelC${colIndex + 1}.xy, final.xy);
`;
}
}
else {
mainLoop += `
if(xC >= 0 && xC < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xC, d1);
if (xC + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
xCOffset = xC + strides[1];
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xCOffset, d1);
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.);
}
xTexelC${colIndex + 1}Ready = 1;
}
xC${colIndex} = vec4(
xTexelC${colIndex}.xy, xTexelC${colIndex + 1}.xy);
`;
if (colIndex + 1 < filterWidth) {
mainLoop += `
xC${colIndex + 1} = vec4(xTexelC${colIndex}.zw, xTexelC${colIndex + 1}.zw);
`;
}
}
}
}
// localize the dotProd accumulation within the loop, the theory is for
// GPU with limited cache, accumulate sum across large amount of
// veriables will cause lots of cache misses. (i.e. 5x5 filter will have
// 50 variables)
if (colIndex < filterWidth) {
mainLoop += `
wTexel = getW(r, ${colIndex}, d1, d2);
dotProd += xC${colIndex}.xxzz * vec4(wTexel.xy, wTexel.xy);
if(d1 + 1 < ${convInfo.inChannels}) {
dotProd += xC${colIndex}.yyww * vec4(wTexel.zw, wTexel.zw);
}
`;
if (colIndex + 1 < filterWidth) {
mainLoop += `
wTexel = getW(r, ${colIndex + 1}, d1, d2);
dotProd += xC${colIndex + 1}.xxzz * vec4(wTexel.xy, wTexel.xy);
if(d1 + 1 < ${convInfo.inChannels}) {
dotProd += xC${colIndex + 1}.yyww * vec4(wTexel.zw, wTexel.zw);
}
`;
}
}
}
mainLoop += `
}
`;
mainLoop += `
}
`;
mainLoop += `
}
`;
let activationSnippet = '', applyActivationSnippet = '';
if (activation) {
if (hasPreluActivation) {
activationSnippet = `vec4 activation(vec4 a) {
vec4 b = getPreluActivationWeightsAtOutCoords();
${activation}
}`;
}
else if (hasLeakyReluAlpha) {
activationSnippet = `vec4 activation(vec4 a) {
vec4 b = getLeakyreluAlphaAtOutCoords();
${activation}
}`;
}
else {
activationSnippet = `vec4 activation(vec4 x) {
${activation}
}`;
}
applyActivationSnippet = `result = activation(result);`;
}
const addBiasSnippet = addBias ? 'result += getBiasAtOutCoords();' : '';
if (addBias) {
this.variableNames.push('bias');
}
if (hasPreluActivation) {
this.variableNames.push('preluActivationWeights');
}
if (hasLeakyReluAlpha) {
this.variableNames.push('leakyreluAlpha');
}
this.userCode = `
${activationSnippet}
void main() {
ivec4 coords = getOutputCoords();
int batch = coords.x;
ivec2 xRCCorner = coords.yz * strides - pads;
int d2 = coords.w;
int xRCorner = xRCCorner.x;
int xCCorner = xRCCorner.y;
//intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
vec4 dotProd = vec4(0.000000000000001);
${mainLoop}
vec4 result = dotProd - vec4(0.000000000000001);
${addBiasSnippet}
${applyActivationSnippet}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Im2ColPackedProgram {
constructor(outputShape, convInfo) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
this.customUniforms = [
{ name: 'inputShape', type: 'ivec4' },
{ name: 'pad', type: 'ivec2' },
{ name: 'stride', type: 'ivec2' },
{ name: 'dilation', type: 'ivec2' },
{ name: 'inChannels', type: 'int' },
{ name: 'itemsPerBlockRow', type: 'int' },
{ name: 'outWidth', type: 'int' },
];
this.outputShape = outputShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
const { dataFormat } = convInfo;
const glsl = getGlslDifferences();
const isChannelsLast = dataFormat === 'channelsLast';
const rowDim = isChannelsLast ? 1 : 2;
const colDim = isChannelsLast ? 2 : 3;
const boundsCheckingSnippet = this.enableShapeUniforms ?
'if(blockIndex < outShape[2] && pos < outShape[1]) {' :
`if(blockIndex < ${outputShape[2]} && pos < ${outputShape[1]}) {`;
let unrolled = ``;
for (let row = 0; row <= 1; row++) {
for (let col = 0; col <= 1; col++) {
unrolled += `
blockIndex = rc.z + ${col};
pos = rc.y + ${row};
${boundsCheckingSnippet}
offsetY = int(blockIndex / outWidth) * stride[0] - pad[0];
d0 = offsetY + dilation[0] * (pos / itemsPerBlockRow);
if(d0 < inputShape[${rowDim}] && d0 >= 0) {
// Use custom imod instead mod. On Intel GPU, mod may generate
// unexpected value.
// https://github.com/tensorflow/tfjs/issues/5447
offsetX = imod(blockIndex, outWidth) * stride[1] - pad[1];
d1 = offsetX + dilation[1] * (imod(pos, itemsPerBlockRow) /
inChannels);
if(d1 < inputShape[${colDim}] && d1 >= 0) {
ch = imod(pos, inChannels);
if (${isChannelsLast}) {
innerDims = vec2(d1, ch);
result[${row * 2 + col}] = getChannel(
getA(rc.x, d0, int(innerDims.x),
int(innerDims.y)), innerDims);
} else {
innerDims = vec2(d0, d1);
result[${row * 2 + col}] = getChannel(
getA(rc.x, ch, int(innerDims.x),
int(innerDims.y)), innerDims);
}
}
}
}
`;
}
}
this.userCode = `
void main() {
ivec3 rc = getOutputCoords();
vec4 result = vec4(0);
int blockIndex, pos, offsetY, d0, offsetX, d1, ch;
vec2 innerDims;
${unrolled}
${glsl.output} = result;
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Both conv2dByMatMul and conv2dWithIm2Row fuse height and width into one
// dimension to compute batchMatMul, so bias and activation weights are also
// supposed to fuse the two dimensions into one.
//
// This function computes the target shape for fusing height and width
// dimensions. Returning null means the shape is already compatible.
//
// Even though the bias is not supposed to be a 3-D or a 4-D (including
// batch) tensor and PReLU activiation weights is not supposed to be a 4-D
// tensor, we still need to support them, because we haven't disabled
// them for NHWC format.
// https://github.com/tensorflow/tfjs/blob/b53bd47e880367ae57493f0ea628abaf08db2d5d/tfjs-core/src/ops/fused/conv2d.ts#L181-L196
function getShapeForBatchMatMul(shape, isChannelsLast) {
const length = shape.length;
if (length >= 3) {
return isChannelsLast ?
[
...shape.slice(0, -3) /* batch */,
shape[length - 3] * shape[length - 2] /* height * width */,
shape[length - 1] /* channel */
] :
[
...shape.slice(0, -3) /* batch */, shape[length - 3] /* channel */,
shape[length - 2] * shape[length - 1] /* height * width */
];
}
else if (!isChannelsLast && length === 1 && shape[0] > 1) {
return [shape[0], 1];
}
else {
return null;
}
}
// For 1x1 kernels that iterate through every point in the input, convolution
// can be expressed as matrix multiplication (without need for memory
// remapping).
function conv2dByMatMul({ x, filter, convInfo, backend, bias = null, preluActivationWeights = null, leakyreluAlpha = 0, activation = null }) {
// Reshapes conv2D input to 2D tensors, uses matMul and then reshape the
// result from 2D to 4D.
const xShape = x.shape;
const xTexData = backend.texData.get(x.dataId);
const sharedMatMulDim = convInfo.inChannels;
const outerShapeX = xShape[0] * xShape[1] * xShape[2];
const outerShapeFilter = convInfo.outChannels;
const isChannelsLast = convInfo.dataFormat === 'channelsLast';
const transposeA = false;
const transposeB = false;
let out;
const intermediates = [];
if (preluActivationWeights != null) {
const targetShape = getShapeForBatchMatMul(preluActivationWeights.shape, isChannelsLast);
if (targetShape != null) {
preluActivationWeights = reshape$1({
inputs: { x: preluActivationWeights },
backend,
attrs: { shape: targetShape }
});
intermediates.push(preluActivationWeights);
}
}
if (bias != null) {
const targetShape = getShapeForBatchMatMul(bias.shape, isChannelsLast);
if (targetShape != null) {
bias = reshape$1({ inputs: { x: bias }, backend, attrs: { shape: targetShape } });
intermediates.push(bias);
}
}
// TODO: Once reduction ops are packed, batchMatMul will always be packed
// and we can remove this condition.
const batchMatMulWillBeUnpacked = (outerShapeX === 1 || outerShapeFilter === 1) &&
sharedMatMulDim > MATMUL_SHARED_DIM_THRESHOLD;
// The algorithm in the if condition assumes (1) the output will be packed,
// (2) x is packed, (3) x isChannelsLast, (4) x's packed texture is already
// on GPU, (5) col is odd, (6) the width, height and inChannels are the same
// for xTexData.shape and xShape.
const canOptimize = !batchMatMulWillBeUnpacked && xTexData.isPacked &&
isChannelsLast && xTexData.texture != null && xShape[2] % 2 !== 0 &&
arraysEqual(xTexData.shape.slice(-3), xShape.slice(-3));
if (canOptimize) {
// We avoid expensive packed 2x2 reshape by padding col count to next,
// even number. When col is odd, the result of packed batchMatMul is
// the same (has the same texture layout and and values in the texture) as
// it is for next even col. We make the odd-cols tensor to look like
// even-cols tensor before the operation and, after the batchMatMul,
// fix the even-cols result to have odd number of cols.
const targetShape = xShape[0] * xShape[1] * (xShape[2] + 1);
const xReshaped = {
dataId: x.dataId,
shape: [1, targetShape, convInfo.inChannels],
dtype: x.dtype
};
// xTexData.shape gets referenced from GPGPUBinary.inShapeInfos.
// Decrementing col count, after batchMatMul->...->compileProgram leads to
// invalid col count within the reference in GPGPUBinary.inShapeInfos.
// Alternative fix would be to provide a copy to GPGPUBinary.inShapeInfos
// in compileProgram method, but that would affect compilation of all
// programs - instead, provide a copy here, with even col count, before
// calling batchMatMul->...->compileProgram and after that, the original
// xTexData.shape is restored.
const originalXTexDataShape = xTexData.shape;
xTexData.shape = xTexData.shape.slice();
xTexData.shape[xTexData.shape.length - 2]++;
assert$1(isReshapeFree(xTexData.shape, xReshaped.shape), () => `packed reshape ${xTexData.shape} to ${xReshaped.shape} isn't free`);
const filterReshaped = reshape$1({
inputs: { x: filter },
backend,
attrs: { shape: [1, convInfo.inChannels, convInfo.outChannels] }
});
intermediates.push(filterReshaped);
const pointwiseConv = batchMatMulImpl({
a: xReshaped,
b: filterReshaped,
backend,
transposeA,
transposeB,
bias,
activation,
preluActivationWeights,
leakyreluAlpha
});
const pointwiseConvTexData = backend.texData.get(pointwiseConv.dataId);
assert$1(pointwiseConvTexData.isPacked, () => 'batchMatMul result is expected to be packed');
// Restore the input shape to original.
xTexData.shape = originalXTexDataShape;
// Set the output shape - there is no need for expensive reshape as data
// layout is already correct.
pointwiseConvTexData.shape = convInfo.outShape;
out = identity({ inputs: { x: pointwiseConv }, backend });
out.shape = convInfo.outShape;
intermediates.push(pointwiseConv);
}
else {
const numCols = convInfo.outHeight * convInfo.outWidth;
const xReshaped = reshape$1({
inputs: { x },
backend,
attrs: {
shape: isChannelsLast ?
[convInfo.batchSize, numCols, convInfo.inChannels] :
[convInfo.batchSize, convInfo.inChannels, numCols]
}
});
const filterReshaped = reshape$1({
inputs: { x: filter },
backend,
attrs: { shape: [1, convInfo.inChannels, convInfo.outChannels] }
});
const result = batchMatMulImpl({
a: isChannelsLast ? xReshaped : filterReshaped,
b: isChannelsLast ? filterReshaped : xReshaped,
transposeA: !isChannelsLast,
transposeB,
backend,
bias,
activation,
preluActivationWeights,
leakyreluAlpha
});
out = reshape$1({ inputs: { x: result }, backend, attrs: { shape: convInfo.outShape } });
intermediates.push(xReshaped);
intermediates.push(filterReshaped);
intermediates.push(result);
}
for (const i of intermediates) {
backend.disposeIntermediateTensorInfo(i);
}
return out;
}
// Implements the im2row algorithm as outlined in "High Performance
// Convolutional Neural Networks for Document Processing" (Suvisoft, 2006)
function conv2dWithIm2Row({ x, filter, convInfo, backend, bias = null, preluActivationWeights = null, leakyreluAlpha = 0, activation = null }) {
// Rearranges conv2d input so each block to be convolved over forms the
// column of a new matrix with shape [filterWidth * filterHeight *
// inChannels, outHeight * outWidth]. The filter is also rearranged so each
// output channel forms a row of a new matrix with shape [outChannels,
// filterWidth * filterHeight * inChannels]. The convolution is then
// computed by multiplying these matrices and reshaping the result.
const { filterWidth, filterHeight, inChannels, outWidth, outHeight, dataFormat } = convInfo;
const isChannelsLast = dataFormat === 'channelsLast';
const sharedDim = filterWidth * filterHeight * inChannels;
const numCols = outHeight * outWidth;
const x2ColShape = [convInfo.batchSize, sharedDim, numCols];
const transposeA = true;
const transposeB = false;
const intermediates = [];
if (preluActivationWeights != null) {
const targetShape = getShapeForBatchMatMul(preluActivationWeights.shape, isChannelsLast);
if (targetShape != null) {
preluActivationWeights = reshape$1({
inputs: { x: preluActivationWeights },
backend,
attrs: { shape: targetShape }
});
intermediates.push(preluActivationWeights);
}
}
if (bias != null) {
const targetShape = getShapeForBatchMatMul(bias.shape, isChannelsLast);
if (targetShape != null) {
bias = reshape$1({ inputs: { x: bias }, backend, attrs: { shape: targetShape } });
intermediates.push(bias);
}
}
const w2Row = reshape$1({
inputs: { x: filter },
backend,
attrs: { shape: [1, sharedDim, sizeFromShape(filter.shape) / sharedDim] }
});
intermediates.push(w2Row);
const im2ColProgram = new Im2ColPackedProgram(x2ColShape, convInfo);
const customValues = [
x.shape, [convInfo.padInfo.top, convInfo.padInfo.left],
[convInfo.strideHeight, convInfo.strideWidth],
[convInfo.dilationHeight, convInfo.dilationWidth], [convInfo.inChannels],
[convInfo.filterWidth * convInfo.inChannels], [convInfo.outWidth]
];
const im2Col = backend.runWebGLProgram(im2ColProgram, [x], 'float32', customValues);
const im2ColReshaped = reshape$1({ inputs: { x: im2Col }, backend, attrs: { shape: x2ColShape } });
intermediates.push(im2Col);
intermediates.push(im2ColReshaped);
const hasBias = bias != null;
const hasPreluActivationWeights = preluActivationWeights != null;
const hasLeakyreluAlpha = activation === 'leakyrelu';
const fusedActivation = activation ? mapActivationToShaderProgram(activation, true) : null;
const matmulProgram = new MatMulPackedProgram(isChannelsLast ? im2ColReshaped.shape :
w2Row.shape, isChannelsLast ? w2Row.shape :
im2ColReshaped.shape, isChannelsLast ? [convInfo.batchSize, numCols, convInfo.outChannels] :
[convInfo.batchSize, convInfo.outChannels, numCols], transposeA, transposeB, hasBias, fusedActivation, hasPreluActivationWeights, hasLeakyreluAlpha);
const inputs = isChannelsLast ? [im2ColReshaped, w2Row] : [w2Row, im2ColReshaped];
if (bias) {
inputs.push(bias);
}
if (hasPreluActivationWeights) {
inputs.push(preluActivationWeights);
}
if (hasLeakyreluAlpha) {
const $leakyreluAlpha = backend.makeTensorInfo([], 'float32', createScalarValue(leakyreluAlpha, 'float32'));
inputs.push($leakyreluAlpha);
intermediates.push($leakyreluAlpha);
}
const product = backend.runWebGLProgram(matmulProgram, inputs, 'float32');
const out = reshape$1({ inputs: { x: product }, backend, attrs: { shape: convInfo.outShape } });
intermediates.push(product);
for (const i of intermediates) {
backend.disposeIntermediateTensorInfo(i);
}
return out;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv2d(args) {
const { inputs, backend, attrs } = args;
const { x, filter } = inputs;
const { strides, pad, dataFormat, dilations, dimRoundingMode } = attrs;
const $dataFormat = convertConv2DDataFormat(dataFormat);
const convInfo = computeConv2DInfo(x.shape, filter.shape, strides, dilations, pad, dimRoundingMode, false /* depthwise */, $dataFormat);
let out;
if (convInfo.filterHeight === 1 && convInfo.filterWidth === 1 &&
convInfo.dilationHeight === 1 && convInfo.dilationWidth === 1 &&
convInfo.strideHeight === 1 && convInfo.strideWidth === 1 &&
(convInfo.padInfo.type === 'SAME' || convInfo.padInfo.type === 'VALID')) {
out = conv2dByMatMul({ x, filter, convInfo, backend });
}
else if (convInfo.strideWidth <= 2 && $dataFormat === 'channelsLast'
&& env().getBool('WEBGL_EXP_CONV')) {
const program = new Conv2DPackedProgram(convInfo);
const customValues = [
[convInfo.padInfo.top, convInfo.padInfo.left],
[convInfo.strideHeight, convInfo.strideWidth],
[convInfo.dilationHeight, convInfo.dilationWidth],
[convInfo.inHeight, convInfo.inWidth]
];
out =
backend.runWebGLProgram(program, [x, filter], 'float32', customValues);
}
else if (env().getBool('WEBGL_CONV_IM2COL')) {
out = conv2dWithIm2Row({ x, filter, convInfo, backend });
}
else {
const program = new Conv2DProgram(convInfo);
out = backend.runWebGLProgram(program, [x, filter], 'float32');
}
const outReshaped = reshape$1({ inputs: { x: out }, backend, attrs: { shape: convInfo.outShape } });
backend.disposeIntermediateTensorInfo(out);
return outReshaped;
}
const conv2DConfig$1 = {
kernelName: Conv2D,
backendName: 'webgl',
kernelFunc: conv2d,
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Conv2DDerFilterProgram {
constructor(convInfo) {
this.variableNames = ['x', 'dy'];
this.outputShape = convInfo.filterShape;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
const isChannelsLast = convInfo.dataFormat === 'channelsLast';
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int wR = coords.x;
int wC = coords.y;
int d1 = coords.z;
int d2 = coords.w;
// Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
for (int b = 0; b < ${convInfo.batchSize}; b++) {
for (int yR = 0; yR < ${convInfo.outHeight}; yR++) {
int xR = wR + yR * ${strideHeight} - ${padTop};
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int yC = 0; yC < ${convInfo.outWidth}; yC++) {
int xC = wC + yC * ${strideWidth} - ${padLeft};
if (xC < 0 || xC >= ${convInfo.inWidth}) {
continue;
}
${isChannelsLast ?
`float dyValue = getDy(b, yR, yC, d2);
float xValue = getX(b, xR, xC, d1);
dotProd += (xValue * dyValue);` :
`float dyValue = getDy(b, d2, yR, yC);
float xValue = getX(b, d1, xR, xC);
dotProd += (xValue * dyValue);`}
}
}
}
setOutput(dotProd);
}
`;
}
}
class Conv2DDerInputProgram {
constructor(convInfo) {
this.variableNames = ['dy', 'W'];
this.outputShape = convInfo.inShape;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const isChannelsLast = convInfo.dataFormat === 'channelsLast';
const padTop = filterHeight - 1 - convInfo.padInfo.top;
const padLeft = filterWidth - 1 - convInfo.padInfo.left;
const rowDim = isChannelsLast ? 1 : 2;
const colDim = isChannelsLast ? 2 : 3;
const channelDim = isChannelsLast ? 3 : 1;
this.userCode = `
const ivec2 pads = ivec2(${padTop}, ${padLeft});
void main() {
ivec4 coords = getOutputCoords();
int batch = coords[0];
int d1 = coords[${channelDim}];
ivec2 dyCorner = ivec2(coords[${rowDim}], coords[${colDim}]) - pads;
int dyRCorner = dyCorner.x;
int dyCCorner = dyCorner.y;
// Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
for (int wR = 0; wR < ${filterHeight}; wR++) {
float dyR = float(dyRCorner + wR) / ${strideHeight}.0;
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 || fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
int wRPerm = ${filterHeight} - 1 - wR;
for (int wC = 0; wC < ${filterWidth}; wC++) {
float dyC = float(dyCCorner + wC) / ${strideWidth}.0;
if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
fract(dyC) > 0.0) {
continue;
}
int idyC = int(dyC);
int wCPerm = ${filterWidth} - 1 - wC;
for (int d2 = 0; d2 < ${convInfo.outChannels}; d2++) {
if (${isChannelsLast}) {
float xValue = getDy(batch, idyR, idyC, d2);
float wValue = getW(wRPerm, wCPerm, d1, d2);
dotProd += xValue * wValue;
} else {
float xValue = getDy(batch, d2, idyR, idyC);
float wValue = getW(wRPerm, wCPerm, d1, d2);
dotProd += xValue * wValue;
}
}
}
}
setOutput(dotProd);
}
`;
}
}
class Conv3DDerFilterProgram {
constructor(convInfo) {
this.variableNames = ['x', 'dy'];
this.outputShape = convInfo.filterShape;
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const padFront = convInfo.padInfo.front;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
this.userCode = `
void main() {
ivec5 coords = getOutputCoords();
int wF = coords.x;
int wR = coords.y;
int wC = coords.z;
int d1 = coords.w;
int d2 = coords.u;
float dotProd = 0.0;
for (int b = 0; b < ${convInfo.batchSize}; b++) {
for (int yF = 0; yF < ${convInfo.outDepth}; yF++) {
int xF = wF + yF * ${strideDepth} - ${padFront};
if (xF < 0 || xF >= ${convInfo.inDepth}) {
continue;
}
for (int yR = 0; yR < ${convInfo.outHeight}; yR++) {
int xR = wR + yR * ${strideHeight} - ${padTop};
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int yC = 0; yC < ${convInfo.outWidth}; yC++) {
int xC = wC + yC * ${strideWidth} - ${padLeft};
if (xC < 0 || xC >= ${convInfo.inWidth}) {
continue;
}
float dyValue = getDy(b, yF, yR, yC, d2);
float xValue = getX(b, xF, xR, xC, d1);
dotProd += (xValue * dyValue);
}
}
}
}
setOutput(dotProd);
}
`;
}
}
class Conv3DDerInputProgram {
constructor(convInfo) {
this.variableNames = ['dy', 'W'];
this.outputShape = convInfo.inShape;
const filterDepth = convInfo.filterDepth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const padFront = filterDepth - 1 - convInfo.padInfo.front;
const padTop = filterHeight - 1 - convInfo.padInfo.top;
const padLeft = filterWidth - 1 - convInfo.padInfo.left;
this.userCode = `
const ivec3 pads = ivec3(${padFront}, ${padTop}, ${padLeft});
void main() {
ivec5 coords = getOutputCoords();
int batch = coords.x;
int d1 = coords.u;
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
int dyFCorner = dyCorner.x;
int dyRCorner = dyCorner.y;
int dyCCorner = dyCorner.z;
float dotProd = 0.0;
for (int wF = 0; wF < ${filterDepth}; wF++) {
float dyF = float(dyFCorner + wF) / ${strideDepth}.0;
if (dyF < 0.0 || dyF >= ${convInfo.outDepth}.0 || fract(dyF) > 0.0) {
continue;
}
int idyF = int(dyF);
int wFPerm = ${filterDepth} - 1 - wF;
for (int wR = 0; wR < ${filterHeight}; wR++) {
float dyR = float(dyRCorner + wR) / ${strideHeight}.0;
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 ||
fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
int wRPerm = ${filterHeight} - 1 - wR;
for (int wC = 0; wC < ${filterWidth}; wC++) {
float dyC = float(dyCCorner + wC) / ${strideWidth}.0;
if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
fract(dyC) > 0.0) {
continue;
}
int idyC = int(dyC);
int wCPerm = ${filterWidth} - 1 - wC;
for (int d2 = 0; d2 < ${convInfo.outChannels}; d2++) {
float xValue = getDy(batch, idyF, idyR, idyC, d2);
float wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
dotProd += xValue * wValue;
}
}
}
}
setOutput(dotProd);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv2DBackpropFilter$1(args) {
const { inputs, backend, attrs } = args;
const { x, dy } = inputs;
const { strides, pad, dataFormat, dimRoundingMode, filterShape } = attrs;
const $dataFormat = convertConv2DDataFormat(dataFormat);
const convInfo = computeConv2DInfo(x.shape, filterShape, strides, 1 /* dilations */, pad, dimRoundingMode, false /* depthwise */, $dataFormat);
const program = new Conv2DDerFilterProgram(convInfo);
return backend.runWebGLProgram(program, [x, dy], 'float32');
}
const conv2DBackpropFilterConfig$1 = {
kernelName: Conv2DBackpropFilter,
backendName: 'webgl',
kernelFunc: conv2DBackpropFilter$1,
};
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Conv2DDerInputPackedProgram {
constructor(convInfo) {
this.variableNames = ['dy', 'W'];
this.packedInputs = true;
this.packedOutput = true;
this.customUniforms = [
{ name: 'strides', type: 'vec2' },
];
this.outputShape = convInfo.inShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const padTop = filterHeight - 1 - convInfo.padInfo.top;
const padLeft = filterWidth - 1 - convInfo.padInfo.left;
this.userCode = `
const ivec2 pads = ivec2(${padTop}, ${padLeft});
void main() {
ivec4 coords = getOutputCoords();
int batch = coords[0];
int d1 = coords[3];
ivec2 dyCorner = ivec2(coords[1], coords[2]) - pads;
int dyRCorner = dyCorner.x;
int dyCCorner = dyCorner.y;
vec4 result = vec4(0.);
for (int wR = 0; wR < ${filterHeight}; wR++) {
float dyR = float(dyRCorner + wR) / strides[0];
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 || fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
int wRPerm = ${filterHeight} - 1 - wR;
for (int wC = 0; wC < ${filterWidth}; wC++) {
int wCPerm = ${filterWidth} - 1 - wC;
float dyC = float(dyCCorner + wC) / strides[1];
bool idyCVal = (dyC >= 0.0) && (dyC < ${convInfo.outWidth}.0)
&& (fract(dyC) == 0.0);
int idyC = int(dyC);
float dyC2 = float(dyCCorner + wC + 1) / strides[1];
bool idyCVal2 = (dyC2 >= 0.0) && (dyC2 < ${convInfo.outWidth}.0)
&& (fract(dyC2) == 0.0);
int idyC2 = int(dyC2);
if (idyCVal && idyCVal2) {
for (int d2 = 0; d2 < ${convInfo.outChannels}; d2 += 2) {
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
vec4 dySample = getDy(batch, idyR, idyC, d2);
vec4 dySample2 = (idyC / 2 == idyC2 / 2) ?
dySample : getDy(batch, idyR, idyC2, d2);
vec2 dyValue = mod(float(idyC), 2.) == 0. ?
dySample.xy : dySample.zw;
result.xy += vec2(dot(dyValue, wValue.xy),
dot(dyValue, wValue.zw));
dyValue = mod(float(idyC2), 2.) == 0. ?
dySample2.xy : dySample2.zw;
result.zw += vec2(dot(dyValue, wValue.xy),
dot(dyValue, wValue.zw));
}
} else if (idyCVal) {
for (int d2 = 0; d2 < ${convInfo.outChannels}; d2 += 2) {
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
vec4 dySample = getDy(batch, idyR, idyC, d2);
vec2 dyValue = mod(float(idyC), 2.) == 0. ?
dySample.xy : dySample.zw;
result.xy += vec2(dot(dyValue, wValue.xy),
dot(dyValue, wValue.zw));
}
} else if (idyCVal2) {
for (int d2 = 0; d2 < ${convInfo.outChannels}; d2 += 2) {
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
vec4 dySample = getDy(batch, idyR, idyC2, d2);
vec2 dyValue = mod(float(idyC2), 2.) == 0. ?
dySample.xy : dySample.zw;
result.zw += vec2(dot(dyValue, wValue.xy),
dot(dyValue, wValue.zw));
}
}
}
}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv2DBackpropInput$1(args) {
const { inputs, backend, attrs } = args;
const { dy, filter } = inputs;
const { inputShape, strides, pad, dataFormat, dimRoundingMode } = attrs;
const $dataFormat = convertConv2DDataFormat(dataFormat);
const convInfo = computeConv2DInfo(inputShape, filter.shape, strides, 1 /* dilations */, pad, dimRoundingMode, false, $dataFormat);
if (env().getBool('WEBGL_PACK_CONV2DTRANSPOSE') &&
$dataFormat === 'channelsLast') {
const customValues = [
[convInfo.strideHeight, convInfo.strideWidth],
];
const program = new Conv2DDerInputPackedProgram(convInfo);
return backend.runWebGLProgram(program, [dy, filter], 'float32', customValues);
}
else {
const program = new Conv2DDerInputProgram(convInfo);
return backend.runWebGLProgram(program, [dy, filter], 'float32');
}
}
const conv2DBackpropInputConfig$1 = {
kernelName: Conv2DBackpropInput,
backendName: 'webgl',
kernelFunc: conv2DBackpropInput$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv3D$1(args) {
const { inputs, backend, attrs } = args;
const { x, filter } = inputs;
const { strides, pad, dilations } = attrs;
const convInfo = computeConv3DInfo(x.shape, filter.shape, strides, dilations, pad);
const program = new Conv3DProgram(convInfo);
return backend.runWebGLProgram(program, [x, filter], 'float32');
}
const conv3DConfig$1 = {
kernelName: Conv3D,
backendName: 'webgl',
kernelFunc: conv3D$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv3DBackpropFilterV2$1(args) {
const { inputs, backend, attrs } = args;
const { x, dy } = inputs;
const { strides, pad, filterShape } = attrs;
const convInfo = computeConv3DInfo(x.shape, filterShape, strides, 1 /* dilations */, pad);
const program = new Conv3DDerFilterProgram(convInfo);
return backend.runWebGLProgram(program, [x, dy], 'float32');
}
const conv3DBackpropFilterV2Config$1 = {
kernelName: Conv3DBackpropFilterV2,
backendName: 'webgl',
kernelFunc: conv3DBackpropFilterV2$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv3DBackpropInput(args) {
const { inputs, backend, attrs } = args;
const { dy, filter } = inputs;
const { pad, strides, inputShape } = attrs;
const convInfo = computeConv3DInfo(inputShape, filter.shape, strides, 1 /* dilations */, pad);
const program = new Conv3DDerInputProgram(convInfo);
return backend.runWebGLProgram(program, [dy, filter], 'float32');
}
const conv3DBackpropInputConfig = {
kernelName: Conv3DBackpropInputV2,
backendName: 'webgl',
kernelFunc: conv3DBackpropInput,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const COS = CHECK_NAN_SNIPPET_UNARY + `
return cos(x);
`;
const COS_PACKED = `
vec4 result = cos(x);
bvec4 isNaN = isnan(x);
${CHECK_NAN_SNIPPET_PACKED}
return result;
`;
const cos$1 = unaryKernelFunc({ opSnippet: COS, packedOpSnippet: COS_PACKED });
const cosConfig$1 = {
kernelName: Cos,
backendName: 'webgl',
kernelFunc: cos$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const COSH = `
float e2x = exp(-x);
return (e2x + 1.0 / e2x) / 2.0;
`;
const cosh$1 = unaryKernelFunc({ opSnippet: COSH });
const coshConfig$1 = {
kernelName: Cosh,
backendName: 'webgl',
kernelFunc: cosh$1,
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class CropAndResizeProgram {
constructor(imageShape, boxShape, cropSize, method, extrapolationValue) {
this.variableNames = ['Image', 'Boxes', 'BoxInd'];
this.outputShape = [];
const [batch, imageHeight, imageWidth, depth] = imageShape;
const [numBoxes,] = boxShape;
const [cropHeight, cropWidth] = cropSize;
this.outputShape = [numBoxes, cropHeight, cropWidth, depth];
const methodId = method === 'bilinear' ? 1 : 0;
const [inputHeightFloat, inputWidthFloat] = [`${imageHeight - 1}.0`, `${imageWidth - 1}.0`];
const [heightRatio, heightScale, inY] = cropHeight > 1 ?
[
`${(imageHeight - 1) / (cropHeight - 1)}`,
'(y2-y1) * height_ratio',
`y1*${inputHeightFloat} + float(y)*(height_scale)`,
] :
[
'0.0',
'0.0',
`0.5 * (y1+y2) * ${inputHeightFloat}`,
];
const [widthRatio, widthScale, inX] = cropWidth > 1 ?
[
`${(imageWidth - 1) / (cropWidth - 1)}`,
'(x2-x1) * width_ratio',
`x1*${inputWidthFloat} + float(x)*(width_scale)`,
] :
[
'0.0',
'0.0',
`0.5 * (x1+x2) * ${inputWidthFloat}`,
];
// Reference implementation
// tslint:disable-next-line:max-line-length
// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/crop_and_resize_op_gpu.cu.cc
this.userCode = `
const float height_ratio = float(${heightRatio});
const float width_ratio = float(${widthRatio});
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int y = coords[1];
int x = coords[2];
int d = coords[3];
// get box vals
float y1 = getBoxes(b,0);
float x1 = getBoxes(b,1);
float y2 = getBoxes(b,2);
float x2 = getBoxes(b,3);
// get image in batch index
int bInd = round(getBoxInd(b));
if(bInd < 0 || bInd >= ${batch}) {
return;
}
float height_scale = ${heightScale};
float width_scale = ${widthScale};
float in_y = ${inY};
if( in_y < 0.0 || in_y > ${inputHeightFloat} ) {
setOutput(float(${extrapolationValue}));
return;
}
float in_x = ${inX};
if( in_x < 0.0 || in_x > ${inputWidthFloat} ) {
setOutput(float(${extrapolationValue}));
return;
}
vec2 sourceFracIndexCR = vec2(in_x,in_y);
if(${methodId} == 1) {
// Compute the four integer indices.
ivec2 sourceFloorCR = ivec2(sourceFracIndexCR);
ivec2 sourceCeilCR = ivec2(ceil(sourceFracIndexCR));
float topLeft = getImage(b, sourceFloorCR.y, sourceFloorCR.x, d);
float bottomLeft = getImage(b, sourceCeilCR.y, sourceFloorCR.x, d);
float topRight = getImage(b, sourceFloorCR.y, sourceCeilCR.x, d);
float bottomRight = getImage(b, sourceCeilCR.y, sourceCeilCR.x, d);
vec2 fracCR = sourceFracIndexCR - vec2(sourceFloorCR);
float top = topLeft + (topRight - topLeft) * fracCR.x;
float bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
float newValue = top + (bottom - top) * fracCR.y;
setOutput(newValue);
} else {
// Compute the coordinators of nearest neighbor point.
ivec2 sourceNearestCR = ivec2(floor(
sourceFracIndexCR + vec2(0.5,0.5)));
float newValue = getImage(b, sourceNearestCR.y, sourceNearestCR.x, d);
setOutput(newValue);
}
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const cropAndResize$1 = (args) => {
const { inputs, backend, attrs } = args;
const { image, boxes, boxInd } = inputs;
const { cropSize, method, extrapolationValue } = attrs;
const program = new CropAndResizeProgram(image.shape, boxes.shape, cropSize, method, extrapolationValue);
return backend.runWebGLProgram(program, [image, boxes, boxInd], 'float32');
};
const cropAndResizeConfig$1 = {
kernelName: CropAndResize,
backendName: 'webgl',
kernelFunc: cropAndResize$1
};
var CumOpType;
(function (CumOpType) {
CumOpType["Prod"] = "*";
CumOpType["Sum"] = "+";
})(CumOpType || (CumOpType = {}));
class CumProgram {
constructor(op, outputShape, exclusive, reverse) {
this.op = op;
this.outputShape = outputShape;
this.variableNames = ['x'];
this.customUniforms = [{ name: 'index', type: 'float' }];
const rank = this.outputShape.length;
const initVal = this.op === CumOpType.Prod ? '1.0' : '0.0';
const val = exclusive ? initVal : `getX(${getCoords(rank, 'coords', this.op)})`;
const length = this.outputShape[this.outputShape.length - 1];
let condition = '';
let idxString = '';
// When exclusive is set, the cum op becomes roll op that copies the
// value from the previous index based on the direction specified by the
// reverse flag.
if (exclusive) {
condition = reverse ? `end != ${length - 1}` : 'end != 0';
idxString = reverse ? 'end + 1' : 'end - 1';
}
else {
condition = reverse ? `end + pow2 < ${length}` : 'end >= pow2';
idxString = (reverse ? 'end + pow2' : 'end - pow2');
}
this.userCode = `
void main() {
${getCoordsDataType(rank)} coords = getOutputCoords();
int end = ${getFinalCoord(rank, 'coords', this.op)};
float val = ${val};
int pow2 = int(pow(2.0, index));
if (${condition}) {
int idx = ${idxString};
${getFinalCoord(rank, 'coords', this.op)} = idx;
val ${this.op}= getX(${getCoords(rank, 'coords', this.op)});
}
setOutput(val);
}
`;
}
}
function getCoords(rank, name, op) {
if (rank === 1) {
return `${name}`;
}
else if (rank === 2) {
return `${name}.x, ${name}.y`;
}
else if (rank === 3) {
return `${name}.x, ${name}.y, ${name}.z`;
}
else if (rank === 4) {
return `${name}.x, ${name}.y, ${name}.z, ${name}.w`;
}
else {
throw new Error(`Cumulative ${op} for rank ${rank} is not yet supported`);
}
}
function getFinalCoord(rank, name, op) {
if (rank === 1) {
return `${name}`;
}
else if (rank === 2) {
return `${name}.y`;
}
else if (rank === 3) {
return `${name}.z`;
}
else if (rank === 4) {
return `${name}.w`;
}
else {
throw new Error(`Cumulative ${op} for rank ${rank} is not yet supported`);
}
}
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function cumImpl(op, x, backend, axis, exclusive, reverse) {
const xRank = x.shape.length;
const permutation = getAxesPermutation([axis], xRank);
let permutedX = x;
if (permutation != null) {
permutedX = transpose({ inputs: { x }, backend, attrs: { perm: permutation } });
}
const permutedAxis = getInnerMostAxes(1, xRank)[0];
if (permutedAxis !== xRank - 1) {
throw new Error(`WebGL cumprod shader expects an inner-most axis=${x.shape.length - 1} ` +
`but got axis=${axis}`);
}
const size = permutedX.shape[permutedAxis];
let result = identity({ inputs: { x: permutedX }, backend });
// Use cum parallel algorithm, inspired by:
// https://developer.nvidia.com/gpugems/gpugems3/part-vi-gpu-computing/chapter-39-parallel-prefix-sum-scan-cuda
// Note: although the algorithm is called sum, it works for any associtative
// operator with an identity.
for (let i = 0; i <= Math.ceil(Math.log2(size)) - 1; i++) {
const program = new CumProgram(op, permutedX.shape, false, reverse);
const customValues = [[i]];
const prevResult = result;
result =
backend.runWebGLProgram(program, [result], result.dtype, customValues);
backend.disposeIntermediateTensorInfo(prevResult);
}
// For exclusive cum, shift the end result in the direction of product or sum
// and add 1 for product or 0 for sum to the front index.
if (exclusive) {
const program = new CumProgram(op, permutedX.shape, exclusive, reverse);
const prevResult = result;
result = backend.runWebGLProgram(program, [result], result.dtype);
backend.disposeIntermediateTensorInfo(prevResult);
}
if (permutation != null) {
const reversePermutation = getUndoAxesPermutation(permutation);
const reverseTransposedResult = transpose({ inputs: { x: result }, backend, attrs: { perm: reversePermutation } });
backend.disposeIntermediateTensorInfo(result);
backend.disposeIntermediateTensorInfo(permutedX);
return reverseTransposedResult;
}
return result;
}
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function cumprod$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, exclusive, reverse } = attrs;
return cumImpl(CumOpType.Prod, x, backend, axis, exclusive, reverse);
}
const cumprodConfig$1 = {
kernelName: Cumprod,
backendName: 'webgl',
kernelFunc: cumprod$1
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function cumsum$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, exclusive, reverse } = attrs;
return cumImpl(CumOpType.Sum, x, backend, axis, exclusive, reverse);
}
const cumsumConfig$1 = {
kernelName: Cumsum,
backendName: 'webgl',
kernelFunc: cumsum$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function denseBincount$1(args) {
const { inputs, backend, attrs } = args;
const { x, weights } = inputs;
const { size, binaryOutput } = attrs;
if (x.shape.length === 1) {
const xVals = backend.readSync(x.dataId);
const weightsVals = backend.readSync(weights.dataId);
const outVals = bincountImplCPU(xVals, weightsVals, weights.dtype, weights.shape, size);
return backend.makeTensorInfo([size], weights.dtype, outVals);
}
else if (x.shape.length === 2) {
const xBuf = backend.bufferSync(x);
const weightsBuf = backend.bufferSync(weights);
const outBuf = bincountReduceImplCPU(xBuf, weightsBuf, size, binaryOutput);
return backend.makeTensorInfo(outBuf.shape, weights.dtype, outBuf.values);
}
throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank` +
`${x.shape.length}.`);
}
const denseBincountConfig$1 = {
kernelName: DenseBincount,
backendName: 'webgl',
kernelFunc: denseBincount$1
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class DepthToSpaceProgram {
constructor(outputShape, blockSize, dataFormat) {
this.variableNames = ['x'];
this.outputShape = [];
this.outputShape = outputShape;
this.blockSize = blockSize;
this.dataFormat = dataFormat;
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int h = ${this.getHeightCoordString()};
int w = ${this.getWidthCoordString()};
int d = ${this.getDepthCoordString()};
int in_h = h / ${blockSize};
int offset_h = imod(h, ${blockSize});
int in_w = w / ${blockSize};
int offset_w = imod(w, ${blockSize});
int offset_d = (offset_h * ${blockSize} + offset_w) *
${this.getOutputDepthSize()};
int in_d = d + offset_d;
float result = ${this.getInputSamplingString()};
setOutput(result);
}
`;
}
getHeightCoordString() {
if (this.dataFormat === 'NHWC') {
return `coords[1]`;
}
else {
return `coords[2]`;
}
}
getWidthCoordString() {
if (this.dataFormat === 'NHWC') {
return `coords[2]`;
}
else {
return `coords[3]`;
}
}
getDepthCoordString() {
if (this.dataFormat === 'NHWC') {
return `coords[3]`;
}
else {
return `coords[1]`;
}
}
getOutputDepthSize() {
if (this.dataFormat === 'NHWC') {
return this.outputShape[3];
}
else {
return this.outputShape[1];
}
}
getInputSamplingString() {
if (this.dataFormat === 'NHWC') {
return `getX(b, in_h, in_w, in_d)`;
}
else {
return `getX(b, in_d, in_h, in_w)`;
}
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthToSpace$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { blockSize, dataFormat } = attrs;
const batchSize = x.shape[0];
const inputHeight = (dataFormat === 'NHWC') ? x.shape[1] : x.shape[2];
const inputWidth = (dataFormat === 'NHWC') ? x.shape[2] : x.shape[3];
const inputDepth = (dataFormat === 'NHWC') ? x.shape[3] : x.shape[1];
const outputHeight = inputHeight * blockSize;
const outputWidth = inputWidth * blockSize;
const outputDepth = inputDepth / (blockSize * blockSize);
const outputShape = (dataFormat === 'NHWC') ?
[batchSize, outputHeight, outputWidth, outputDepth] :
[batchSize, outputDepth, outputHeight, outputWidth];
const program = new DepthToSpaceProgram(outputShape, blockSize, dataFormat);
return backend.runWebGLProgram(program, [x], x.dtype);
}
const depthToSpaceConfig$1 = {
kernelName: DepthToSpace,
backendName: 'webgl',
kernelFunc: depthToSpace$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class DepthwiseConv2DProgram {
constructor(convInfo, addBias = false, activation = null, hasPreluActivation = false, hasLeakyReluAlpha = false) {
this.variableNames = ['x', 'W'];
this.customUniforms = [
{ name: 'pads', type: 'ivec2' },
{ name: 'strides', type: 'ivec2' },
{ name: 'dilations', type: 'ivec2' },
{ name: 'inDims', type: 'ivec2' },
];
this.outputShape = convInfo.outShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const channelMul = convInfo.outChannels / convInfo.inChannels;
let activationSnippet = '', applyActivationSnippet = '';
if (activation) {
if (hasPreluActivation) {
activationSnippet = `float activation(float a) {
float b = getPreluActivationWeightsAtOutCoords();
${activation}
}`;
}
else if (hasLeakyReluAlpha) {
activationSnippet = `float activation(float a) {
float b = getLeakyreluAlphaAtOutCoords();
${activation}
}`;
}
else {
activationSnippet = `
float activation(float x) {
${activation}
}
`;
}
applyActivationSnippet = `result = activation(result);`;
}
const addBiasSnippet = addBias ? 'result += getBiasAtOutCoords();' : '';
if (addBias) {
this.variableNames.push('bias');
}
if (hasPreluActivation) {
this.variableNames.push('preluActivationWeights');
}
if (hasLeakyReluAlpha) {
this.variableNames.push('leakyreluAlpha');
}
this.userCode = `
${activationSnippet}
void main() {
ivec4 coords = getOutputCoords();
int batch = coords.x;
ivec2 xRCCorner = coords.yz * strides - pads;
int d2 = coords.w;
int d1 = d2 / ${channelMul};
int q = d2 - d1 * ${channelMul};
int xRCorner = xRCCorner.x;
int xCCorner = xRCCorner.y;
// Convolve x(?, ?, d1) with w(:, :, d1, q) to get y(yR, yC, d2).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
// TO DO(dsmilkov): Flatten the two for loops and vec4 the operations.
for (int wR = 0; wR < ${filterHeight}; wR++) {
int xR = xRCorner + wR * dilations[0];
if (xR < 0 || xR >= inDims[0]) {
continue;
}
for (int wC = 0; wC < ${filterWidth}; wC++) {
int xC = xCCorner + wC * dilations[1];
if (xC < 0 || xC >= inDims[1]) {
continue;
}
float xVal = getX(batch, xR, xC, d1);
float wVal = getW(wR, wC, d1, q);
dotProd += xVal * wVal;
}
}
float result = dotProd;
${addBiasSnippet}
${applyActivationSnippet}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class DepthwiseConvPacked2DProgram {
constructor(convInfo, addBias = false, activation = null, hasPreluActivation = false, hasLeakyReluAlpha = false) {
this.variableNames = ['x', 'W'];
this.packedInputs = true;
this.packedOutput = true;
this.customUniforms = [
{ name: 'pads', type: 'ivec2' },
{ name: 'strides', type: 'ivec2' },
{ name: 'dilations', type: 'ivec2' },
{ name: 'inDims', type: 'ivec2' },
];
this.outputShape = convInfo.outShape;
this.enableShapeUniforms = useShapeUniforms(this.outputShape.length);
const channelMul = convInfo.outChannels / convInfo.inChannels;
const padLeft = convInfo.padInfo.left;
const strideWidth = convInfo.strideWidth;
const dilationWidth = convInfo.dilationWidth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const texelsAcross = filterWidth;
let mainLoop = `
int xR; int xC; int xCOffset;
vec4 wTexel; vec4 previous; vec4 final;`;
for (let c = 0; c < filterWidth; c++) {
mainLoop += `
vec4 xTexelC${c * 2};
int xTexelC${c * 2}Ready;
vec4 xTexelC${c * 2 + 1};
int xTexelC${c * 2 + 1}Ready;
vec4 xC${c};`;
}
/**
* This vectorized implementation works by gathering the values needed for
* each output channel's dot product into vec4's and then multiplying them
* all together (this happens in the final double for-loop below). Most of
* the main loop consists of constructing these vec4's with the minimum
* number of texture2D calls, which means making use of all four returned
* values from a texture2D call at once.
*/
mainLoop += `
for (int r = 0; r < ${filterHeight}; r++) {
`;
for (let c = 0; c < filterWidth; c++) {
mainLoop += `
xTexelC${c * 2} = vec4(0.0);
xTexelC${c * 2}Ready = 0;
xTexelC${c * 2 + 1} = vec4(0.0);
xTexelC${c * 2 + 1}Ready = 0;
xC${c} = vec4(0.0);`;
}
mainLoop += `
xR = xRCorner + r * dilations[0];
if (xR >=0 && xR < inDims[0]) {
`;
for (let texelC = 0; texelC < (texelsAcross + 1) / 2; texelC++) {
const colIndex = texelC * 2;
mainLoop += `
xC = xCCorner + ${colIndex * dilationWidth};
`;
if (strideWidth === 1) {
if (colIndex < filterWidth) {
// If padding is odd, the outer texels have to be composed.
if (padLeft % 2 === 1) {
// TODO: Ensure vec4 previous does not result in redundant sample,
// and avoid setting xTexelRC's that exceed the boundary in the
// first place rather than resetting them to vec4(0)).
// To compute xCOffset:
// - If padding is odd, we must add 1 to ensure we ask for an
// even-numbered row.
// - We subtract 2 to access the previous texel.
mainLoop += `
xCOffset = xC + 1;
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
`;
// This texel has been read in previous iteration if the dilation
// is 1.
if (dilationWidth === 1 && colIndex > 0) {
mainLoop += `
xC${colIndex} = vec4(xTexelC${colIndex - 2}.zw, xTexelC${colIndex}.xy);
`;
}
else {
mainLoop += `
xCOffset = xC + 1 - 2;
if (xCOffset >= 0 && xCOffset < inDims[1]) {
previous = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
previous.zw = vec2(0.0);
}
xC${colIndex} = vec4(previous.zw, xTexelC${colIndex}.xy);
} else {
xC${colIndex} = vec4(0.0, 0.0, xTexelC${colIndex}.xy);
}
`;
}
}
else {
// Padding is even, so xRC corresponds to a single texel.
mainLoop += `
if (xC >= 0 && xC < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xC, d1);
if (xC + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
xC${colIndex} = xTexelC${colIndex};
`;
}
if (colIndex + 1 < filterWidth) {
// If dilation is even, the second entry should match the first
// (either both are composed or both are single samples). But if
// dilation is odd, then the second entry should be the opposite
// of the first (if the first is composed, the second is a single
// sample, and vice versa.)
const nextTexelOffset = padLeft % 2 === 0 ?
nearestLargerEven(dilationWidth) :
dilationWidth;
if ((dilationWidth % 2 === 0 && padLeft % 2 === 1) ||
(dilationWidth % 2 !== 0 && padLeft % 2 !== 1)) {
mainLoop += `
xCOffset = xC + imod(pads[1], 2) + ${nextTexelOffset};
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.0);
}
xTexelC${colIndex + 1}Ready = 1;
}
`;
// If dilation > 1 then the xRC's will not be able to share any
// values, so each xRC will require two unique calls to getX.
if (dilationWidth > 1) {
mainLoop += `
xCOffset -= 2;
if (xCOffset >= 0 && xCOffset < inDims[1]) {
previous = getX(batch, xR, xCOffset, d1);
xC${colIndex + 1} = vec4(previous.zw, xTexelC${colIndex + 1}.xy);
} else {
xC${colIndex + 1} = vec4(0.0, 0.0, xTexelC${colIndex + 1}.xy);
}
`;
}
else {
mainLoop += `
xC${colIndex + 1} = vec4(xTexelC${colIndex}.zw, xTexelC${colIndex + 1}.xy);
`;
}
}
else {
// If dilation is 1 and padding is odd, we have already read the
// texel when constructing the previous x value. Here we can
// simply skip the texture read.
if (nextTexelOffset === 1) {
mainLoop += `
xC${colIndex + 1} = xTexelC${colIndex};
`;
}
else {
mainLoop += `
xCOffset = xC + ${nextTexelOffset};
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xCOffset, d1);
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.0);
}
xTexelC${colIndex + 1}Ready = 1;
}
xC${colIndex + 1} = xTexelC${colIndex + 1};
`;
}
}
}
}
}
else { // stride === 2
if (colIndex < filterWidth) {
// Depending on whether padLeft is even or odd, we want either the
// xy or zw channels from X texels for xC${colIndex}. If padLeft is
// even, xC${colIndex +1} is simply the zw channels of texels we've
// already sampled. But if padLeft is odd, xC{$c + 1}.zw will
// need to come from the xy channels of a new texel, hence the `
// vec4
// final` initialized below.
if (padLeft % 2 === 1) {
mainLoop += `
xCOffset = xC + 1 - strides[1];
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xCOffset, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xC + 1, d1);
// Need to manually clear unused channels in case
// we're reading from recycled texture.
if (xC + 2 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.0);
}
xTexelC${colIndex + 1}Ready = 1;
}
xC${colIndex} = vec4(xTexelC${colIndex}.zw, xTexelC${colIndex + 1}.zw);
`;
if (colIndex + 1 < filterWidth) {
mainLoop += `
final = vec4(0.0);
xCOffset = xC + 1 + strides[1];
if(xCOffset >= 0 && xCOffset < inDims[1]) {
final = getX(batch, xR, xCOffset, d1);
}
xC${colIndex + 1} = vec4(xTexelC${colIndex + 1}.xy, final.xy);
`;
}
}
else {
mainLoop += `
if(xC >= 0 && xC < inDims[1] && xTexelC${colIndex}Ready == 0) {
xTexelC${colIndex} = getX(batch, xR, xC, d1);
if (xC + 1 >= inDims[1]) {
xTexelC${colIndex}.zw = vec2(0.0);
}
xTexelC${colIndex}Ready = 1;
}
xCOffset = xC + strides[1];
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${colIndex + 1}Ready == 0) {
xTexelC${colIndex + 1} = getX(batch, xR, xCOffset, d1);
if (xCOffset + 1 >= inDims[1]) {
xTexelC${colIndex + 1}.zw = vec2(0.);
}
xTexelC${colIndex + 1}Ready = 1;
}
xC${colIndex} = vec4(
xTexelC${colIndex}.xy, xTexelC${colIndex + 1}.xy);
`;
if (colIndex + 1 < filterWidth) {
mainLoop += `
xC${colIndex + 1} = vec4(xTexelC${colIndex}.zw, xTexelC${colIndex + 1}.zw);
`;
}
}
}
}
// localize the dotProd accumulation within the loop, the theory is for
// GPU with limited cache, accumulate sum across large amount of
// veriables will cause lots of cache misses. (i.e. 5x5 filter will have
// 50 variables)
if (colIndex < filterWidth) {
mainLoop += `
wTexel = getW(r, ${colIndex}, d1, q);
dotProd += xC${colIndex} * vec4(wTexel.xz, wTexel.xz);
`;
if (colIndex + 1 < filterWidth) {
mainLoop += `
wTexel = getW(r, ${colIndex + 1}, d1, q);
dotProd += xC${colIndex + 1} * vec4(wTexel.xz, wTexel.xz);
`;
}
}
}
mainLoop += `
}
`;
mainLoop += `
}
`;
let activationSnippet = '', applyActivationSnippet = '';
if (activation) {
if (hasPreluActivation) {
activationSnippet = `vec4 activation(vec4 a) {
vec4 b = getPreluActivationWeightsAtOutCoords();
${activation}
}`;
}
else if (hasLeakyReluAlpha) {
activationSnippet = `vec4 activation(vec4 a) {
vec4 b = getLeakyreluAlphaAtOutCoords();
${activation}
}`;
}
else {
activationSnippet = `vec4 activation(vec4 x) {
${activation}
}`;
}
applyActivationSnippet = `result = activation(result);`;
}
const addBiasSnippet = addBias ? 'result += getBiasAtOutCoords();' : '';
if (addBias) {
this.variableNames.push('bias');
}
if (hasPreluActivation) {
this.variableNames.push('preluActivationWeights');
}
if (hasLeakyReluAlpha) {
this.variableNames.push('leakyreluAlpha');
}
this.userCode = `
${activationSnippet}
void main() {
ivec4 coords = getOutputCoords();
int batch = coords.x;
ivec2 xRCCorner = coords.yz * strides - pads;
int d2 = coords.w;
int d1 = d2 / ${channelMul};
int q = d2 - d1 * ${channelMul};
int xRCorner = xRCCorner.x;
int xCCorner = xRCCorner.y;
//intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
vec4 dotProd = vec4(0.000000000000001);
${mainLoop}
vec4 result = dotProd - vec4(0.000000000000001);
${addBiasSnippet}
${applyActivationSnippet}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNative$1(args) {
const { inputs, backend, attrs } = args;
const { x, filter } = inputs;
const { strides, pad, dilations, dimRoundingMode } = attrs;
let $dilations = dilations;
if ($dilations == null) {
$dilations = [1, 1];
}
assert$1(eitherStridesOrDilationsAreOne(strides, $dilations), () => 'Error in depthwiseConv2d: Either strides or dilations must be ' +
`1. Got strides ${strides} and dilations '${$dilations}'`);
const convInfo = computeConv2DInfo(x.shape, filter.shape, strides, $dilations, pad, dimRoundingMode, true /* depthwise */);
let program;
if (env().getBool('WEBGL_PACK_DEPTHWISECONV') && convInfo.strideWidth <= 2 &&
convInfo.outChannels / convInfo.inChannels === 1) {
program = new DepthwiseConvPacked2DProgram(convInfo);
}
else {
program = new DepthwiseConv2DProgram(convInfo);
}
const customValues = [
[convInfo.padInfo.top, convInfo.padInfo.left],
[convInfo.strideHeight, convInfo.strideWidth],
[convInfo.dilationHeight, convInfo.dilationWidth],
[convInfo.inHeight, convInfo.inWidth]
];
return backend.runWebGLProgram(program, [x, filter], 'float32', customValues);
}
const depthwiseConv2dNativeConfig$1 = {
kernelName: DepthwiseConv2dNative,
backendName: 'webgl',
kernelFunc: depthwiseConv2dNative$1,
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class DepthwiseConv2DDerFilterProgram {
constructor(convInfo) {
this.variableNames = ['x', 'dy'];
this.outputShape = convInfo.filterShape;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
const channelMul = convInfo.outChannels / convInfo.inChannels;
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int wR = coords.x;
int wC = coords.y;
int d1 = coords.z;
int dm = coords.w;
int d2 = d1 * ${channelMul} + dm;
float dotProd = 0.0;
// TO DO: Vec4 over the batch size
for (int b = 0; b < ${convInfo.batchSize}; b++) {
for (int yR = 0; yR < ${convInfo.outHeight}; yR++) {
int xR = wR + yR * ${strideHeight} - ${padTop};
if (xR < 0 || xR >= ${convInfo.inHeight}) {
continue;
}
for (int yC = 0; yC < ${convInfo.outWidth}; yC++) {
int xC = wC + yC * ${strideWidth} - ${padLeft};
if (xC < 0 || xC >= ${convInfo.inWidth}) {
continue;
}
float dyValue = getDy(b, yR, yC, d2);
float xValue = getX(b, xR, xC, d1);
dotProd += (xValue * dyValue);
}
}
}
setOutput(dotProd);
}
`;
}
}
class DepthwiseConv2DDerInputProgram {
constructor(convInfo) {
this.variableNames = ['dy', 'W'];
this.outputShape = convInfo.inShape;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const padTop = filterHeight - 1 - convInfo.padInfo.top;
const padLeft = filterWidth - 1 - convInfo.padInfo.left;
const channelMul = convInfo.outChannels / convInfo.inChannels;
this.userCode = `
const ivec2 pads = ivec2(${padTop}, ${padLeft});
void main() {
ivec4 coords = getOutputCoords();
int batch = coords[0];
int d1 = coords[3];
ivec2 dyCorner = coords.yz - pads;
int dyRCorner = dyCorner.x;
int dyCCorner = dyCorner.y;
float dotProd = 0.0;
for (int wR = 0; wR < ${filterHeight}; wR++) {
float dyR = float(dyRCorner + wR) / ${strideHeight}.0;
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 || fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
int wRPerm = ${filterHeight} - 1 - wR;
for (int wC = 0; wC < ${filterWidth}; wC++) {
float dyC = float(dyCCorner + wC) / ${strideWidth}.0;
if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
fract(dyC) > 0.0) {
continue;
}
int idyC = int(dyC);
int wCPerm = ${filterWidth} - 1 - wC;
// TO DO: Vec4 over the channelMul
for (int dm = 0; dm < ${channelMul}; dm++) {
int d2 = d1 * ${channelMul} + dm;
float xValue = getDy(batch, idyR, idyC, d2);
float wValue = getW(wRPerm, wCPerm, d1, dm);
dotProd += xValue * wValue;
}
}
}
setOutput(dotProd);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNativeBackpropFilter$1(args) {
const { inputs, backend, attrs } = args;
const { x, dy } = inputs;
const { strides, dilations, pad, dimRoundingMode, filterShape } = attrs;
const convInfo = computeConv2DInfo(x.shape, filterShape, strides, dilations, pad, dimRoundingMode, true /* depthwise */);
const program = new DepthwiseConv2DDerFilterProgram(convInfo);
return backend.runWebGLProgram(program, [x, dy], 'float32');
}
const depthwiseConv2dNativeBackpropFilterConfig$1 = {
kernelName: DepthwiseConv2dNativeBackpropFilter,
backendName: 'webgl',
kernelFunc: depthwiseConv2dNativeBackpropFilter$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNativeBackpropInput$1(args) {
const { inputs, backend, attrs } = args;
const { dy, filter } = inputs;
const { strides, dilations, pad, dimRoundingMode, inputShape } = attrs;
const convInfo = computeConv2DInfo(inputShape, filter.shape, strides, dilations, pad, dimRoundingMode, true /* depthwise */);
const program = new DepthwiseConv2DDerInputProgram(convInfo);
return backend.runWebGLProgram(program, [dy, filter], 'float32');
}
const depthwiseConv2dNativeBackpropInputConfig$1 = {
kernelName: DepthwiseConv2dNativeBackpropInput,
backendName: 'webgl',
kernelFunc: depthwiseConv2dNativeBackpropInput$1
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class DiagProgram {
constructor(size) {
this.variableNames = ['X'];
this.outputShape = [size, size];
this.userCode = `
void main() {
ivec2 coords = getOutputCoords();
float val = coords[0] == coords[1] ? getX(coords[0]) : 0.0;
setOutput(val);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function diag$1(args) {
const { inputs, backend } = args;
const { x } = inputs;
const outShape = [...x.shape, ...x.shape];
const xSize = sizeFromShape(x.shape);
const flat = reshape$1({ inputs: { x }, backend, attrs: { shape: [xSize] } });
const program = new DiagProgram(xSize);
const res = backend.runWebGLProgram(program, [flat], flat.dtype);
const out = reshape$1({ inputs: { x: res }, backend, attrs: { shape: outShape } });
backend.disposeIntermediateTensorInfo(flat);
backend.disposeIntermediateTensorInfo(res);
return out;
}
const diagConfig$1 = {
kernelName: Diag,
backendName: 'webgl',
kernelFunc: diag$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class Dilation2DProgram {
constructor(convInfo) {
this.variableNames = ['x', 'W'];
this.outputShape = convInfo.outShape;
const { inHeight, inWidth, padInfo, strideHeight, strideWidth, filterHeight, filterWidth, dilationHeight, dilationWidth } = convInfo;
const { top: padTop, left: padLeft } = padInfo;
this.userCode = `
const ivec2 strides = ivec2(${strideHeight}, ${strideWidth});
const ivec2 pads = ivec2(${padTop}, ${padLeft});
const float neg_infinity = -3.4e38;
void main() {
ivec4 coords = getOutputCoords();
int batch = coords.x;
int d1 = coords.w;
ivec2 outTopLeftCorner =
coords.yz * strides - pads;
int hBeg = outTopLeftCorner.x;
int wBeg = outTopLeftCorner.y;
float curVal = neg_infinity;
for (int h = 0; h < ${filterHeight}; h++) {
int hIn = hBeg + h * ${dilationHeight};
if (hIn >= 0 && hIn < ${inHeight}) {
for (int w = 0; w < ${filterWidth}; w++) {
int wIn = wBeg + w * ${dilationWidth};
if (wIn >= 0 && wIn < ${inWidth}) {
float xVal = getX(batch, hIn, wIn, d1);
float wVal = getW(h, w, d1);
float val = xVal + wVal;
if (val > curVal) {
curVal = val;
}
}
}
}
}
float result = curVal;
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function dilation2D(args) {
const { inputs, backend, attrs } = args;
const { x, filter } = inputs;
const { strides, pad, dilations } = attrs;
const convInfo = computeDilation2DInfo(x.shape, filter.shape, strides, pad, 'NHWC' /* dataFormat */, dilations);
let out;
const program = new Dilation2DProgram(convInfo);
out = backend.runWebGLProgram(program, [x, filter], 'float32');
const outReshaped = reshape$1({ inputs: { x: out }, backend, attrs: { shape: convInfo.outShape } });
backend.disposeIntermediateTensorInfo(out);
return outReshaped;
}
const dilation2DConfig$1 = {
kernelName: Dilation2D,
backendName: 'webgl',
kernelFunc: dilation2D,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function einsum$1(args) {
const { inputs, backend, attrs } = args;
const { equation } = attrs;
const tensors = inputs;
const { allDims, summedDims, idDims } = decodeEinsumEquation(equation, tensors.length);
checkEinsumDimSizes(allDims.length, idDims, tensors);
const { path, steps } = getEinsumComputePath(summedDims, idDims);
const nSteps = steps.length;
let out = null;
let numDimsRemaining = allDims.length;
const tensorsToDispose = [];
for (let i = 0; i < nSteps; ++i) {
for (const idTerm of steps[i]) {
const { permutationIndices: perm, expandDims: dimsToExpand } = getEinsumPermutation(numDimsRemaining, idDims[idTerm]);
let x;
if (isIdentityPermutation(perm)) {
x = tensors[idTerm];
}
else {
x = transpose({ inputs: { x: tensors[idTerm] }, backend, attrs: { perm } });
tensorsToDispose.push(x);
}
const targetShape = x.shape.slice();
for (let k = 0; k < dimsToExpand.length; ++k) {
targetShape.splice(dimsToExpand[k], 0, 1);
}
if (!arraysEqual(x.shape, targetShape)) {
x = reshape$1({ inputs: { x }, backend, attrs: { shape: targetShape } });
tensorsToDispose.push(x);
}
if (out === null) {
out = x;
}
else {
// tslint:disable-next-line: no-unnecessary-type-assertion
out = multiply({ inputs: { a: x, b: out }, backend });
tensorsToDispose.push(out);
}
}
if (i < nSteps - 1) {
if (path[i] >= 0) {
out = sum$1({
inputs: { x: out },
backend,
attrs: {
axis: path[i] - (allDims.length - numDimsRemaining),
keepDims: false
}
});
tensorsToDispose.push(out);
}
numDimsRemaining--;
}
}
// Clean up intermediate tensors.
for (const tensorInfo of tensorsToDispose) {
if (tensorInfo === out) {
continue;
}
backend.disposeIntermediateTensorInfo(tensorInfo);
}
return out;
}
const einsumConfig$1 = {
kernelName: Einsum,
backendName: 'webgl',
kernelFunc: einsum$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ELU = `return (x >= 0.0) ? x : (exp(x) - 1.0);`;
const ELU_PACKED = `
vec4 result;
result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);
return result;
`;
const elu$2 = unaryKernelFunc({ opSnippet: ELU, packedOpSnippet: ELU_PACKED });
const eluConfig$1 = {
kernelName: Elu$1,
backendName: 'webgl',
kernelFunc: elu$2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ELU_DER = `return (b >= 0.0) ? a : a * (b + 1.0);`;
const ELU_DER_PACKED = `
vec4 bGTEZero = vec4(greaterThanEqual(b, vec4(0.)));
return (bGTEZero * a) + ((vec4(1.0) - bGTEZero) * (a * (b + vec4(1.0))));
`;
const eluGrad$1 = (args) => {
const { inputs, backend } = args;
const { dy, y } = inputs;
const program = env().getBool('WEBGL_PACK_BINARY_OPERATIONS') ?
new BinaryOpPackedProgram(ELU_DER_PACKED, dy.shape, y.shape) :
new BinaryOpProgram(ELU_DER, dy.shape, y.shape);
return backend.runWebGLProgram(program, [dy, y], dy.dtype);
};
const eluGradConfig$2 = {
kernelName: EluGrad,
backendName: 'webgl',
kernelFunc: eluGrad$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const PACKED_EQUAL = `
return vec4(equal(a, b));
`;
const EQUAL = `return float(a == b);`;
const equal = binaryKernelFunc({
opSnippet: EQUAL,
packedOpSnippet: PACKED_EQUAL,
dtype: 'bool',
cpuKernelImpl: equalImplCPU,
});
const equalConfig = {
kernelName: Equal,
backendName: 'webgl',
kernelFunc: equal
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ERF = `
// Error function is calculated approximately with elementary function.
// See "Handbook of Mathematical Functions with Formulas,
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
float p = ${ERF_P};
float a1 = ${ERF_A1};
float a2 = ${ERF_A2};
float a3 = ${ERF_A3};
float a4 = ${ERF_A4};
float a5 = ${ERF_A5};
float sign = sign(x);
x = abs(x);
float t = 1.0 / (1.0 + p * x);
return sign * (1.0 - (((((a5*t + a4)*t) + a3)*t + a2)*t + a1)*t*exp(-x*x));
`;
const erf$1 = unaryKernelFunc({ opSnippet: ERF });
const erfConfig$1 = {
kernelName: Erf,
backendName: 'webgl',
kernelFunc: erf$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const EXP = CHECK_NAN_SNIPPET_UNARY + `
return exp(x);
`;
const EXP_PACKED = `
vec4 result = exp(x);
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : result.r;
result.g = isNaN.g ? x.g : result.g;
result.b = isNaN.b ? x.b : result.b;
result.a = isNaN.a ? x.a : result.a;
return result;
`;
const exp = unaryKernelFunc({
opSnippet: EXP,
packedOpSnippet: EXP_PACKED,
cpuKernelImpl: expImplCPU,
dtype: 'float32',
});
const expConfig = {
kernelName: Exp,
backendName: 'webgl',
kernelFunc: exp
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function expandDims$2(args) {
const { inputs, attrs, backend } = args;
const { dim } = attrs;
const { input } = inputs;
const inputRank = input.shape.length;
const newShape = input.shape.slice();
let $dim = dim;
if (dim < 0) {
// Negative value is counted from the tail of rank.
assert$1(-(inputRank + 1) <= dim, () => `Axis must be in the interval [${-(inputRank + 1)}, ${inputRank}]`);
$dim = inputRank + dim + 1;
}
newShape.splice($dim, 0, 1);
return reshape$1({ inputs: { x: input }, backend, attrs: { shape: newShape } });
}
const expandDimsConfig$1 = {
kernelName: ExpandDims,
backendName: 'webgl',
kernelFunc: expandDims$2,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const EXPM1 = `return exp(x) - 1.0;`;
const expm1 = unaryKernelFunc({ opSnippet: EXPM1, packedOpSnippet: EXPM1, cpuKernelImpl: expm1ImplCPU });
const expm1Config = {
kernelName: Expm1,
backendName: 'webgl',
kernelFunc: expm1
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class FFTProgram {
constructor(component, inputShape, inverse) {
this.variableNames = ['real', 'imag'];
const innerDim = inputShape[1];
this.outputShape = inputShape;
const exponentMultiplierSnippet = inverse ? `2.0 * ${Math.PI}` : `-2.0 * ${Math.PI}`;
const resultDenominator = inverse ? `${innerDim}.0` : '1.0';
let opString;
if (component === 'real') {
opString = 'return real * expR - imag * expI;';
}
else if (component === 'imag') {
opString = 'return real * expI + imag * expR;';
}
else {
throw new Error(`FFT component must be either "real" or "imag", got ${component}.`);
}
this.userCode = `
const float exponentMultiplier = ${exponentMultiplierSnippet};
float unaryOpComplex(float real, float expR, float imag, float expI) {
${opString}
}
float mulMatDFT(int batch, int index) {
float indexRatio = float(index) / float(${innerDim});
float exponentMultiplierTimesIndexRatio =
exponentMultiplier * indexRatio;
float result = 0.0;
for (int i = 0; i < ${innerDim}; i++) {
// x = (-2|2 * PI / N) * index * i;
float x = exponentMultiplierTimesIndexRatio * float(i);
float expR = cos(x);
float expI = sin(x);
float real = getReal(batch, i);
float imag = getImag(batch, i);
result +=
unaryOpComplex(real, expR, imag, expI) / ${resultDenominator};
}
return result;
}
void main() {
ivec2 coords = getOutputCoords();
setOutput(mulMatDFT(coords[0], coords[1]));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fftImpl$1(x, inverse, backend) {
const xData = backend.texData.get(x.dataId);
const inputSize = sizeFromShape(x.shape);
// Collapse all outer dimensions to a single batch dimension.
const innerDimensionSize = x.shape[x.shape.length - 1];
const batch = inputSize / innerDimensionSize;
const input2D = reshape$1({ inputs: { x }, backend, attrs: { shape: [batch, innerDimensionSize] } });
const xShape = input2D.shape;
const realProgram = new FFTProgram('real', xShape, inverse);
const imagProgram = new FFTProgram('imag', xShape, inverse);
const inputs = [
{
dataId: xData.complexTensorInfos.real.dataId,
dtype: xData.complexTensorInfos.real.dtype,
shape: xShape
},
{
dataId: xData.complexTensorInfos.imag.dataId,
dtype: xData.complexTensorInfos.imag.dtype,
shape: xShape
}
];
const realPart = backend.runWebGLProgram(realProgram, inputs, 'float32');
const imagPart = backend.runWebGLProgram(imagProgram, inputs, 'float32');
const complexOutput = complex({ inputs: { real: realPart, imag: imagPart }, backend });
backend.disposeIntermediateTensorInfo(realPart);
backend.disposeIntermediateTensorInfo(imagPart);
const complexOutputReshaped = reshape$1({ inputs: { x: complexOutput }, backend, attrs: { shape: x.shape } });
backend.disposeIntermediateTensorInfo(input2D);
backend.disposeIntermediateTensorInfo(complexOutput);
return complexOutputReshaped;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fft$1(args) {
const { inputs, backend } = args;
const { input } = inputs;
return fftImpl$1(input, false /* inverse */, backend);
}
const fftConfig$1 = {
kernelName: FFT,
backendName: 'webgl',
kernelFunc: fft$1
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class FillProgram {
constructor(shape, value) {
this.outputShape = [];
this.customUniforms = [{ name: 'value', type: 'float' }];
this.variableNames = ['x'];
this.outputShape = shape;
this.userCode = `
void main() {
// Input can be obtained from uniform value.
setOutput(value);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fill$1(args) {
const { backend, attrs } = args;
const { shape, value } = attrs;
let { dtype } = attrs;
dtype = dtype || inferDtype(value);
if (dtype === 'string') {
// String type should be handled in CPU memory.
const values = getArrayFromDType(dtype, sizeFromShape(shape));
values.fill(value);
return backend.makeTensorInfo(shape, dtype, values);
}
else {
const program = new FillProgram(shape, value);
const customValues = [[value]];
return backend.runWebGLProgram(program, [], dtype, customValues);
}
}
const fillConfig$1 = {
kernelName: Fill,
backendName: 'webgl',
kernelFunc: fill$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class FlipLeftRightProgram {
constructor(imageShape) {
this.variableNames = ['Image'];
this.outputShape = [];
const imageWidth = imageShape[2];
this.outputShape = imageShape;
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int x = coords[2];
int coordX = ${imageWidth} - x - 1;
float outputValue;
if(coordX >= 0 && coordX < ${imageWidth}) {
outputValue = getImage(coords[0], coords[1], coordX, coords[3]);
} else {
outputValue = getImage(coords[0], coords[1], coords[2], coords[3]);
}
setOutput(outputValue);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const flipLeftRightConfig$1 = {
kernelName: FlipLeftRight,
backendName: 'webgl',
kernelFunc: ({ inputs, backend }) => {
const { image } = inputs;
const webglBackend = backend;
const program = new FlipLeftRightProgram(image.shape);
const output = webglBackend.runWebGLProgram(program, [image], image.dtype);
return output;
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const FLOOR = `return floor(x);`;
const floor = unaryKernelFunc({ opSnippet: FLOOR, packedOpSnippet: FLOOR, cpuKernelImpl: floorImplCPU });
const floorConfig = {
kernelName: Floor,
backendName: 'webgl',
kernelFunc: floor,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// We use native integer division to deal with floating point imprecision. Since
// we implement floor division and glsl implements truncated division, we
// correct for this by subtracting 1 from result when the result is negative and
// there is a remainder.
const INT_DIV = `
float s = sign(a) * sign(b);
int ia = round(a);
int ib = round(b);
if (ib != 0) {
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
return float(idiv(ia, ib, s));
} else {
return NAN;
}
`;
const INT_DIV_PACKED = `
ivec4 ia = round(a);
ivec4 ib = round(b);
bvec4 cond = notEqual(ib, ivec4(0));
ivec4 result = ivec4(0);
vec4 s = sign(a) * sign(b);
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
if (cond[0]) {
result[0] = idiv(ia[0], ib[0], s[0]);
}
if (cond[1]) {
result[1] = idiv(ia[1], ib[1], s[1]);
}
if (cond[2]) {
result[2] = idiv(ia[2], ib[2], s[2]);
}
if (cond[3]) {
result[3] = idiv(ia[3], ib[3], s[3]);
}
return vec4(result);
`;
const floorDiv = binaryKernelFunc({ opSnippet: INT_DIV, packedOpSnippet: INT_DIV_PACKED, dtype: 'int32' });
const floorDivConfig = {
kernelName: FloorDiv,
backendName: 'webgl',
kernelFunc: floorDiv
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class FromPixelsProgram {
constructor(outputShape) {
this.variableNames = ['A'];
const glsl = getGlslDifferences();
const [height, width,] = outputShape;
this.outputShape = outputShape;
this.userCode = `
void main() {
ivec3 coords = getOutputCoords();
int texR = coords[0];
int texC = coords[1];
int depth = coords[2];
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${width}.0, ${height}.0);
vec4 values = ${glsl.texture2D}(A, uv);
float value;
if (depth == 0) {
value = values.r;
} else if (depth == 1) {
value = values.g;
} else if (depth == 2) {
value = values.b;
} else if (depth == 3) {
value = values.a;
}
setOutput(floor(value * 255.0 + 0.5));
}
`;
}
}
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class FromPixelsPackedProgram {
constructor(outputShape) {
this.variableNames = ['A'];
this.packedInputs = false;
this.packedOutput = true;
const glsl = getGlslDifferences();
const [height, width,] = outputShape;
this.outputShape = outputShape;
this.userCode = `
void main() {
ivec3 coords = getOutputCoords();
int texR = coords[0];
int texC = coords[1];
int depth = coords[2];
vec4 result = vec4(0.);
for(int row=0; row<=1; row++) {
for(int col=0; col<=1; col++) {
texC = coords[1] + row;
depth = coords[2] + col;
vec2 uv = (vec2(texC, texR) + halfCR) /
vec2(${width}.0, ${height}.0);
vec4 values = ${glsl.texture2D}(A, uv);
float value;
if (depth == 0) {
value = values.r;
} else if (depth == 1) {
value = values.g;
} else if (depth == 2) {
value = values.b;
} else if (depth == 3) {
value = values.a;
}
result[row * 2 + col] = floor(value * 255.0 + 0.5);
}
}
${glsl.output} = result;
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const fromPixelsConfig = {
kernelName: FromPixels,
backendName: 'webgl',
kernelFunc: fromPixels,
};
let fromPixels2DContext;
let willReadFrequently = env().getBool('CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU');
function fromPixels(args) {
const { inputs, backend, attrs } = args;
let { pixels } = inputs;
const { numChannels } = attrs;
const isVideo = typeof (HTMLVideoElement) !== 'undefined' &&
pixels instanceof HTMLVideoElement;
const isImage = typeof (HTMLImageElement) !== 'undefined' &&
pixels instanceof HTMLImageElement;
const [width, height] = isVideo ?
[
pixels.videoWidth,
pixels.videoHeight
] :
[pixels.width, pixels.height];
const texShape = [height, width];
const outShape = [height, width, numChannels];
if (isImage || isVideo) {
const newWillReadFrequently = env().getBool('CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU');
if (fromPixels2DContext == null ||
newWillReadFrequently !== willReadFrequently) {
willReadFrequently = newWillReadFrequently;
fromPixels2DContext =
document.createElement('canvas').getContext('2d', { willReadFrequently });
}
fromPixels2DContext.canvas.width = width;
fromPixels2DContext.canvas.height = height;
fromPixels2DContext.drawImage(pixels, 0, 0, width, height);
pixels = fromPixels2DContext.canvas;
}
const tempPixelHandle = backend.makeTensorInfo(texShape, 'int32');
// This is a byte texture with pixels.
backend.texData.get(tempPixelHandle.dataId).usage = TextureUsage.PIXELS;
backend.gpgpu.uploadPixelDataToTexture(backend.getTexture(tempPixelHandle.dataId), pixels);
const program = env().getBool('WEBGL_PACK') ?
new FromPixelsPackedProgram(outShape) :
new FromPixelsProgram(outShape);
const res = backend.runWebGLProgram(program, [tempPixelHandle], 'int32');
backend.disposeData(tempPixelHandle.dataId);
return res;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fusedConv2d(args) {
const { inputs, backend, attrs } = args;
const { x, filter, bias, preluActivationWeights } = inputs;
const { strides, pad, dataFormat, dilations, dimRoundingMode, activation, leakyreluAlpha } = attrs;
const $dataFormat = convertConv2DDataFormat(dataFormat);
const convInfo = computeConv2DInfo(x.shape, filter.shape, strides, dilations, pad, dimRoundingMode, false /* depthwise */, $dataFormat);
let out;
const intermediates = [];
const hasBias = bias != null;
const hasPreluActivationWeights = preluActivationWeights != null;
const hasLeakyreluAlpha = activation === 'leakyrelu';
const prepareInputs = () => {
const inputs = [x, filter];
// If the input is a 1-D tensor, align it with the channels.
//
// For fusedConv2d, the inputs (x, W, bias, preluActivationWeights) are
// supposed to be aligned with the dataFormat. The 4-D tensor inputs or
// scalar inputs are originally aligned, but the 1-D tensor inputs are
// supposed to be aligned with the channels (only bias and PReLU activation
// weights could be a 1-D tensor).
const alignInputWithDataFormat = (input, dataFormat) => {
if (dataFormat === 'NCHW' && input.shape.length === 1 &&
input.shape[0] !== 1) {
const alignedInput = reshape$1({
inputs: { x: input },
backend,
attrs: { shape: [input.shape[0], 1, 1] }
});
intermediates.push(alignedInput);
return alignedInput;
}
return input;
};
if (hasBias) {
inputs.push(alignInputWithDataFormat(bias, dataFormat));
}
if (hasPreluActivationWeights) {
inputs.push(alignInputWithDataFormat(preluActivationWeights, dataFormat));
}
if (hasLeakyreluAlpha) {
const $leakyreluAlpha = backend.makeTensorInfo([], 'float32', createScalarValue(leakyreluAlpha, 'float32'));
inputs.push($leakyreluAlpha);
intermediates.push($leakyreluAlpha);
}
return inputs;
};
if (convInfo.filterHeight === 1 && convInfo.filterWidth === 1 &&
convInfo.dilationHeight === 1 && convInfo.dilationWidth === 1 &&
convInfo.strideHeight === 1 && convInfo.strideWidth === 1 &&
(convInfo.padInfo.type === 'SAME' || convInfo.padInfo.type === 'VALID')) {
out = conv2dByMatMul({
x,
filter,
convInfo,
backend,
bias,
activation,
preluActivationWeights,
leakyreluAlpha
});
}
else if (convInfo.strideWidth <= 2 && $dataFormat === 'channelsLast'
&& env().getBool('WEBGL_EXP_CONV')) {
const fusedActivation = activation ? mapActivationToShaderProgram(activation, true) : null;
const program = new Conv2DPackedProgram(convInfo, hasBias, fusedActivation, hasPreluActivationWeights, hasLeakyreluAlpha);
const customValues = [
[convInfo.padInfo.top, convInfo.padInfo.left],
[convInfo.strideHeight, convInfo.strideWidth],
[convInfo.dilationHeight, convInfo.dilationWidth],
[convInfo.inHeight, convInfo.inWidth]
];
const inputs = prepareInputs();
out = backend.runWebGLProgram(program, inputs, 'float32', customValues);
}
else if (env().getBool('WEBGL_CONV_IM2COL')) {
out = conv2dWithIm2Row({
x,
filter,
convInfo,
backend,
bias,
activation,
preluActivationWeights,
leakyreluAlpha
});
}
else {
const fusedActivation = activation ? mapActivationToShaderProgram(activation, false) : null;
const program = new Conv2DProgram(convInfo, hasBias, fusedActivation, hasPreluActivationWeights, hasLeakyreluAlpha);
const inputs = prepareInputs();
out = backend.runWebGLProgram(program, inputs, 'float32');
}
const outReshaped = reshape$1({ inputs: { x: out }, backend, attrs: { shape: convInfo.outShape } });
intermediates.push(out);
intermediates.forEach(t => backend.disposeIntermediateTensorInfo(t));
return outReshaped;
}
const fusedConv2DConfig$1 = {
kernelName: FusedConv2D,
backendName: 'webgl',
kernelFunc: fusedConv2d,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fusedDepthwiseConv2D$1(args) {
const { inputs, backend, attrs } = args;
const { x, filter, bias, preluActivationWeights } = inputs;
const { strides, pad, dilations, dimRoundingMode, activation, leakyreluAlpha } = attrs;
const intermediates = [];
let $dilations = dilations;
if ($dilations == null) {
$dilations = [1, 1];
}
assert$1(eitherStridesOrDilationsAreOne(strides, $dilations), () => 'Error in depthwiseConv2d: Either strides or dilations must be ' +
`1. Got strides ${strides} and dilations '${$dilations}'`);
const convInfo = computeConv2DInfo(x.shape, filter.shape, strides, $dilations, pad, dimRoundingMode, true /* depthwise */);
const shouldPackDepthwiseConv = env().getBool('WEBGL_PACK_DEPTHWISECONV') &&
convInfo.strideWidth <= 2 &&
convInfo.outChannels / convInfo.inChannels === 1;
const fusedActivation = activation ?
mapActivationToShaderProgram(activation, shouldPackDepthwiseConv) :
null;
const programInputs = [x, filter];
const hasBias = bias != null;
const hasPreluActivationWeights = preluActivationWeights != null;
const hasLeakyreluAlpha = activation === 'leakyrelu';
if (hasBias) {
programInputs.push(bias);
}
if (hasPreluActivationWeights) {
programInputs.push(preluActivationWeights);
}
if (hasLeakyreluAlpha) {
const $leakyreluAlpha = backend.makeTensorInfo([], 'float32', createScalarValue(leakyreluAlpha, 'float32'));
programInputs.push($leakyreluAlpha);
intermediates.push($leakyreluAlpha);
}
let program;
if (shouldPackDepthwiseConv) {
program = new DepthwiseConvPacked2DProgram(convInfo, hasBias, fusedActivation, hasPreluActivationWeights, hasLeakyreluAlpha);
}
else {
program = new DepthwiseConv2DProgram(convInfo, hasBias, fusedActivation, hasPreluActivationWeights, hasLeakyreluAlpha);
}
const customValues = [
[convInfo.padInfo.top, convInfo.padInfo.left],
[convInfo.strideHeight, convInfo.strideWidth],
[convInfo.dilationHeight, convInfo.dilationWidth],
[convInfo.inHeight, convInfo.inWidth]
];
const result = backend.runWebGLProgram(program, programInputs, 'float32', customValues);
intermediates.forEach(t => backend.disposeIntermediateTensorInfo(t));
return result;
}
const fusedDepthwiseConv2DConfig$1 = {
kernelName: FusedDepthwiseConv2D,
backendName: 'webgl',
kernelFunc: fusedDepthwiseConv2D$1,
};
class GatherNDProgram {
constructor(sliceDim, strides, shape, paramsShape) {
this.sliceDim = sliceDim;
this.strides = strides;
this.paramsShape = paramsShape;
this.variableNames = ['x', 'indices'];
this.outputShape = shape;
const dtype = getCoordsDataType(shape.length);
let mainLoop = `
int index;`;
for (let j = 0; j < this.sliceDim; j++) {
mainLoop += `
index = round(getIndices(coords[0], ${j}));
out_of_bounds = out_of_bounds || index < 0;
out_of_bounds = out_of_bounds || index >= ${this.paramsShape[j]};
flattenIndex += index * ${this.strides[j]};`;
}
this.userCode = `
void main() {
${dtype} coords = getOutputCoords();
int flattenIndex = 0;
bool out_of_bounds = false;
${mainLoop}
setOutput(out_of_bounds ? 0.0 : getX(flattenIndex, coords[1]));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function gatherNd$1(args) {
const { inputs, backend } = args;
const { params, indices } = inputs;
const indicesShape = indices.shape;
const sliceRank = indicesShape[indicesShape.length - 1];
const paramsSize = sizeFromShape(params.shape);
const [resultShape, numSlices, sliceSize, strides] = prepareAndValidate(params, indices);
const flattenIndices = reshape$1({ inputs: { x: indices }, backend, attrs: { shape: [numSlices, sliceRank] } });
const flattenX = reshape$1({
inputs: { x: params },
backend,
attrs: { shape: [(sizeFromShape(params.shape) / sliceSize), sliceSize] }
});
if (backend.shouldExecuteOnCPU([params, indices]) ||
params.dtype === 'string') {
const indicesData = backend.readSync(indices.dataId);
const paramsBuf = backend.bufferSync(params);
const outValue = gatherNdImplCPU(indicesData, paramsBuf, params.dtype, numSlices, sliceRank, sliceSize, strides, params.shape, paramsSize);
return backend.makeTensorInfo(resultShape, params.dtype, outValue.values);
}
const program = new GatherNDProgram(sliceRank, strides, [numSlices, sliceSize], params.shape);
const res = backend.runWebGLProgram(program, [flattenX, flattenIndices], flattenX.dtype);
const reshaped = reshape$1({ inputs: { x: res }, backend, attrs: { shape: resultShape } });
backend.disposeIntermediateTensorInfo(flattenIndices);
backend.disposeIntermediateTensorInfo(flattenX);
backend.disposeIntermediateTensorInfo(res);
return reshaped;
}
const gatherNdConfig$1 = {
kernelName: GatherNd,
backendName: 'webgl',
kernelFunc: gatherNd$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class GatherProgram {
constructor(aShape, outputShape) {
this.variableNames = ['A', 'indices'];
this.outputShape = outputShape;
this.rank = outputShape.length;
const dtype = getCoordsDataType(this.rank);
const sourceCoords = getSourceCoords$1(aShape);
this.userCode = `
void main() {
${dtype} resRC = getOutputCoords();
int index = int(getIndices(resRC.x, resRC.z));
float inBounds = (index >= 0) && (index < ${aShape[2]}) ? 1.0 : 0.0;
setOutput(inBounds * getA(${sourceCoords}));
}
`;
}
}
// The input and output are always flattened into rank 4 tensors.
function getSourceCoords$1(aShape, axis) {
const currentCoords = ['resRC.x', 'resRC.y', 'resRC.z', 'resRC.w'];
const sourceCoords = [];
for (let i = 0; i < aShape.length; i++) {
if (i === 2) {
sourceCoords.push('index');
}
else {
sourceCoords.push(`${currentCoords[i]}`);
}
}
return sourceCoords.join();
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function gatherV2$1(args) {
const { inputs, backend, attrs } = args;
const { x, indices } = inputs;
const { axis, batchDims } = attrs;
const parsedAxis = parseAxisParam(axis, x.shape)[0];
if (env().get('DEBUG')) {
// In debug mode, throw error when any index is out of bound.
// Otherwise, just fill out of bounds with zeroes.
const indicesVals = backend.readSync(indices.dataId);
const axisDim = x.shape[parsedAxis];
for (let i = 0; i < indicesVals.length; ++i) {
const index = indicesVals[i];
assert$1(index <= axisDim - 1 && index >= 0, () => `GatherV2: the index value ${index} is not in [0, ${axisDim - 1}]`);
}
}
const shapeInfo = collectGatherOpShapeInfo(x, indices, parsedAxis, batchDims);
const indicesSize = sizeFromShape(indices.shape);
const toDispose = [];
const flattenX = reshape$1({
inputs: { x },
backend,
attrs: {
shape: [
shapeInfo.batchSize, shapeInfo.outerSize, shapeInfo.dimSize,
shapeInfo.sliceSize
]
}
});
const flattenIndex = reshape$1({
inputs: { x: indices },
backend,
attrs: { shape: [shapeInfo.batchSize, indicesSize / shapeInfo.batchSize] }
});
toDispose.push(flattenX);
toDispose.push(flattenIndex);
const flattenOutputShape = [
shapeInfo.batchSize, shapeInfo.outerSize, indicesSize / shapeInfo.batchSize,
shapeInfo.sliceSize
];
if (backend.shouldExecuteOnCPU([x, indices]) || x.dtype === 'string') {
const indicesBuf = backend.bufferSync(flattenIndex);
const xBuf = backend.bufferSync(flattenX);
const outBuf = gatherV2ImplCPU(xBuf, indicesBuf, flattenOutputShape);
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return backend.makeTensorInfo(shapeInfo.outputShape, outBuf.dtype, outBuf.values);
}
const program = new GatherProgram(flattenX.shape, flattenOutputShape);
const res = backend.runWebGLProgram(program, [flattenX, flattenIndex], flattenX.dtype);
toDispose.push(res);
const reshaped = reshape$1({ inputs: { x: res }, backend, attrs: { shape: shapeInfo.outputShape } });
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return reshaped;
}
const gatherV2Config$1 = {
kernelName: GatherV2,
backendName: 'webgl',
kernelFunc: gatherV2$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const GREATER = `return float(a > b);`;
const GREATER_PACKED = `
return vec4(greaterThan(a, b));
`;
const greater = binaryKernelFunc({
opSnippet: GREATER,
packedOpSnippet: GREATER_PACKED,
cpuKernelImpl: greaterImplCPU,
dtype: 'bool'
});
const greaterConfig = {
kernelName: Greater,
backendName: 'webgl',
kernelFunc: greater
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const GREATER_EQUAL = `return float(a >= b);`;
const GREATER_EQUAL_PACKED = `
return vec4(greaterThanEqual(a, b));
`;
const greaterEqual = binaryKernelFunc({
opSnippet: GREATER_EQUAL,
packedOpSnippet: GREATER_EQUAL_PACKED,
dtype: 'bool',
cpuKernelImpl: greaterEqualImplCPU
});
const greaterEqualConfig = {
kernelName: GreaterEqual,
backendName: 'webgl',
kernelFunc: greaterEqual
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function ifft$1(args) {
const { inputs, backend } = args;
const { input } = inputs;
return fftImpl$1(input, true /* inverse */, backend);
}
const ifftConfig$1 = {
kernelName: IFFT,
backendName: 'webgl',
kernelFunc: ifft$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const IS_FINITE = `return float(!isnan(x) && !isinf(x));`;
const isFinite$2 = unaryKernelFunc({ opSnippet: IS_FINITE, dtype: 'bool' });
const isFiniteConfig$1 = {
kernelName: IsFinite,
backendName: 'webgl',
kernelFunc: isFinite$2,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const IS_INF = `return float(isinf(x));`;
const isInf$1 = unaryKernelFunc({ opSnippet: IS_INF, dtype: 'bool' });
const isInfConfig$1 = {
kernelName: IsInf,
backendName: 'webgl',
kernelFunc: isInf$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const IS_NAN = `return float(isnan(x));`;
const isNaN$2 = unaryKernelFunc({ opSnippet: IS_NAN, dtype: 'bool' });
const isNaNConfig$1 = {
kernelName: IsNan,
backendName: 'webgl',
kernelFunc: isNaN$2,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LESS = `return float(a < b);`;
const LESS_PACKED = `
return vec4(lessThan(a, b));
`;
const less = binaryKernelFunc({
opSnippet: LESS,
packedOpSnippet: LESS_PACKED,
cpuKernelImpl: lessImplCPU,
dtype: 'bool'
});
const lessConfig = {
kernelName: Less,
backendName: 'webgl',
kernelFunc: less
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LESS_EQUAL = `return float(a <= b);`;
const LESS_EQUAL_PACKED = `
return vec4(lessThanEqual(a, b));
`;
const lessEqual = binaryKernelFunc({
opSnippet: LESS_EQUAL,
packedOpSnippet: LESS_EQUAL_PACKED,
cpuKernelImpl: lessEqualImplCPU,
dtype: 'bool'
});
const lessEqualConfig = {
kernelName: LessEqual,
backendName: 'webgl',
kernelFunc: lessEqual
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function linSpace$1(args) {
const { backend, attrs } = args;
const { start, stop, num } = attrs;
// TODO: Use CPU implementation due to the precision problem in Safari.
const outVals = linSpaceImplCPU(start, stop, num);
return backend.makeTensorInfo([outVals.length], 'float32', outVals);
}
const linSpaceConfig$1 = {
kernelName: LinSpace,
backendName: 'webgl',
kernelFunc: linSpace$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Windows chrome return 0 if the input is negative value. We will specifically
// return NaN if the input is 0 to solve compatiblity issue.
const LOG = CHECK_NAN_SNIPPET_UNARY + `
return x < 0.0 ? 0./0. : log(x);
`;
const LOG_PACKED = `
vec4 result = log(x);
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : (x.r < 0.0 ? 0./0. : result.r);
result.g = isNaN.g ? x.g : (x.g < 0.0 ? 0./0. : result.g);
result.b = isNaN.b ? x.b : (x.b < 0.0 ? 0./0. : result.b);
result.a = isNaN.a ? x.a : (x.a < 0.0 ? 0./0. : result.a);
return result;
`;
const log = unaryKernelFunc({ opSnippet: LOG, packedOpSnippet: LOG_PACKED, cpuKernelImpl: logImplCPU });
const logConfig = {
kernelName: Log,
backendName: 'webgl',
kernelFunc: log
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LOG1P = CHECK_NAN_SNIPPET_UNARY + `
return log(1.0 + x);
`;
const log1p$1 = unaryKernelFunc({ opSnippet: LOG1P });
const log1pConfig$1 = {
kernelName: Log1p,
backendName: 'webgl',
kernelFunc: log1p$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LOGICAL_AND = `return float(a >= 1.0 && b >= 1.0);`;
const LOGICAL_AND_PACKED = `
return vec4(
vec4(greaterThanEqual(a, vec4(1.0))) *
vec4(greaterThanEqual(b, vec4(1.0))));
`;
const logicalAnd$1 = binaryKernelFunc({
opSnippet: LOGICAL_AND,
packedOpSnippet: LOGICAL_AND_PACKED,
dtype: 'bool'
});
const logicalAndConfig$1 = {
kernelName: LogicalAnd,
backendName: 'webgl',
kernelFunc: logicalAnd$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LOGICAL_NOT = `return float(!(x >= 1.0));`;
const logicalNot$1 = unaryKernelFunc({ opSnippet: LOGICAL_NOT });
const logicalNotConfig$1 = {
kernelName: LogicalNot,
backendName: 'webgl',
kernelFunc: logicalNot$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const LOGICAL_OR = `return float(a >= 1.0 || b >= 1.0);`;
const LOGICAL_OR_PACKED = `
return min(
vec4(greaterThanEqual(a, vec4(1.0))) +
vec4(greaterThanEqual(b, vec4(1.0))),
vec4(1.0));
`;
const logicalOr$1 = binaryKernelFunc({ opSnippet: LOGICAL_OR, packedOpSnippet: LOGICAL_OR_PACKED, dtype: 'bool' });
const logicalOrConfig$1 = {
kernelName: LogicalOr,
backendName: 'webgl',
kernelFunc: logicalOr$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class LRNProgram {
constructor(xShape, radius, bias, alpha, beta) {
this.variableNames = ['x'];
this.outputShape = [];
const rad = radius;
const maxD = xShape[3] - 1;
this.outputShape = xShape;
// optimize pow(bias + alpha * sum, -beta)
// src: https://github.com/tensorflow/tensorflow/..
// blob/26033a1644a9c4a5fbe3170ab2e864b6a4ccd4ca/..
// tensorflow/core/kernels/mkl_lrn_op.cc#L320
let powOperator;
const basis = `float(${bias}) + float(${alpha}) * sum`;
if (beta === 0.5) {
powOperator = `inversesqrt(${basis})`;
}
else if (beta === 1.0) {
powOperator = `1.0/(${basis})`;
}
else {
powOperator = `exp(log(${basis}) * float(-${beta}));`;
}
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int r = coords[1];
int c = coords[2];
int d = coords[3];
float x = getX(b, r, c, d);
float sum = 0.0;
for (int j = -${rad}; j <= ${rad}; j++) {
int idx = d + j;
if (idx >= 0 && idx <= ${maxD}) {
float z = getX(b, r, c, idx);
sum += z * z;
}
}
float val = x * ${powOperator};
setOutput(val);
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class LRNPackedProgram {
constructor(xShape, radius, bias, alpha, beta) {
this.variableNames = ['x'];
this.outputShape = [];
this.packedInputs = true;
this.packedOutput = true;
const rad = radius;
const maxD = xShape[3] - 1;
this.outputShape = xShape;
// optimize pow(bias + alpha * sum, -beta)
// src: https://github.com/tensorflow/tensorflow/..
// blob/26033a1644a9c4a5fbe3170ab2e864b6a4ccd4ca/..
// tensorflow/core/kernels/mkl_lrn_op.cc#L320
let powOperator;
const basis = `float(${bias}) + float(${alpha}) * sum`;
if (beta === 0.5) {
powOperator = `inversesqrt(${basis})`;
}
else if (beta === 1.0) {
powOperator = `1.0/(${basis})`;
}
else {
powOperator = `exp(log(${basis}) * float(-${beta}));`;
}
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int b = coords.x;
int r = coords.y;
int c = coords.z;
int d = coords.w;
bool hasNextCol = d < ${this.outputShape[3]};
bool hasNextRow = c < ${this.outputShape[2]};
vec4 sum = vec4(0.);
vec4 xFragAtOutputCoords = getX(b, r, c, d);
vec4 xAtOutputCoords = vec4(
getChannel(xFragAtOutputCoords, vec2(c, d)),
hasNextCol ?
getChannel(xFragAtOutputCoords, vec2(c, d + 1)) : 0.0,
hasNextRow ?
getChannel(xFragAtOutputCoords , vec2(c + 1, d)) : 0.0,
(hasNextRow && hasNextCol) ?
getChannel(xFragAtOutputCoords, vec2(c + 1, d + 1)) : 0.0
);
int firstChannel = d - ${rad};
vec2 cache = vec2(0.);
if(firstChannel >= 0){
vec4 firstChannelFrag = getX(b, r, c, firstChannel);
cache.x = getChannel(firstChannelFrag, vec2(c, firstChannel));
if(hasNextRow){
cache.y = getChannel(firstChannelFrag, vec2(c + 1, firstChannel));
}
}
ivec2 depth = ivec2(d, d + 1);
for (int j = - ${rad}; j <= ${rad}; j++) {
ivec2 idx = depth + j;
bvec2 aboveLowerBound = greaterThanEqual(idx, ivec2(0));
bvec2 belowUpperBound = lessThanEqual(idx, ivec2(${maxD}));
bool depthInRange = aboveLowerBound.x && belowUpperBound.x;
bool depthPlusOneInRange = aboveLowerBound.y && belowUpperBound.y;
if(depthInRange || depthPlusOneInRange){
vec4 z = vec4(0.);
vec4 xFragAtCurrentDepth;
z.xz = cache.xy;
if(depthPlusOneInRange && hasNextCol){
xFragAtCurrentDepth = idx.y != d ?
getX(b, r, c, idx.y) : xFragAtOutputCoords;
z.y = getChannel(xFragAtCurrentDepth, vec2(c, idx.y));
if(hasNextRow){
z.w = getChannel(xFragAtCurrentDepth, vec2(c + 1, idx.y));
}
}
cache.xy = z.yw;
sum += z * z;
}
}
vec4 result = xAtOutputCoords * ${powOperator};
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const lrn = (args) => {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { depthRadius, bias, alpha, beta } = attrs;
const program = env().getBool('WEBGL_PACK_NORMALIZATION') ?
new LRNPackedProgram(x.shape, depthRadius, bias, alpha, beta) :
new LRNProgram(x.shape, depthRadius, bias, alpha, beta);
return backend.runWebGLProgram(program, [x], x.dtype);
};
// tslint:disable-next-line: variable-name
const LRNConfig$1 = {
kernelName: LRN,
backendName: 'webgl',
kernelFunc: lrn
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class LRNGradProgram {
constructor(inputShape, depthRadius, bias, alpha, beta) {
this.variableNames = ['inputImage', 'outputImage', 'dy'];
this.outputShape = [];
this.outputShape = inputShape;
this.depth = inputShape[3];
this.depthRadius = depthRadius;
this.bias = bias;
this.alpha = alpha;
this.beta = beta;
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int r = coords[1];
int c = coords[2];
float result = 0.0;
for (int d = 0; d < ${this.depth}; ++d) {
int depthBegin = int(max(0.0, float(d - ${depthRadius})));
int depthEnd = int(min(float(${this.depth}),
float(d + ${depthRadius} + 1)));
const int MIN_DEPTH_BEGIN = 0;
const int MAX_DEPTH_END = ${this.depth};
float norm = 0.0;
for (int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k) {
if (k < depthBegin){
continue;
}
else if (k >= depthBegin && k < depthEnd) {
norm += getInputImage(b, r, c, k) * getInputImage(b, r, c, k);
}
else {
break;
}
}
norm = float(${alpha}) * norm + float(${bias});
for(int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k){
if (k < depthBegin){
continue;
}
else if (k >= depthBegin && k < depthEnd){
float dyi = -2.0 * float(${alpha})
* float(${beta})
* getInputImage(b, r, c, k) * getOutputImage(b, r, c, d)
/ norm;
if (k == d) {
dyi += pow(norm, -1.0 * ${beta});
}
if (k == coords[3]) {
dyi *= getDy(b, r, c, d);
result += dyi;
}
}
else {
break;
}
}
}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const lrnGrad = (args) => {
const { inputs, backend, attrs } = args;
const { x, y, dy } = inputs;
const { depthRadius, bias, alpha, beta } = attrs;
const program = new LRNGradProgram(x.shape, depthRadius, bias, alpha, beta);
return backend.runWebGLProgram(program, [x, y, dy], x.dtype);
};
// tslint:disable-next-line: variable-name
const LRNGradConfig$1 = {
kernelName: LRNGrad,
backendName: 'webgl',
kernelFunc: lrnGrad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxImpl(x, reduceShape, outShape, backend) {
const inSize = sizeFromShape(reduceShape);
const xSize = sizeFromShape(x.shape);
const batchSize = xSize / inSize;
const reshapedInput = reshape$1({ inputs: { x }, attrs: { shape: [batchSize, inSize] }, backend });
const reduced = reduce(reshapedInput, x.dtype, 'max', backend);
const reshapedOutput = reshape$1({ inputs: { x: reduced }, attrs: { shape: outShape }, backend });
backend.disposeIntermediateTensorInfo(reshapedInput);
backend.disposeIntermediateTensorInfo(reduced);
return reshapedOutput;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function max$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { reductionIndices, keepDims } = attrs;
const xRank = x.shape.length;
const origAxes = parseAxisParam(reductionIndices, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
const maxInputIsTransposed = permutedAxes != null;
const shouldExecuteOnCPU = backend.shouldExecuteOnCPU([x]);
let maxInput = x;
if (maxInputIsTransposed) {
if (shouldExecuteOnCPU) {
const xTexData = backend.texData.get(maxInput.dataId);
const values = xTexData.values;
const newShape = new Array(xRank);
for (let i = 0; i < newShape.length; i++) {
newShape[i] = x.shape[permutedAxes[i]];
}
const maxInputValues = transposeImplCPU(values, x.shape, x.dtype, permutedAxes, newShape);
maxInput = backend.makeTensorInfo(newShape, x.dtype);
const maxInputData = backend.texData.get(maxInput.dataId);
maxInputData.values = maxInputValues;
}
else {
maxInput = transposeImpl(x, permutedAxes, backend);
}
axes = getInnerMostAxes(axes.length, xRank);
}
assertAxesAreInnerMostDims('max', axes, xRank);
const [maxOutShape, reduceShape] = computeOutAndReduceShapes(maxInput.shape, axes);
let outShape = maxOutShape;
if (keepDims) {
// rather than reshape at the end, set the target shape here.
outShape = expandShapeToKeepDim(maxOutShape, origAxes);
}
let out;
if (shouldExecuteOnCPU) {
const xTexData = backend.texData.get(maxInput.dataId);
const values = xTexData.values;
const outValues = maxImplCPU(values, sizeFromShape(reduceShape), outShape, x.dtype);
out = backend.makeTensorInfo(outShape, x.dtype);
const outData = backend.texData.get(out.dataId);
outData.values = outValues;
}
else {
out = maxImpl(maxInput, reduceShape, outShape, backend);
}
if (maxInputIsTransposed) {
backend.disposeIntermediateTensorInfo(maxInput);
}
return out;
}
const maxConfig$1 = {
kernelName: Max,
backendName: 'webgl',
kernelFunc: max$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const MAXIMUM = CHECK_NAN_SNIPPET + `
return max(a, b);
`;
const MAXIMUM_PACKED = `
vec4 result = vec4(max(a, b));
bvec4 isNaNA = isnan(a);
bvec4 isNaNB = isnan(b);
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
` +
CHECK_NAN_SNIPPET_PACKED + `
return result;
`;
const maximum = binaryKernelFunc({
opSnippet: MAXIMUM,
packedOpSnippet: MAXIMUM_PACKED,
cpuKernelImpl: maximumImplCPU
});
const maximumConfig = {
kernelName: Maximum,
backendName: 'webgl',
kernelFunc: maximum
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPool$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
assertNotComplex$1(x, 'maxPool');
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const dilations = 1;
assert$1(eitherStridesOrDilationsAreOne(strides, dilations), () => 'Error in maxPool: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
const convInfo = computePool2DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode);
if (convInfo.filterWidth === 1 && convInfo.filterHeight === 1 &&
arraysEqual(convInfo.inShape, convInfo.outShape)) {
return identity({ inputs: { x }, backend });
}
const maxPoolProgram = new Pool2DProgram(convInfo, 'max', false);
return backend.runWebGLProgram(maxPoolProgram, [x], x.dtype);
}
const maxPoolConfig$1 = {
kernelName: MaxPool,
backendName: 'webgl',
kernelFunc: maxPool$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPool3d(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { filterSize, strides, pad, dataFormat, dimRoundingMode } = attrs;
const dilations = [1, 1, 1];
const convInfo = computePool3DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode, dataFormat);
const maxPoolProgram = new Pool3DProgram(convInfo, 'max', false);
return backend.runWebGLProgram(maxPoolProgram, [x], x.dtype);
}
const maxPool3DConfig$1 = {
kernelName: MaxPool3D,
backendName: 'webgl',
kernelFunc: maxPool3d
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class MaxPool2DBackpropProgram {
constructor(convInfo) {
this.variableNames = ['dy', 'maxPos'];
this.outputShape = convInfo.inShape;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationHeight = convInfo.dilationHeight;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const lastIndex = effectiveFilterHeight * effectiveFilterWidth - 1;
this.userCode = `
const ivec2 pads = ivec2(${padTop}, ${padLeft});
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
ivec2 dyRCCorner = coords.yz - pads;
int dyRCorner = dyRCCorner.x;
int dyCCorner = dyRCCorner.y;
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
float dyR = float(dyRCorner + wR) / ${strideHeight}.0;
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 || fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
for (int wC = 0; wC < ${effectiveFilterWidth}; wC++) {
float dyC = float(dyCCorner + wC) / ${strideWidth}.0;
if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
fract(dyC) > 0.0) {
continue;
}
int idyC = int(dyC);
float dyValue = getDy(b, idyR, idyC, d);
int maxPosValue = ${lastIndex} - int(getMaxPos(b, idyR, idyC, d));
// Get the current value, check it against the value from the
// position matrix.
int curPosValue = wR * ${effectiveFilterWidth} + wC;
float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);
dotProd += dyValue * mask;
}
}
setOutput(dotProd);
}
`;
}
}
class MaxPool3DBackpropProgram {
constructor(convInfo) {
this.variableNames = ['dy', 'maxPos'];
this.outputShape = convInfo.inShape;
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterDepth = convInfo.effectiveFilterDepth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padFront = effectiveFilterDepth - 1 - convInfo.padInfo.front;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const lastIndex = effectiveFilterDepth * effectiveFilterHeight * effectiveFilterWidth - 1;
this.userCode = `
const ivec3 pads = ivec3(${padFront}, ${padTop}, ${padLeft});
void main() {
ivec5 coords = getOutputCoords();
int batch = coords.x;
int ch = coords.u;
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
int dyDCorner = dyCorner.x;
int dyRCorner = dyCorner.y;
int dyCCorner = dyCorner.z;
// Convolve dy(?, ?, ?, ch) with pos mask(:, :, :, d) to get
// dx(xD, xR, xC, ch).
// ? = to be determined. : = across all values in that axis.
float dotProd = 0.0;
for (int wD = 0; wD < ${effectiveFilterDepth};
wD += ${dilationDepth}) {
float dyD = float(dyDCorner + wD) / ${strideDepth}.0;
if (dyD < 0.0 || dyD >= ${convInfo.outDepth}.0 || fract(dyD) > 0.0) {
continue;
}
int idyD = int(dyD);
for (int wR = 0; wR < ${effectiveFilterHeight};
wR += ${dilationHeight}) {
float dyR = float(dyRCorner + wR) / ${strideHeight}.0;
if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 ||
fract(dyR) > 0.0) {
continue;
}
int idyR = int(dyR);
for (int wC = 0; wC < ${effectiveFilterWidth};
wC += ${dilationWidth}) {
float dyC = float(dyCCorner + wC) / ${strideWidth}.0;
if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
fract(dyC) > 0.0) {
continue;
}
int idyC = int(dyC);
float dyValue = getDy(batch, idyD, idyR, idyC, ch);
int maxPosValue = ${lastIndex} -
int(getMaxPos(batch, idyD, idyR, idyC, ch));
// Get the current value, check it against the value from the
// position matrix.
int curPosValue =
wD * ${effectiveFilterHeight} * ${effectiveFilterWidth} +
wR * ${effectiveFilterWidth} + wC;
float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);
dotProd += dyValue * mask;
}
}
}
setOutput(dotProd);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPool3DGrad$1(args) {
const { inputs, backend, attrs } = args;
const { dy, input } = inputs;
const x = input;
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const dilations = [1, 1, 1];
const convInfo = computePool3DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode);
const maxPool3dPositionsProgram = new Pool3DProgram(convInfo, 'max', true /* get positions */);
const maxPool3dPositions = backend.runWebGLProgram(maxPool3dPositionsProgram, [x], x.dtype);
const maxPoolBackpropProgram = new MaxPool3DBackpropProgram(convInfo);
const result = backend.runWebGLProgram(maxPoolBackpropProgram, [dy, maxPool3dPositions], x.dtype);
backend.disposeIntermediateTensorInfo(maxPool3dPositions);
return result;
}
const maxPool3DGradConfig$2 = {
kernelName: MaxPool3DGrad,
backendName: 'webgl',
kernelFunc: maxPool3DGrad$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPoolGrad$2(args) {
const { inputs, backend, attrs } = args;
const { dy, input, output } = inputs;
const x = input;
assertNotComplex$1([input, output], 'maxPoolGrad');
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const convInfo = computePool2DInfo(x.shape, filterSize, strides, 1 /* dilations */, pad, dimRoundingMode);
const getPositions = true;
const maxPoolPositionsProgram = new Pool2DProgram(convInfo, 'max', getPositions);
const maxPoolPositions = backend.runWebGLProgram(maxPoolPositionsProgram, [x], x.dtype);
const maxPoolBackPropProgram = new MaxPool2DBackpropProgram(convInfo);
const result = backend.runWebGLProgram(maxPoolBackPropProgram, [dy, maxPoolPositions], x.dtype);
backend.disposeIntermediateTensorInfo(maxPoolPositions);
return result;
}
const maxPoolGradConfig$2 = {
kernelName: MaxPoolGrad,
backendName: 'webgl',
kernelFunc: maxPoolGrad$2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPoolWithArgmaxImpl$1(x, includeBatchInIndex, convInfo, backend) {
let program = new Pool2DProgram(convInfo, 'max', false);
const poolOutput = backend.runWebGLProgram(program, [x], 'float32');
program = new Pool2DProgram(convInfo, 'max', true, true, includeBatchInIndex);
const indexOutput = backend.runWebGLProgram(program, [x], 'float32');
return [poolOutput, indexOutput];
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const maxPoolWithArgmaxConfig$1 = {
kernelName: MaxPoolWithArgmax,
backendName: 'webgl',
kernelFunc: ({ inputs, attrs, backend }) => {
const { x } = inputs;
const { filterSize, strides, pad, includeBatchInIndex } = attrs;
const webglBackend = backend;
assert$1(x.shape.length === 4, () => `Error in maxPool: input must be rank 4 but got rank ${x.shape.length}.`);
const dilations = [1, 1];
assert$1(eitherStridesOrDilationsAreOne(strides, dilations), () => 'Error in maxPool: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
const convInfo = computePool2DInfo(x.shape, filterSize, strides, dilations, pad);
const [result, indexes] = maxPoolWithArgmaxImpl$1(x, includeBatchInIndex, convInfo, webglBackend);
return [result, indexes];
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function meanImpl(x, reduceShape, outShape, backend) {
const inSize = sizeFromShape(reduceShape);
const xSize = sizeFromShape(x.shape);
const batchSize = xSize / inSize;
const reshapedInput = reshape$1({ inputs: { x }, attrs: { shape: [batchSize, inSize] }, backend });
const reduced = reduce(reshapedInput, 'float32', 'mean', backend);
const reshapedOutput = reshape$1({ inputs: { x: reduced }, attrs: { shape: outShape }, backend });
backend.disposeIntermediateTensorInfo(reshapedInput);
backend.disposeIntermediateTensorInfo(reduced);
return reshapedOutput;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const meanConfig$1 = {
kernelName: Mean,
backendName: 'webgl',
kernelFunc: ({ inputs, attrs, backend }) => {
const { x } = inputs;
const { keepDims, axis } = attrs;
const webglBackend = backend;
const xRank = x.shape.length;
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
const meanInputIsTransposed = permutedAxes != null;
const shouldExecuteOnCPU = webglBackend.shouldExecuteOnCPU([x]);
const intermediates = [];
let meanInput = x;
if (meanInputIsTransposed) {
if (shouldExecuteOnCPU) {
const xTexData = webglBackend.texData.get(meanInput.dataId);
const values = xTexData.values;
const newShape = new Array(xRank);
for (let i = 0; i < newShape.length; i++) {
newShape[i] = x.shape[permutedAxes[i]];
}
const meanInputValues = transposeImplCPU(values, x.shape, x.dtype, permutedAxes, newShape);
meanInput = webglBackend.makeTensorInfo(newShape, x.dtype);
const meanInputData = webglBackend.texData.get(meanInput.dataId);
meanInputData.values = meanInputValues;
}
else {
meanInput = transposeImpl(x, permutedAxes, webglBackend);
}
intermediates.push(meanInput);
axes = getInnerMostAxes(axes.length, xRank);
}
assertAxesAreInnerMostDims('sum', axes, xRank);
const [meanOutShape, reduceShape] = computeOutAndReduceShapes(meanInput.shape, axes);
let outShape = meanOutShape;
if (keepDims) {
// rather than reshape at the end, set the target shape here.
outShape = expandShapeToKeepDim(meanOutShape, origAxes);
}
const out = meanImpl(meanInput, reduceShape, outShape, webglBackend);
for (const i of intermediates) {
webglBackend.disposeIntermediateTensorInfo(i);
}
return out;
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function min$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
const xRank = x.shape.length;
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
let permutedX = x;
if (permutedAxes != null) {
permutedX = transpose({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
axes = getInnerMostAxes(axes.length, x.shape.length);
}
assertAxesAreInnerMostDims('min', axes, xRank);
const [outShape, reduceShape] = computeOutAndReduceShapes(permutedX.shape, axes);
const inSize = sizeFromShape(reduceShape);
const a2D = reshape$1({ inputs: { x: permutedX }, backend, attrs: { shape: [-1, inSize] } });
const reduced = reduce(a2D, a2D.dtype, 'min', backend);
let res;
if (keepDims) {
const newShape = expandShapeToKeepDim(outShape, origAxes);
res = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: newShape } });
}
else {
res = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: outShape } });
}
backend.disposeIntermediateTensorInfo(a2D);
backend.disposeIntermediateTensorInfo(reduced);
if (permutedAxes != null) {
backend.disposeIntermediateTensorInfo(permutedX);
}
return res;
}
const minConfig$1 = {
kernelName: Min,
backendName: 'webgl',
kernelFunc: min$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const MINIMUM = CHECK_NAN_SNIPPET + `
return min(a, b);
`;
const MINIMUM_PACKED = `
vec4 result = vec4(min(a, b));
bvec4 isNaNA = isnan(a);
bvec4 isNaNB = isnan(b);
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
` +
CHECK_NAN_SNIPPET_PACKED + `
return result;
`;
const minimum = binaryKernelFunc({
opSnippet: MINIMUM,
packedOpSnippet: MINIMUM_PACKED,
cpuKernelImpl: minimumImplCPU
});
const minimumConfig = {
kernelName: Minimum,
backendName: 'webgl',
kernelFunc: minimum
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class MirrorPadProgram {
constructor(xShape, paddings, mode) {
this.variableNames = ['x'];
this.outputShape = paddings.map((p, i) => p[0] /* beforePad */ + xShape[i] + p[1] /* afterPad */);
const rank = xShape.length;
const dtype = getCoordsDataType(rank);
const start = paddings.map(p => p[0]).join(',');
const end = paddings.map((p, i) => p[0] + xShape[i]).join(',');
const unpackedCoords = ['coords[0]', 'coords[1]', 'coords[2]', 'coords[3]'].slice(0, rank);
const offset = mode === 'reflect' ? 0 : 1;
if (rank === 1) {
this.userCode = `
int start = ${start};
int end = ${end};
void main() {
int outC = getOutputCoords();
if (outC < start) {
outC = start * 2 - outC - ${offset};
} else if(outC >= end) {
outC = (end - 1) * 2 - outC + ${offset};
}
setOutput(getX(outC - start));
}
`;
return;
}
this.userCode = `
${dtype} start = ${dtype}(${start});
${dtype} end = ${dtype}(${end});
void main() {
${dtype} outC = getOutputCoords();
for (int i = 0; i < ${rank}; i++) {
if (outC[i] < start[i]) {
outC[i] = start[i] * 2 - outC[i] - ${offset};
} else if(outC[i] >= end[i]) {
outC[i] = (end[i] - 1) * 2 - outC[i] + ${offset};
}
}
${dtype} coords = outC - start;
setOutput(getX(${unpackedCoords}));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Example shader code for
* `mirrorPad(tf.tensor1d([1, 2, 3], 'int32'), [[2, 2]], 'reflect')`
* ```
* const int start = int(2);
* const int end = int(5);
*
* void main() {
* int outputLoc = getOutputCoords();
* vec4 result = vec4(0.);
*
* int rc = outputLoc;
*
* int source = rc;
* if (source < start) {
* source = start * 2 - source - 0;
* } else if (source >= end) {
* source = (end - 1) * 2 - source + 0;
* }
* source -= start;
*
* result[0] = getChannel(getX(source), source);
* rc += 1;
* if(rc < 6) {
* int source = rc;
* if (source < start) {
* source = start * 2 - source - 0;
* } else if (source >= end) {
* source = (end - 1) * 2 - source + 0;
* }
* source -= start;
*
* result[1] = getChannel(getX(source), source);
* }
*
* setOutput(result);
* }
* ```
*/
class MirrorPadPackedProgram {
constructor(xShape, paddings, mode) {
this.variableNames = ['x'];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = paddings.map((p, i) => p[0] /* beforePad */ + xShape[i] + p[1] /* afterPad */);
const rank = xShape.length;
const dtype = getCoordsDataType(rank);
const start = paddings.map(p => p[0]).join(',');
const end = paddings.map((p, i) => p[0] + xShape[i]).join(',');
const coords = getChannels('rc', rank);
const source = getChannels('source', rank);
const cLimit = `${coords[rank - 1]} < ${this.outputShape[rank - 1]}`;
const innerDims = rank === 1 ? 'source' : `vec2(${source.slice(-2).join()})`;
const offset = mode === 'reflect' ? 0 : 1;
let mainLoop = '';
if (rank === 1) {
const padSetup = `
${dtype} source = rc;
if (source < start) {
source = start * 2 - source - ${offset};
} else if (source >= end) {
source = (end - 1) * 2 - source + ${offset};
}
source -= start;
`;
mainLoop = `
${dtype} rc = outputLoc;
${padSetup}
result[0] = getChannel(getX(${source.join()}), ${innerDims});
${coords[rank - 1]} += 1;
if(${cLimit}) {
${padSetup}
result[1] = getChannel(getX(${source.join()}), ${innerDims});
}
`;
}
else {
const padSetup = `
${dtype} source = rc;
${dtype} lt = ${dtype}(lessThan(source, start));
${dtype} gte = ${dtype}(greaterThanEqual(source, end));
${dtype} orig = 1 - (lt + gte);
source = orig * source +
lt * (start * 2 - source - ${offset}) +
gte * ((end - 1) * 2 - source + ${offset});
source -= start;
`;
mainLoop = `
${dtype} rc = outputLoc;
${padSetup}
result[0] = getChannel(getX(${source.join()}), ${innerDims});
${coords[rank - 1]} += 1;
if(${cLimit}) {
${padSetup}
result[1] = getChannel(getX(${source.join()}), ${innerDims});
}
rc = outputLoc;
${coords[rank - 2]} += 1;
if(${coords[rank - 2]} < ${this.outputShape[rank - 2]}) {
${padSetup}
result[2] = getChannel(getX(${source.join()}), ${innerDims});
${coords[rank - 1]} += 1;
if(${cLimit}) {
${padSetup}
result[3] = getChannel(getX(${source.join()}), ${innerDims});
}
}
`;
}
this.userCode = `
const ${dtype} start = ${dtype}(${start});
const ${dtype} end = ${dtype}(${end});
void main() {
${dtype} outputLoc = getOutputCoords();
vec4 result = vec4(0.);
${mainLoop}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const mirrorPadKernelFunc = ({ inputs, backend, attrs }) => {
const { x } = inputs;
const { paddings, mode } = attrs;
const program = env().getBool('WEBGL_PACK_ARRAY_OPERATIONS') ?
new MirrorPadPackedProgram(x.shape, paddings, mode) :
new MirrorPadProgram(x.shape, paddings, mode);
const output = backend.runWebGLProgram(program, [x], x.dtype);
return output;
};
const mirrorPadConfig$1 = {
kernelName: MirrorPad,
backendName: 'webgl',
kernelFunc: mirrorPadKernelFunc,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const MOD = `if (b == 0.0) return NAN;
return mod(a, b);`;
const MOD_PACKED = `
vec4 result = mod(a, b);
bvec4 isNaN = equal(b, vec4(0.0));
` +
CHECK_NAN_SNIPPET_PACKED + `
return result;
`;
const mod$1 = binaryKernelFunc({
opSnippet: MOD,
packedOpSnippet: MOD_PACKED,
});
const modConfig$1 = {
kernelName: Mod,
backendName: 'webgl',
kernelFunc: mod$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class MultinomialProgram {
constructor(batchSize, numOutcomes, numSamples) {
this.variableNames = ['probs'];
this.customUniforms = [{ name: 'seed', type: 'float' }];
this.outputShape = [batchSize, numSamples];
this.userCode = `
void main() {
ivec2 coords = getOutputCoords();
int batch = coords[0];
float r = random(seed);
float cdf = 0.0;
for (int i = 0; i < ${numOutcomes - 1}; i++) {
cdf += getProbs(batch, i);
if (r < cdf) {
setOutput(float(i));
return;
}
}
// If no other event happened, last event happened.
setOutput(float(${numOutcomes - 1}));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Without the equality check div produces 0.9999 for a = b, which when
// floored can cause errors.
const DIV = `
if (a == b) {
return 1.0;
};
return a / b;`;
// We do the same as in ./binaryop_gpu, with vec4 and ivec4.
// On Linux, the vectorized implementation produces NaNs when a and b are 0.
const DIV_PACKED = `
// vec4 one = vec4(equal(a, b));
// return one + (vec4(1.0) - one) * a / b;
vec4 result = a / b;
if(a.x == b.x) {
result.x = 1.;
}
if(a.y == b.y) {
result.y = 1.;
}
if(a.z == b.z) {
result.z = 1.;
}
if(a.w == b.w) {
result.w = 1.;
}
return result;
`;
const realDiv = binaryKernelFunc({ opSnippet: DIV, packedOpSnippet: DIV_PACKED, checkOutOfBounds: true });
const realDivConfig$1 = {
kernelName: RealDiv,
backendName: 'webgl',
kernelFunc: realDiv,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SUB = 'return a - b;';
const sub = binaryKernelFunc({
opSnippet: SUB,
packedOpSnippet: SUB,
supportsComplex: true,
cpuKernelImpl: subImplCPU
});
const subConfig = {
kernelName: Sub,
backendName: 'webgl',
kernelFunc: sub
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function softmax$1(args) {
const { inputs, backend, attrs } = args;
const { logits } = inputs;
const { dim } = attrs;
const axes = parseAxisParam([dim], logits.shape);
const maxLogit = max$1({
inputs: { x: logits },
backend,
attrs: { reductionIndices: axes, keepDims: false }
});
const expandedShape = expandShapeToKeepDim(maxLogit.shape, axes);
const maxLogitsReshaped = reshape$1({ inputs: { x: maxLogit }, backend, attrs: { shape: expandedShape } });
const a = sub({ inputs: { a: logits, b: maxLogitsReshaped }, backend });
const b = exp({ inputs: { x: a }, backend });
const sumExp = sum$1({ inputs: { x: b }, backend, attrs: { axis: axes, keepDims: false } });
const sumExpReshaped = reshape$1({ inputs: { x: sumExp }, backend, attrs: { shape: expandedShape } });
const res = realDiv({ inputs: { a: b, b: sumExpReshaped }, backend });
backend.disposeIntermediateTensorInfo(maxLogit);
backend.disposeIntermediateTensorInfo(maxLogitsReshaped);
backend.disposeIntermediateTensorInfo(a);
backend.disposeIntermediateTensorInfo(b);
backend.disposeIntermediateTensorInfo(sumExp);
backend.disposeIntermediateTensorInfo(sumExpReshaped);
return res;
}
const softmaxConfig$1 = {
kernelName: Softmax$1,
backendName: 'webgl',
kernelFunc: softmax$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function multinomial$1(args) {
const { inputs, backend, attrs } = args;
const { logits } = inputs;
const { numSamples, seed, normalized } = attrs;
const probs = normalized ?
logits :
softmax$1({ inputs: { logits }, backend, attrs: { dim: logits.shape.length - 1 } });
const batchSize = probs.shape[0];
const numOutcomes = probs.shape[1];
const program = new MultinomialProgram(batchSize, numOutcomes, numSamples);
const customValues = [[seed]];
const res = backend.runWebGLProgram(program, [probs], 'int32', customValues);
if (!normalized) {
backend.disposeIntermediateTensorInfo(probs);
}
return res;
}
const multinomialConfig$1 = {
kernelName: Multinomial,
backendName: 'webgl',
kernelFunc: multinomial$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const NEG = CHECK_NAN_SNIPPET$1 + `
return -x;
`;
const NEG_PACKED = `
vec4 result = -x;
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : result.r;
result.g = isNaN.g ? x.g : result.g;
result.b = isNaN.b ? x.b : result.b;
result.a = isNaN.a ? x.a : result.a;
return result;
`;
// This doesn't use unaryKernelFunc because negImplCPU is not of type
// SimpleUnaryKernelImplCPU.
function neg(args) {
const { inputs, backend } = args;
const { x } = inputs;
if (backend.shouldExecuteOnCPU([x])) {
const xData = backend.texData.get(x.dataId);
const [outValues, newShape] = negImplCPU(xData.values, x.shape, x.dtype);
return backend.makeTensorInfo(newShape, x.dtype, outValues);
}
let program;
if (env().getBool('WEBGL_PACK_UNARY_OPERATIONS')) {
program = new UnaryOpPackedProgram(x.shape, NEG_PACKED);
}
else {
program = new UnaryOpProgram(x.shape, NEG);
}
return backend.runWebGLProgram(program, [x], x.dtype);
}
const negConfig = {
kernelName: Neg,
backendName: 'webgl',
kernelFunc: neg
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const nonMaxSuppressionV3Impl$1 = nonMaxSuppressionV3Impl$2;
function nonMaxSuppressionV3$1(args) {
warn('tf.nonMaxSuppression() in webgl locks the UI thread. ' +
'Call tf.nonMaxSuppressionAsync() instead');
const { inputs, backend, attrs } = args;
const { boxes, scores } = inputs;
const { maxOutputSize, iouThreshold, scoreThreshold } = attrs;
const boxesVals = backend.readSync(boxes.dataId);
const scoresVals = backend.readSync(scores.dataId);
const { selectedIndices } = nonMaxSuppressionV3Impl$1(boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold);
return backend.makeTensorInfo([selectedIndices.length], 'int32', new Int32Array(selectedIndices));
}
const nonMaxSuppressionV3Config$1 = {
kernelName: NonMaxSuppressionV3,
backendName: 'webgl',
kernelFunc: nonMaxSuppressionV3$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const nonMaxSuppressionV4Impl$1 = nonMaxSuppressionV4Impl$2;
function nonMaxSuppressionV4$1(args) {
warn('tf.nonMaxSuppression() in webgl locks the UI thread. ' +
'Call tf.nonMaxSuppressionAsync() instead');
const { inputs, backend, attrs } = args;
const { boxes, scores } = inputs;
const { maxOutputSize, iouThreshold, scoreThreshold, padToMaxOutputSize } = attrs;
const boxesVals = backend.readSync(boxes.dataId);
const scoresVals = backend.readSync(scores.dataId);
const { selectedIndices, validOutputs } = nonMaxSuppressionV4Impl$1(boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold, padToMaxOutputSize);
return [
backend.makeTensorInfo([selectedIndices.length], 'int32', new Int32Array(selectedIndices)),
backend.makeTensorInfo([], 'int32', new Int32Array([validOutputs]))
];
}
const nonMaxSuppressionV4Config$1 = {
kernelName: NonMaxSuppressionV4,
backendName: 'webgl',
kernelFunc: nonMaxSuppressionV4$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const nonMaxSuppressionV5Impl$1 = nonMaxSuppressionV5Impl$2;
function nonMaxSuppressionV5$1(args) {
warn('tf.nonMaxSuppression() in webgl locks the UI thread. ' +
'Call tf.nonMaxSuppressionAsync() instead');
const { inputs, backend, attrs } = args;
const { boxes, scores } = inputs;
const { maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma } = attrs;
const boxesVals = backend.readSync(boxes.dataId);
const scoresVals = backend.readSync(scores.dataId);
const maxOutputSizeVal = maxOutputSize;
const iouThresholdVal = iouThreshold;
const scoreThresholdVal = scoreThreshold;
const softNmsSigmaVal = softNmsSigma;
const { selectedIndices, selectedScores } = nonMaxSuppressionV5Impl$1(boxesVals, scoresVals, maxOutputSizeVal, iouThresholdVal, scoreThresholdVal, softNmsSigmaVal);
return [
backend.makeTensorInfo([selectedIndices.length], 'int32', new Int32Array(selectedIndices)),
backend.makeTensorInfo([selectedScores.length], 'float32', new Float32Array(selectedScores))
];
}
const nonMaxSuppressionV5Config$1 = {
kernelName: NonMaxSuppressionV5,
backendName: 'webgl',
kernelFunc: nonMaxSuppressionV5$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class OneHotProgram {
constructor(numIndices, depth, onValue, offValue) {
this.variableNames = ['indices'];
this.outputShape = [numIndices, depth];
this.userCode = `
void main() {
ivec2 coords = getOutputCoords();
int index = round(getIndices(coords.x));
setOutput(mix(float(${offValue}), float(${onValue}),
float(index == coords.y)));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const oneHot$1 = (args) => {
const { inputs, backend, attrs } = args;
const { indices } = inputs;
const { dtype, depth, onValue, offValue } = attrs;
const indicesSize = sizeFromShape(indices.shape);
const program = new OneHotProgram(indicesSize, depth, onValue, offValue);
const reshaped = reshape$1({ inputs: { x: indices }, backend, attrs: { shape: [indicesSize] } });
const result = backend.runWebGLProgram(program, [reshaped], dtype);
backend.disposeIntermediateTensorInfo(reshaped);
const outShape = [...indices.shape, depth];
const out = reshape$1({ inputs: { x: result }, backend, attrs: { shape: outShape } });
backend.disposeIntermediateTensorInfo(result);
return out;
};
const oneHotConfig$1 = {
kernelName: OneHot,
backendName: 'webgl',
kernelFunc: oneHot$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function zerosLike$1(args) {
const { inputs, backend } = args;
const { x } = inputs;
if (x.dtype === 'complex64') {
const realPart = real({ inputs: { input: x }, backend });
const r = zerosLike$1({ inputs: { x: realPart }, backend });
const imagPart = imag$1({ inputs: { input: x }, backend });
const i = zerosLike$1({ inputs: { x: imagPart }, backend });
const result = complex({ inputs: { real: r, imag: i }, backend });
backend.disposeIntermediateTensorInfo(realPart);
backend.disposeIntermediateTensorInfo(r);
backend.disposeIntermediateTensorInfo(imagPart);
backend.disposeIntermediateTensorInfo(i);
return result;
}
else {
return fill$1({
attrs: {
shape: x.shape,
dtype: x.dtype,
value: x.dtype === 'string' ? '' : 0
},
backend
});
}
}
const zerosLikeConfig$1 = {
kernelName: ZerosLike,
backendName: 'webgl',
kernelFunc: zerosLike$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function onesLike$1(args) {
const { inputs, backend } = args;
const { x } = inputs;
if (x.dtype === 'string') {
throw new Error('onesLike is not supported under string dtype');
}
else if (x.dtype === 'complex64') {
const realPart = real({ inputs: { input: x }, backend });
const r = onesLike$1({ inputs: { x: realPart }, backend });
const imagPart = imag$1({ inputs: { input: x }, backend });
const i = zerosLike$1({ inputs: { x: imagPart }, backend });
const result = complex({ inputs: { real: r, imag: i }, backend });
backend.disposeIntermediateTensorInfo(realPart);
backend.disposeIntermediateTensorInfo(r);
backend.disposeIntermediateTensorInfo(imagPart);
backend.disposeIntermediateTensorInfo(i);
return result;
}
else {
// TODO(cais, smilkov): Add WebGL shader for onesLike:
// https://github.com/tensorflow/tfjs/issues/1293
return fill$1({ attrs: { shape: x.shape, dtype: x.dtype, value: 1 }, backend });
}
}
const onesLikeConfig$1 = {
kernelName: OnesLike,
backendName: 'webgl',
kernelFunc: onesLike$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function pack$1(args) {
const { inputs, backend, attrs } = args;
const { axis } = attrs;
if (inputs.length === 1) {
return expandDims$2({ inputs: { input: inputs[0] }, backend, attrs: { dim: axis } });
}
const shape = inputs[0].shape;
const dtype = inputs[0].dtype;
inputs.forEach(t => {
assertShapesMatch(shape, t.shape, 'All tensors passed to stack must have matching shapes');
assert$1(dtype === t.dtype, () => 'All tensors passed to stack must have matching dtypes');
});
const intermediateTensorInfos = [];
const expandedTensors = inputs.map(t => {
const expandedT = expandDims$2({ inputs: { input: t }, backend, attrs: { dim: axis } });
intermediateTensorInfos.push(expandedT);
return expandedT;
});
const result = concat$1({ inputs: expandedTensors, backend, attrs: { axis } });
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return result;
}
const packConfig$1 = {
kernelName: Pack,
backendName: 'webgl',
kernelFunc: pack$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class PadProgram {
constructor(xShape, paddings, constantValue) {
this.variableNames = ['x'];
this.customUniforms = [{ name: 'value', type: 'float' }];
this.outputShape = paddings.map((p, i) => p[0] /* beforePad */ + xShape[i] + p[1] /* afterPad */);
const rank = xShape.length;
const type = getCoordsDataType(rank);
const start = paddings.map(p => p[0]).join(',');
const end = paddings.map((p, i) => p[0] + xShape[i]).join(',');
const unpackedCoords = ['coords[0]', 'coords[1]', 'coords[2]', 'coords[3]'].slice(0, rank);
if (rank === 1) {
this.userCode = `
int start = ${start};
int end = ${end};
void main() {
int outC = getOutputCoords();
if (outC < start || outC >= end) {
setOutput(value);
} else {
setOutput(getX(outC - start));
}
}
`;
return;
}
this.userCode = `
${type} start = ${type}(${start});
${type} end = ${type}(${end});
void main() {
${type} outC = getOutputCoords();
if (any(lessThan(outC, start)) || any(greaterThanEqual(outC, end))) {
setOutput(value);
} else {
${type} coords = outC - start;
setOutput(getX(${unpackedCoords}));
}
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class PadPackedProgram {
constructor(xShape, paddings, constantValue) {
this.variableNames = ['x'];
this.packedInputs = true;
this.packedOutput = true;
this.customUniforms = [{ name: 'value', type: 'float' }];
this.outputShape = paddings.map((p, i) => p[0] /* beforePad */ + xShape[i] + p[1] /* afterPad */);
const rank = xShape.length;
const dtype = getCoordsDataType(rank);
const start = paddings.map(p => p[0]).join(',');
const end = paddings.map((p, i) => p[0] + xShape[i]).join(',');
const coords = getChannels('rc', rank);
const source = getChannels('source', rank);
const cLimit = `${coords[rank - 1]} < ${this.outputShape[rank - 1]}`;
const innerDims = rank === 1 ? 'source' : `vec2(${source.slice(-2).join()})`;
const componentSetup = [
`${dtype} rc = outputLoc;`, `${coords[rank - 1]} += 1;
if(${cLimit}) {
`,
rank === 1 ? '' : `}
rc = outputLoc;
${coords[rank - 2]} += 1;
if(${coords[rank - 2]} < ${this.outputShape[rank - 2]}) {`,
rank === 1 ? '' : ` ${coords[rank - 1]} += 1;
if(${cLimit}) {`
];
const paddingArea = rank === 1 ?
'rc < start || rc >= end' :
'any(lessThan(rc, start)) || any(greaterThanEqual(rc, end))';
let mainLoop = '';
for (let i = 0, j = rank === 1 ? 2 : 4; i < j; i++) {
mainLoop += `
${componentSetup[i]}
if (${paddingArea}) {
result[${i}] = float(value);
} else {
${dtype} source = rc - start;
result[${i}] = getChannel(getX(${source.join()}), ${innerDims});
}
`;
}
mainLoop += (rank === 1 ? `} ` : `}}`);
this.userCode = `
const ${dtype} start = ${dtype}(${start});
const ${dtype} end = ${dtype}(${end});
void main() {
${dtype} outputLoc = getOutputCoords();
vec4 result = vec4(0.);
${mainLoop}
setOutput(result);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const padV2$1 = (args) => {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { paddings, constantValue } = attrs;
if (sizeFromShape(x.shape) === 0) {
// Short-circuit the computation, since x doesn't have value, only
// the shape is used to compute output shape to pad.
const outputShape = paddings.map((p, i) => p[0] /* beforePad */ + x.shape[i] + p[1] /* afterPad */);
return fill$1({
backend,
attrs: { shape: outputShape, value: constantValue, dtype: x.dtype }
});
}
const program = env().getBool('WEBGL_PACK_ARRAY_OPERATIONS') ?
new PadPackedProgram(x.shape, paddings, constantValue) :
new PadProgram(x.shape, paddings, constantValue);
const customValues = [[constantValue]];
return backend.runWebGLProgram(program, [x], x.dtype, customValues);
};
const padV2Config$1 = {
kernelName: PadV2,
backendName: 'webgl',
kernelFunc: padV2$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const POW = `
if(a < 0.0 && floor(b) < b){
return NAN;
}
if (b == 0.0) {
return 1.0;
}
return (round(mod(b, 2.0)) != 1) ?
pow(abs(a), b) : sign(a) * pow(abs(a), b);
`;
const POW_PACKED = `
// isModRound1 has 1 for components with round(mod(b, 2.0)) == 1, 0 otherwise.
vec4 isModRound1 = vec4(equal(round(mod(b, 2.0)), ivec4(1)));
vec4 multiplier = sign(a) * isModRound1 + (vec4(1.0) - isModRound1);
vec4 result = multiplier * pow(abs(a), b);
// Ensure that a^0 = 1, including 0^0 = 1 as this correspond to TF and JS
bvec4 isExpZero = equal(b, vec4(0.0));
result.r = isExpZero.r ? 1.0 : result.r;
result.g = isExpZero.g ? 1.0 : result.g;
result.b = isExpZero.b ? 1.0 : result.b;
result.a = isExpZero.a ? 1.0 : result.a;
bvec4 isNaN1 = lessThan(a, vec4(0.0));
bvec4 isNaN2 = lessThan(floor(b), b);
bvec4 isNaN = bvec4(isNaN1.x && isNaN2.x, isNaN1.y && isNaN2.y, isNaN1.z && isNaN2.z, isNaN1.w && isNaN2.w);
` +
CHECK_NAN_SNIPPET_PACKED + `
return result;
`;
const pow$1 = binaryKernelFunc({ opSnippet: POW, packedOpSnippet: POW_PACKED });
const powConfig$1 = {
kernelName: Pow,
backendName: 'webgl',
kernelFunc: pow$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function prod(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
const xRank = x.shape.length;
const toDispose = [];
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
let permutedX = x;
if (permutedAxes != null) {
permutedX = transpose({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
axes = getInnerMostAxes(axes.length, xRank);
toDispose.push(permutedX);
}
assertAxesAreInnerMostDims('prod', axes, xRank);
let res;
if (backend.shouldExecuteOnCPU([permutedX])) {
const xVals = backend.texData.get(permutedX.dataId).values;
const { outVals, outShape, outDtype } = prodImplCPU(permutedX.shape, permutedX.dtype, xVals, axes);
res = backend.makeTensorInfo(outShape, outDtype, outVals);
}
else {
const [outShape, reduceShape] = computeOutAndReduceShapes(permutedX.shape, axes);
const inSize = sizeFromShape(reduceShape);
const a2D = reshape$1({ inputs: { x: permutedX }, backend, attrs: { shape: [-1, inSize] } });
const outputDType = sumOutType(x.dtype);
const reduced = reduce(a2D, outputDType, 'prod', backend);
res = reshape$1({ inputs: { x: reduced }, backend, attrs: { shape: outShape } });
toDispose.push(a2D);
toDispose.push(reduced);
}
if (keepDims) {
toDispose.push(res);
const newShape = expandShapeToKeepDim(res.shape, origAxes);
res = reshape$1({ inputs: { x: res }, backend, attrs: { shape: newShape } });
}
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return res;
}
const prodConfig = {
kernelName: Prod,
backendName: 'webgl',
kernelFunc: prod
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function raggedGather$1(args) {
const { inputs, backend, attrs } = args;
const { paramsNestedSplits, paramsDenseValues, indices } = inputs;
const { outputRaggedRank } = attrs;
const $paramsNestedSplits = paramsNestedSplits.map(t => backend.readSync(t.dataId));
const $paramsNestedSplitsShapes = paramsNestedSplits.map(t => t.shape);
const $paramsDenseValues = backend.readSync(paramsDenseValues.dataId);
const $indices = backend.readSync(indices.dataId);
const [outputNestedSplits, outputDenseValues, outputDenseValuesShape] = raggedGatherImplCPU($paramsNestedSplits, $paramsNestedSplitsShapes, $paramsDenseValues, paramsDenseValues.shape, paramsDenseValues.dtype, $indices, indices.shape, outputRaggedRank);
const outputNestedSplitsTensors = outputNestedSplits.map((splits) => backend.makeTensorInfo([splits.length], 'int32', splits));
const outputDenseValuesTensor = backend.makeTensorInfo(outputDenseValuesShape, paramsDenseValues.dtype, outputDenseValues);
return outputNestedSplitsTensors.concat([outputDenseValuesTensor]);
}
const raggedGatherConfig$1 = {
kernelName: RaggedGather,
backendName: 'webgl',
kernelFunc: raggedGather$1,
};
/**
* @license
* Copyright 2022 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function raggedRange$1(args) {
const { inputs, backend } = args;
const { starts, limits, deltas } = inputs;
const $starts = backend.readSync(starts.dataId);
const $limits = backend.readSync(limits.dataId);
const $deltas = backend.readSync(deltas.dataId);
const [rtNestedSplitsData, rtDenseValuesData] = raggedRangeImplCPU($starts, starts.shape, starts.dtype, $limits, limits.shape, $deltas, deltas.shape);
const rtNestedSplits = backend.makeTensorInfo([rtNestedSplitsData.length], 'int32', rtNestedSplitsData);
const rtDenseValues = backend.makeTensorInfo([rtDenseValuesData.length], starts.dtype, rtDenseValuesData);
return [rtNestedSplits, rtDenseValues];
}
const raggedRangeConfig$1 = {
kernelName: RaggedRange,
backendName: 'webgl',
kernelFunc: raggedRange$1,
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function raggedTensorToTensor$1(args) {
const { inputs, backend, attrs } = args;
const { shape, values, defaultValue, rowPartitionTensors } = inputs;
const { rowPartitionTypes } = attrs;
const $shape = backend.readSync(shape.dataId);
const $values = backend.readSync(values.dataId);
const $defaultValue = backend.readSync(defaultValue.dataId);
const $rowPartitionValues = rowPartitionTensors.map(t => backend.readSync(t.dataId));
const rowPartitionValuesShapes = rowPartitionTensors.map(t => t.shape);
const [outputShape, output] = raggedTensorToTensorImplCPU($shape, shape.shape, $values, values.shape, values.dtype, $defaultValue, defaultValue.shape, $rowPartitionValues, rowPartitionValuesShapes, rowPartitionTypes);
return backend.makeTensorInfo(outputShape, values.dtype, output);
}
const raggedTensorToTensorConfig$1 = {
kernelName: RaggedTensorToTensor,
backendName: 'webgl',
kernelFunc: raggedTensorToTensor$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const range$2 = (args) => {
const { backend, attrs } = args;
const { start, stop, step, dtype } = attrs;
const values = rangeImplCPU(start, stop, step, dtype);
return backend.makeTensorInfo([values.length], dtype, values);
};
const rangeConfig$1 = {
kernelName: Range,
backendName: 'webgl',
kernelFunc: range$2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const RECIPROCAL = `return 1.0 / x;`;
const reciprocal$1 = unaryKernelFunc({ opSnippet: RECIPROCAL });
const reciprocalConfig$1 = {
kernelName: Reciprocal,
backendName: 'webgl',
kernelFunc: reciprocal$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const RELU = CHECK_NAN_SNIPPET$1 + `
return (x < 0.0) ? 0.0 : x;
`;
const RELU_PACKED = `
vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : result.r;
result.g = isNaN.g ? x.g : result.g;
result.b = isNaN.b ? x.b : result.b;
result.a = isNaN.a ? x.a : result.a;
return result;
`;
const relu$1 = unaryKernelFunc({ opSnippet: RELU, packedOpSnippet: RELU_PACKED });
const reluConfig$1 = {
kernelName: Relu$1,
backendName: 'webgl',
kernelFunc: relu$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const RELU6 = CHECK_NAN_SNIPPET$1 + `
return (x < 0.0) ? 0.0 : min(6.0, x);
`;
const RELU6_PACKED = `
vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : result.r;
result.g = isNaN.g ? x.g : result.g;
result.b = isNaN.b ? x.b : result.b;
result.a = isNaN.a ? x.a : result.a;
return result;
`;
const relu6$1 = unaryKernelFunc({ opSnippet: RELU6, packedOpSnippet: RELU6_PACKED });
const relu6Config$1 = {
kernelName: Relu6$1,
backendName: 'webgl',
kernelFunc: relu6$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ResizeBilinearProgram {
constructor(inputShape, newHeight, newWidth, alignCorners, halfPixelCenters) {
this.variableNames = ['A'];
this.outputShape = [];
const [batch, oldHeight, oldWidth, depth] = inputShape;
this.outputShape = [batch, newHeight, newWidth, depth];
const effectiveInSize = [
(alignCorners && newHeight > 1) ? oldHeight - 1 : oldHeight,
(alignCorners && newWidth > 1) ? oldWidth - 1 : oldWidth
];
const effectiveOutSize = [
(alignCorners && newHeight > 1) ? newHeight - 1 : newHeight,
(alignCorners && newWidth > 1) ? newWidth - 1 : newWidth
];
let sourceFracIndexRC;
if (halfPixelCenters) {
sourceFracIndexRC =
`(vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC` +
` - vec2(0.5)`;
}
else {
sourceFracIndexRC = `vec2(yRC) * effectiveInputOverOutputRatioRC`;
}
this.userCode = `
const vec2 effectiveInputOverOutputRatioRC = vec2(
${effectiveInSize[0] / effectiveOutSize[0]},
${effectiveInSize[1] / effectiveOutSize[1]});
const vec2 inputShapeRC = vec2(${oldHeight}.0, ${oldWidth}.0);
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
ivec2 yRC = coords.yz;
// Fractional source index.
vec2 sourceFracIndexRC = ${sourceFracIndexRC};
// Compute the four integer indices.
ivec2 sourceFloorRC = ivec2(max(sourceFracIndexRC, vec2(0.0)));
ivec2 sourceCeilRC = ivec2(
min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));
float topLeft = getA(b, sourceFloorRC.x, sourceFloorRC.y, d);
float bottomLeft = getA(b, sourceCeilRC.x, sourceFloorRC.y, d);
float topRight = getA(b, sourceFloorRC.x, sourceCeilRC.y, d);
float bottomRight = getA(b, sourceCeilRC.x, sourceCeilRC.y, d);
vec2 fracRC = sourceFracIndexRC - vec2(sourceFloorRC);
float top = topLeft + (topRight - topLeft) * fracRC.y;
float bottom = bottomLeft + (bottomRight - bottomLeft) * fracRC.y;
float newValue = top + (bottom - top) * fracRC.x;
setOutput(newValue);
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ResizeBilinearPackedProgram {
constructor(inputShape, newHeight, newWidth, alignCorners, halfPixelCenters) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = [];
const [batch, oldHeight, oldWidth, depth] = inputShape;
this.outputShape = [batch, newHeight, newWidth, depth];
const effectiveInSize = [
(alignCorners && newHeight > 1) ? oldHeight - 1 : oldHeight,
(alignCorners && newWidth > 1) ? oldWidth - 1 : oldWidth
];
const effectiveOutSize = [
(alignCorners && newHeight > 1) ? newHeight - 1 : newHeight,
(alignCorners && newWidth > 1) ? newWidth - 1 : newWidth
];
let sourceFracIndexRC;
if (halfPixelCenters) {
sourceFracIndexRC = `(vec3(yRC) + vec3(0.5)) * ` +
`effectiveInputOverOutputRatioRC - vec3(0.5)`;
}
else {
sourceFracIndexRC = `vec3(yRC) * effectiveInputOverOutputRatioRC`;
}
this.userCode = `
const vec3 effectiveInputOverOutputRatioRC = vec3(
${effectiveInSize[0] / effectiveOutSize[0]},
${effectiveInSize[1] / effectiveOutSize[1]},
${effectiveInSize[1] / effectiveOutSize[1]});
const vec3 inputShapeRC = vec3(${oldHeight}.0, ${oldWidth}.0,
${oldWidth}.0);
float getAValue(int b, int r, int c, int d) {
return getChannel(getA(b, r, c, d), vec2(c, d));
}
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
// Calculate values for next column in yRC.z.
ivec3 yRC = coords.yzz + ivec3(0, 0, 1);
// Fractional source index.
vec3 sourceFracIndexRC = ${sourceFracIndexRC};
// Compute the four integer indices.
ivec3 sourceFloorRC = ivec3(max(sourceFracIndexRC, vec3(0.0)));
ivec3 sourceCeilRC = ivec3(
min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));
// Should we calculate next column and row elements in 2x2 packed cell.
bool hasNextCol = d < ${depth - 1};
bool hasNextRow = coords.z < ${newWidth - 1};
// In parallel, construct four corners for all four components in
// packed 2x2 cell.
vec4 topLeft = vec4(
getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d),
hasNextCol ? getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d + 1)
: 0.0,
hasNextRow ? getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d)
: 0.0,
(hasNextRow && hasNextCol) ?
getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d + 1) : 0.0);
vec4 bottomLeft = vec4(
getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d),
hasNextCol ? getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d + 1)
: 0.0,
hasNextRow ? getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d)
: 0.0,
(hasNextRow && hasNextCol) ?
getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d + 1) : 0.0);
vec4 topRight = vec4(
getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d),
hasNextCol ? getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d + 1)
: 0.0,
hasNextRow ? getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d)
: 0.0,
(hasNextRow && hasNextCol) ?
getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d + 1) : 0.0);
vec4 bottomRight = vec4(
getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d),
hasNextCol ? getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d + 1)
: 0.0,
hasNextRow ? getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d)
: 0.0,
(hasNextRow && hasNextCol) ?
getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d + 1) : 0.0);
vec3 fracRC = sourceFracIndexRC - vec3(sourceFloorRC);
vec4 top = mix(topLeft, topRight, fracRC.yyzz);
vec4 bottom = mix(bottomLeft, bottomRight, fracRC.yyzz);
vec4 newValue = mix(top, bottom, fracRC.x);
setOutput(newValue);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeBilinear$1(args) {
const { inputs, backend, attrs } = args;
const { images } = inputs;
const { alignCorners, halfPixelCenters, size } = attrs;
const [newHeight, newWidth] = size;
const program = env().getBool('WEBGL_PACK_IMAGE_OPERATIONS') ?
new ResizeBilinearPackedProgram(images.shape, newHeight, newWidth, alignCorners, halfPixelCenters) :
new ResizeBilinearProgram(images.shape, newHeight, newWidth, alignCorners, halfPixelCenters);
return backend.runWebGLProgram(program, [images], 'float32');
}
const resizeBilinearConfig$1 = {
kernelName: ResizeBilinear,
backendName: 'webgl',
kernelFunc: resizeBilinear$1
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ResizeBilinearBackpropProgram {
constructor(dyShape, inputShape, alignCorners) {
this.variableNames = ['dy'];
this.outputShape = [];
this.outputShape = inputShape;
const [, xHeight, xWidth,] = inputShape;
const [, yHeight, yWidth] = dyShape;
// In the backwards pass, we want to find the pixels that were generated for
// each pixel in the input image the forward pass and add the corresponding
// coefficient from dy to the gradient (with some interpolation).
const effectiveXSize = [
(alignCorners && yHeight > 1) ? xHeight - 1 : xHeight,
(alignCorners && yWidth > 1) ? xWidth - 1 : xWidth
];
const effectiveYSize = [
(alignCorners && yHeight > 1) ? yHeight - 1 : yHeight,
(alignCorners && yWidth > 1) ? yWidth - 1 : yWidth
];
const heightScale = effectiveXSize[0] / effectiveYSize[0];
const widthScale = effectiveXSize[1] / effectiveYSize[1];
const invHeightScale = 1 / heightScale;
const invWidthScale = 1 / widthScale;
// This defines the size of the window of values around a particular
// index in dy that we want to search for contributions to dx.
const winHeight = (Math.ceil(invHeightScale) * 2) + 2;
const winWidth = (Math.ceil(invWidthScale) * 2) + 2;
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
int r = coords[1];
int c = coords[2];
float accumulator = 0.0;
const float heightScale = float(${heightScale});
const float widthScale = float(${widthScale});
const float invHeightScale = float(${invHeightScale});
const float invWidthScale = float(${invWidthScale});
const int winHeight = int(${winHeight});
const int winWidth = int(${winWidth});
// Compute bounds for where in dy we will look
float startRLerp = floor(float(r) * invHeightScale);
int startDyR = int(startRLerp - float(winHeight / 2));
float startCLerp = floor(float(c) * invWidthScale);
int startDyC = int(startCLerp - float(winWidth / 2));
// Loop over dy
for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
int dyR = dyROffset + startDyR;
// Guard against the window exceeding the bounds of dy
if (dyR < 0 || dyR >= ${yHeight}) {
continue;
}
for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
int dyC = dyCOffset + startDyC;
// Guard against the window exceeding the bounds of dy
if (dyC < 0 || dyC >= ${yWidth}) {
continue;
}
float dxR = float(dyR) * heightScale;
int topDxRIndex = int(floor(dxR));
int bottomDxRIndex = int(min(ceil(dxR), ${xHeight - 1}.0));
float dxRLerp = dxR - float(topDxRIndex);
float inverseDxRLerp = 1.0 - dxRLerp;
float dxC = float(dyC) * widthScale;
int leftDxCIndex = int(floor(dxC));
int rightDxCIndex = int(min(ceil(dxC), ${xWidth - 1}.0));
float dxCLerp = dxC - float(leftDxCIndex);
float inverseDxCLerp = 1.0 - dxCLerp;
if (r == topDxRIndex && c == leftDxCIndex) {
// topLeft
accumulator +=
getDy(b, dyR, dyC, d) * inverseDxRLerp * inverseDxCLerp;
}
if (r == topDxRIndex && c == rightDxCIndex) {
// topRight
accumulator += getDy(b, dyR, dyC, d) * inverseDxRLerp * dxCLerp;
}
if (r == bottomDxRIndex && c == leftDxCIndex) {
// bottomLeft
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * inverseDxCLerp;
}
if (r == bottomDxRIndex && c == rightDxCIndex) {
// bottomRight
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * dxCLerp;
}
}
}
// End loop over dy
setOutput(accumulator);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeBilinearGrad$1(args) {
const { inputs, backend, attrs } = args;
const { images, dy } = inputs;
const { alignCorners } = attrs;
const program = new ResizeBilinearBackpropProgram(dy.shape, images.shape, alignCorners);
return backend.runWebGLProgram(program, [dy], dy.dtype);
}
const resizeBilinearGradConfig$2 = {
kernelName: ResizeBilinearGrad,
backendName: 'webgl',
kernelFunc: resizeBilinearGrad$1
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ResizeNearestNeighborProgram {
constructor(inputShape, newHeight, newWidth, alignCorners, halfPixelCenters) {
this.variableNames = ['A'];
this.outputShape = [];
const [batch, oldHeight, oldWidth, depth] = inputShape;
this.outputShape = [batch, newHeight, newWidth, depth];
const effectiveInSize = [
(alignCorners && newHeight > 1) ? oldHeight - 1 : oldHeight,
(alignCorners && newWidth > 1) ? oldWidth - 1 : oldWidth
];
const effectiveOutSize = [
(alignCorners && newHeight > 1) ? newHeight - 1 : newHeight,
(alignCorners && newWidth > 1) ? newWidth - 1 : newWidth
];
// When align corners is false, we rounds the value with floor.
const roundBase = alignCorners ? '0.5' : '0.0';
let sourceFracIndexRC;
if (halfPixelCenters) {
sourceFracIndexRC =
`max((vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC` +
`, vec2(0.0))`;
}
else {
sourceFracIndexRC = `vec2(yRC) * effectiveInputOverOutputRatioRC`;
}
this.userCode = `
const vec2 effectiveInputOverOutputRatioRC = vec2(
${effectiveInSize[0] / effectiveOutSize[0]},
${effectiveInSize[1] / effectiveOutSize[1]});
const vec2 inputShapeRC = vec2(${oldHeight}.0, ${oldWidth}.0);
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
ivec2 yRC = coords.yz;
// Fractional source index.
vec2 sourceFracIndexRC = ${sourceFracIndexRC};
// Compute the coordinators of nearest neighbor point.
ivec2 sourceNearestRC = ivec2(
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${roundBase})));
float newValue = getA(b, sourceNearestRC.x, sourceNearestRC.y, d);
setOutput(newValue);
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ResizeNearestNeighborPackedProgram {
constructor(inputShape, newHeight, newWidth, alignCorners, halfPixelCenters) {
this.variableNames = ['A'];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = [];
const [batch, oldHeight, oldWidth, depth] = inputShape;
this.outputShape = [batch, newHeight, newWidth, depth];
const effectiveInSize = [
(alignCorners && newHeight > 1) ? oldHeight - 1 : oldHeight,
(alignCorners && newWidth > 1) ? oldWidth - 1 : oldWidth
];
const effectiveOutSize = [
(alignCorners && newHeight > 1) ? newHeight - 1 : newHeight,
(alignCorners && newWidth > 1) ? newWidth - 1 : newWidth
];
// When align corners is false, we rounds the value with floor.
const roundBase = alignCorners ? '0.5' : '0.0';
let sourceFracIndexRC;
if (halfPixelCenters) {
sourceFracIndexRC = `max((vec3(yRC) + vec3(0.5)) * ` +
`effectiveInputOverOutputRatioRC, vec3(0.0))`;
}
else {
sourceFracIndexRC = `vec3(yRC) * effectiveInputOverOutputRatioRC`;
}
this.userCode = `
const vec3 effectiveInputOverOutputRatioRC = vec3(
${effectiveInSize[0] / effectiveOutSize[0]},
${effectiveInSize[1] / effectiveOutSize[1]},
${effectiveInSize[1] / effectiveOutSize[1]});
const vec3 inputShapeRC = vec3(${oldHeight}.0, ${oldWidth}.0,
${oldWidth}.0);
float getAValue(int b, int r, int c, int d) {
return getChannel(getA(b, r, c, d), vec2(c, d));
}
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
// Calculate values for next column in yRC.z.
ivec3 yRC = coords.yzz + ivec3(0, 0, 1);
// Fractional source index.
vec3 sourceFracIndexRC = ${sourceFracIndexRC};
// Compute the coordinators of nearest neighbor point.
ivec3 sourceNearestRC = ivec3(
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${roundBase})));
// Should we calculate next column and row elements in 2x2 packed cell.
bool hasNextCol = d < ${depth - 1};
bool hasNextRow = coords.z < ${newWidth - 1};
vec4 newValue = vec4(
getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d),
hasNextCol ? getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d + 1)
: 0.0,
hasNextRow ? getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d)
: 0.0,
(hasNextRow && hasNextCol) ?
getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d + 1) : 0.0);
setOutput(newValue);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeNearestNeighbor$1(args) {
const { inputs, backend, attrs } = args;
const { images } = inputs;
const { alignCorners, halfPixelCenters, size } = attrs;
const [newHeight, newWidth] = size;
const program = env().getBool('WEBGL_PACK_IMAGE_OPERATIONS') ?
new ResizeNearestNeighborPackedProgram(images.shape, newHeight, newWidth, alignCorners, halfPixelCenters) :
new ResizeNearestNeighborProgram(images.shape, newHeight, newWidth, alignCorners, halfPixelCenters);
return backend.runWebGLProgram(program, [images], images.dtype);
}
const resizeNearestNeighborConfig$1 = {
kernelName: ResizeNearestNeighbor,
backendName: 'webgl',
kernelFunc: resizeNearestNeighbor$1
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ResizeNearestNeigborBackpropProgram {
constructor(dyShape, inputShape, alignCorners) {
this.variableNames = ['dy'];
this.outputShape = [];
this.outputShape = inputShape;
const [, xHeight, xWidth,] = inputShape;
const [, yHeight, yWidth] = dyShape;
// In the backwards pass, we want to find the pixels that were generated for
// each pixel in the input image the forward pass and add the corresponding
// coefficient from dy to the gradient (with some interpolation).
const effectiveXSize = [
(alignCorners && yHeight > 1) ? xHeight - 1 : xHeight,
(alignCorners && yWidth > 1) ? xWidth - 1 : xWidth
];
const effectiveYSize = [
(alignCorners && yHeight > 1) ? yHeight - 1 : yHeight,
(alignCorners && yWidth > 1) ? yWidth - 1 : yWidth
];
const heightScale = effectiveXSize[0] / effectiveYSize[0];
const widthScale = effectiveXSize[1] / effectiveYSize[1];
const invHeightScale = 1 / heightScale;
const invWidthScale = 1 / widthScale;
// This defines the size of the window of values around a particular
// index in dy that we want to search for contributions to dx.
const winHeight = (Math.ceil(invHeightScale) * 2) + 2;
const winWidth = (Math.ceil(invWidthScale) * 2) + 2;
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int b = coords[0];
int d = coords[3];
int r = coords[1];
int c = coords[2];
float accumulator = 0.0;
const float heightScale = float(${heightScale});
const float widthScale = float(${widthScale});
const float invHeightScale = float(${invHeightScale});
const float invWidthScale = float(${invWidthScale});
const int winHeight = int(${winHeight});
const int winWidth = int(${winWidth});
// Compute bounds for where in dy we will look
float startRLerp = floor(float(r) * invHeightScale);
int startDyR = int(floor(startRLerp - float(winHeight / 2)));
float startCLerp = floor(float(c) * invWidthScale);
int startDyC = int(floor(startCLerp - float(winWidth / 2)));
// Loop over dy
for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
int dyR = dyROffset + startDyR;
// Guard against the window exceeding the bounds of dy
if (dyR < 0 || dyR >= ${yHeight}) {
continue;
}
for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
int dyC = dyCOffset + startDyC;
// Guard against the window exceeding the bounds of dy
if (dyC < 0 || dyC >= ${yWidth}) {
continue;
}
float sourceFracRow =
float(${effectiveXSize[0]}) *
(float(dyR) / float(${effectiveYSize[0]}));
float sourceFracCol =
float(${effectiveXSize[1]}) *
(float(dyC) / float(${effectiveYSize[1]}));
int sourceNearestRow = int(min(
float(int(${xHeight}) - 1),
${alignCorners} ? float(round(sourceFracRow)) :
float(floor(sourceFracRow))));
int sourceNearestCol = int(min(
float(int(${xWidth}) - 1),
${alignCorners} ? float(round(sourceFracCol)) :
float(floor(sourceFracCol))));
if (r == sourceNearestRow && c == sourceNearestCol) {
accumulator += getDy(b, dyR, dyC, d);
}
}
}
// End loop over dy
setOutput(accumulator);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeNearestNeighborGrad$1(args) {
const { inputs, backend, attrs } = args;
const { images, dy } = inputs;
const { alignCorners } = attrs;
const program = new ResizeNearestNeigborBackpropProgram(dy.shape, images.shape, alignCorners);
return backend.runWebGLProgram(program, [dy], dy.dtype);
}
const resizeNearestNeighborGradConfig$2 = {
kernelName: ResizeNearestNeighborGrad,
backendName: 'webgl',
kernelFunc: resizeNearestNeighborGrad$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ReverseProgram {
constructor(xShape, axis) {
this.variableNames = ['x'];
const rank = xShape.length;
if (rank > 4) {
throw new Error(`WebGL backend: Reverse of rank-${rank} tensor is not yet supported`);
}
this.outputShape = xShape;
if (rank === 1) {
this.userCode = `
void main() {
int coord = getOutputCoords();
setOutput(getX(${xShape[0]} - coord - 1));
}
`;
return;
}
const getInCoord = (i) => {
if (axis.indexOf(i) !== -1 && xShape[i] !== 1) {
return `${xShape[i]} - coords[${i}] - 1`;
}
return `coords[${i}]`;
};
const inCoords = xShape.map((_, i) => getInCoord(i)).join(',');
const type = getCoordsDataType(rank);
this.userCode = `
void main() {
${type} coords = getOutputCoords();
setOutput(getX(${inCoords}));
}
`;
}
}
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ReversePackedProgram {
constructor(xShape, axis) {
this.variableNames = ['x'];
this.packedInputs = true;
this.packedOutput = true;
const rank = xShape.length;
if (rank > 4) {
throw new Error(`WebGL backend: Reverse of rank-${rank} tensor is not yet supported`);
}
this.outputShape = xShape;
const channels = getChannels('rc', rank);
const nextColumn = `${channels[rank - 1]} + 1 < ${this.outputShape[rank - 1]}`;
const nextRow = `${channels[rank - 2]} + 1 < ${this.outputShape[rank - 2]}`;
const type = getCoordsDataType(rank);
if (rank === 1) {
this.userCode = `
void main(){
int rc = getOutputCoords();
vec4 result = vec4(0.);
result.r = getChannel(getX(${xShape[0]} - rc - 1),
${xShape[0]} - rc - 1);
if(${nextColumn}){
result.g = getChannel(getX(${xShape[0]} - (rc + 1) - 1),
${xShape[0]} - (rc + 1) - 1);
}
setOutput(result);
}
`;
}
else {
this.userCode = `
void main() {
${type} rc = getOutputCoords();
vec4 result = vec4(0.);
result.r = ${getR(channels.slice())};
if(${nextColumn}){
result.g = ${getG(channels.slice())};
}
if(${nextRow}) {
result.b = ${getB(channels.slice())};
if(${nextColumn}) {
result.a = ${getA(channels.slice())};
}
}
setOutput(result);
}
`;
}
function getR(channels) {
return getChannel(channels);
}
function getG(channels) {
channels[rank - 1] = '(' + channels[rank - 1] + ` + 1)`;
return getChannel(channels);
}
function getB(channels) {
channels[rank - 2] = '(' + channels[rank - 2] + ` + 1)`;
return getChannel(channels);
}
function getA(channels) {
channels[rank - 1] = '(' + channels[rank - 1] + ` + 1)`;
channels[rank - 2] = '(' + channels[rank - 2] + ` + 1)`;
return getChannel(channels);
}
function getChannel(channels) {
const inCoordsArray = xShape.map((_, i) => getInCoord(i, channels));
const inCoords = inCoordsArray.join(',');
const innerDims = inCoordsArray.slice(-2).join(',');
return `getChannel(getX(${inCoords}), vec2(${innerDims}))`;
}
function getInCoord(i, channels1) {
if (axis.indexOf(i) !== -1 && xShape[i] !== 1) {
return `${xShape[i]} - ${channels1[i]} - 1`;
}
else {
return `${channels1[i]}`;
}
}
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function reverse$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { dims } = attrs;
const xRank = x.shape.length;
const $dims = parseAxisParam(dims, x.shape);
if (xRank === 0) {
return identity({ inputs: { x }, backend });
}
const program = env().getBool('WEBGL_PACK_ARRAY_OPERATIONS') ?
new ReversePackedProgram(x.shape, $dims) :
new ReverseProgram(x.shape, $dims);
return backend.runWebGLProgram(program, [x], x.dtype);
}
const reverseConfig$1 = {
kernelName: Reverse,
backendName: 'webgl',
kernelFunc: reverse$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class RotateProgram {
constructor(imageShape, fillValue) {
this.variableNames = ['Image'];
this.outputShape = [];
this.customUniforms = [{ name: 'params', type: 'vec4' }];
const imageHeight = imageShape[1];
const imageWidth = imageShape[2];
this.outputShape = imageShape;
let fillSnippet = '';
if (typeof fillValue === 'number') {
fillSnippet = `float outputValue = ${fillValue.toFixed(2)};`;
}
else {
fillSnippet = `
vec3 fill = vec3(${fillValue.join(',')});
float outputValue = fill[coords[3]];`;
}
this.userCode = `
void main() {
ivec4 coords = getOutputCoords();
int x = coords[2];
int y = coords[1];
float coordXFloat = (float(x) - params[0]) * params[3] -
(float(y) - params[1]) * params[2];
float coordYFloat = (float(x) - params[0]) * params[2] +
(float(y) - params[1]) * params[3];
int coordX = int(round(coordXFloat + params[0]));
int coordY = int(round(coordYFloat + params[1]));
${fillSnippet}
if(coordX >= 0 && coordX < ${imageWidth} && coordY >= 0 && coordY < ${imageHeight}) {
outputValue = getImage(coords[0], coordY, coordX, coords[3]);
}
setOutput(outputValue);
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const rotateWithOffsetConfig$1 = {
kernelName: RotateWithOffset,
backendName: 'webgl',
kernelFunc: ({ inputs, attrs, backend }) => {
const { image } = inputs;
const { radians, fillValue, center } = attrs;
const webglBackend = backend;
const program = new RotateProgram(image.shape, fillValue);
const [centerX, centerY] = getImageCenter(center, image.shape[1], image.shape[2]);
const customValues = [[centerX, centerY, Math.sin(radians), Math.cos(radians)]];
const output = webglBackend.runWebGLProgram(program, [image], image.dtype, customValues);
return output;
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ROUND = `
// OpenGL ES does not support round function.
// The algorithm is based on banker's rounding.
float base = floor(x);
if ((x - base) < 0.5) {
return floor(x);
} else if ((x - base) > 0.5) {
return ceil(x);
} else {
if (mod(base, 2.0) == 0.0) {
return base;
} else {
return base + 1.0;
}
}
`;
const round$1 = unaryKernelFunc({ opSnippet: ROUND });
const roundConfig$1 = {
kernelName: Round,
backendName: 'webgl',
kernelFunc: round$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const RSQRT = `return inversesqrt(x);`;
const rsqrt = unaryKernelFunc({ opSnippet: RSQRT, cpuKernelImpl: rsqrtImplCPU });
const rsqrtConfig = {
kernelName: Rsqrt,
backendName: 'webgl',
kernelFunc: rsqrt
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ScatterProgram {
constructor(updateSize, sliceDim, indicesRank, updatesRank, strides, shape, summingDupeIndex = true, defaultIsTensor = false) {
this.variableNames = ['updates', 'indices', 'defaultValue'];
this.outputShape = shape;
const stridesType = getCoordsDataType(strides.length);
const dtype = getCoordsDataType(shape.length);
let indicesString = '';
if (indicesRank === 1) {
indicesString = 'i';
}
else if (indicesRank === 2) {
indicesString = 'i, j';
}
const indicesSnippet = `getIndices(${indicesString})`;
let updatesString = '';
if (updatesRank === 1) {
updatesString = 'i';
}
else if (updatesRank === 2) {
updatesString = 'i, coords[1]';
}
const updatesSnippet = `getUpdates(${updatesString})`;
let defaultValuesString = '';
if (defaultIsTensor) {
defaultValuesString = 'coords[0], coords[1]';
}
const defaultValueSnippet = `getDefaultValue(${defaultValuesString})`;
const strideString = sliceDim > 1 ? 'strides[j]' : 'strides';
this.userCode = `
${stridesType} strides = ${stridesType}(${strides});
void main() {
${dtype} coords = getOutputCoords();
float sum = 0.0;
bool found = false;
for (int i = 0; i < ${updateSize}; i++) {
int flattenedIndex = 0;
for (int j = 0; j < ${sliceDim}; j++) {
int index = round(${indicesSnippet});
flattenedIndex += index * ${strideString};
}
if (flattenedIndex == coords[0]) {
sum += ${updatesSnippet};
found = true;
}
}
setOutput(mix(${defaultValueSnippet}, sum, float(found)));
}
`;
}
}
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class ScatterPackedProgram {
constructor(updateSize, sliceDim, indicesRank, updatesRank, strides, shape, summingDupeIndex = true, defaultIsTensor = false) {
this.variableNames = ['updates', 'indices', 'defaultValue'];
this.packedInputs = true;
this.packedOutput = true;
this.outputShape = shape;
const stridesType = getCoordsDataType(strides.length);
const dtype = getCoordsDataType(shape.length);
let indicesString = '';
if (indicesRank === 1) {
indicesString = 'i';
}
else if (indicesRank === 2) {
indicesString = 'i, j';
}
const indicesSnippet = `getIndices(${indicesString})`;
let updatesString = '';
if (updatesRank === 1) {
updatesString = 'i';
}
else if (updatesRank === 2) {
updatesString = 'i, coords[1]';
}
const updatesSnippet = `getUpdates(${updatesString})`;
let defaultValuesString = '';
if (defaultIsTensor) {
defaultValuesString = 'coords[0], coords[1]';
}
const defaultValueSnippet = `getDefaultValue(${defaultValuesString})`;
const strideString = sliceDim > 1 ? 'strides[j]' : 'strides';
const strideString2 = sliceDim > 1 ? 'strides[j + 1]' : 'strides';
this.userCode = `
${stridesType} strides = ${stridesType}(${strides});
void main() {
${dtype} coords = getOutputCoords();
vec4 sum = vec4(0.);
vec4 found = vec4(0.);
for (int i = 0; i < ${updateSize}; i+=2) {
ivec2 flattenedIndex = ivec2(0);
for (int j = 0; j < ${sliceDim}; j+=2) {
ivec4 index = round(${indicesSnippet});
flattenedIndex += index.xz * ${strideString};
if (j + 1 < ${sliceDim}) {
flattenedIndex += index.yw * ${strideString2};
}
}
if (flattenedIndex[0] == coords[0] || flattenedIndex[1] == coords[0] ||
flattenedIndex[0] == coords[0] + 1 || flattenedIndex[1] == coords[0] + 1) {
vec4 updVals = ${updatesSnippet};
if (flattenedIndex[0] == coords[0]) {
sum.xy += updVals.xy;
found.xy = vec2(1.);
} else if (flattenedIndex[0] == coords[0] + 1) {
sum.zw += updVals.xy;
found.zw = vec2(1.);
}
if (flattenedIndex[1] == coords[0]) {
sum.xy += updVals.zw;
found.xy = vec2(1.);
} else if (flattenedIndex[1] == coords[0] + 1) {
sum.zw += updVals.zw;
found.zw = vec2(1.);
}
}
}
setOutput(mix(${defaultValueSnippet}, sum, found));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function scatterNd$1(args) {
const { inputs, backend, attrs } = args;
const { indices, updates } = inputs;
const { shape } = attrs;
const { sliceRank, numUpdates, sliceSize, strides, outputSize } = calculateShapes(updates, indices, shape);
const flattenShape = [outputSize / sliceSize, sliceSize];
if (outputSize === 0) {
return backend.makeTensorInfo(shape, indices.dtype);
}
const flattenIndices = reshape$1({ inputs: { x: indices }, backend, attrs: { shape: [numUpdates, sliceRank] } });
const flattenX = reshape$1({ inputs: { x: updates }, backend, attrs: { shape: [numUpdates, sliceSize] } });
const defaultValue = backend.makeTensorInfo([], 'float32', new Float32Array([0])); // scalar(0)
let program;
if (env().getBool('WEBGL_PACK')) {
program = new ScatterPackedProgram(numUpdates, sliceRank, flattenIndices.shape.length, flattenX.shape.length, strides, flattenShape);
}
else {
program = new ScatterProgram(numUpdates, sliceRank, flattenIndices.shape.length, flattenX.shape.length, strides, flattenShape);
}
const res = backend.runWebGLProgram(program, [flattenX, flattenIndices, defaultValue], flattenX.dtype);
const reshaped = reshape$1({ inputs: { x: res }, backend, attrs: { shape } });
backend.disposeIntermediateTensorInfo(flattenIndices);
backend.disposeIntermediateTensorInfo(flattenX);
backend.disposeIntermediateTensorInfo(res);
backend.disposeIntermediateTensorInfo(defaultValue);
return reshaped;
}
const scatterNdConfig$1 = {
kernelName: ScatterNd,
backendName: 'webgl',
kernelFunc: scatterNd$1
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class SearchSortedProgram {
constructor(batchSize, numInputs, numValues, side) {
this.variableNames = ['sortedSequence', 'values'];
this.customUniforms = [{ name: 'numInputs', type: 'int' }];
this.outputShape = [batchSize, numValues];
const webGL2LoopHead = 'while (left < right) {';
// WebGL1 doesn't accept non constant loop conditions, so upper bound loop
// iterations.
const webGL1LoopHead = `for (int i = 0; i < ${Math.ceil(Math.log2(numInputs + 1))}; ++i) { if (left >= right) break;`;
const loopHead = env().getNumber('WEBGL_VERSION') === 2 ? webGL2LoopHead :
webGL1LoopHead;
// left corresponds to lower bound and right to upper bound.
const boundComparator = side === 'left' ? '<' : '<=';
this.userCode = `
int findBound(int batch, float value) {
int left = 0;
int right = numInputs;
int mid;
${loopHead}
mid = (left + right) / 2;
if (getSortedSequence(batch, mid) ${boundComparator} value) {
left = mid + 1;
} else {
right = mid;
}
}
return right;
}
void main() {
ivec2 coords = getOutputCoords();
int batch = coords[0];
int valueIndex = coords[1];
float value = getValues(batch, valueIndex);
setOutput(float(findBound(batch, value)));
}
`;
}
}
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function searchSorted$1(args) {
const { inputs, backend, attrs } = args;
const { sortedSequence, values } = inputs;
const { side } = attrs;
const program = new SearchSortedProgram(sortedSequence.shape[0], sortedSequence.shape[1], values.shape[1], side);
const customValues = [[sortedSequence.shape[1]]];
return backend.runWebGLProgram(program, [sortedSequence, values], 'int32', customValues);
}
const searchSortedConfig$1 = {
kernelName: SearchSorted,
backendName: 'webgl',
kernelFunc: searchSorted$1,
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class SelectProgram {
constructor(cRank, shape, rank) {
this.variableNames = ['c', 'a', 'b'];
this.outputShape = shape;
let cCoords;
let abCoords;
if (rank > 4) {
throw Error(`Where for rank ${rank} is not yet supported`);
}
if (rank === 1) {
abCoords = `resRC`;
cCoords = `resRC`;
}
else {
const currentCoords = ['resRC.x', 'resRC.y', 'resRC.z', 'resRC.w'];
const cCoordVars = [];
const abCoordVars = [];
for (let i = 0; i < shape.length; i++) {
abCoordVars.push(`${currentCoords[i]}`);
if (i < cRank) {
cCoordVars.push(`${currentCoords[i]}`);
}
}
cCoords = cCoordVars.join();
abCoords = abCoordVars.join();
}
const dtype = getCoordsDataType(rank);
this.userCode = `
void main() {
${dtype} resRC = getOutputCoords();
float cVal = getC(${cCoords});
if (cVal >= 1.0) {
setOutput(getA(${abCoords}));
} else {
setOutput(getB(${abCoords}));
}
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function select$1(args) {
const { inputs, backend } = args;
const { condition, t, e } = inputs;
const program = new SelectProgram(condition.shape.length, t.shape, t.shape.length);
return backend.runWebGLProgram(program, [condition, t, e], upcastType(t.dtype, e.dtype));
}
const selectConfig$1 = {
kernelName: Select,
backendName: 'webgl',
kernelFunc: select$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SELU = `
// Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
// see: https://arxiv.org/abs/1706.02515
float scaleAlpha = ${SELU_SCALEALPHA};
float scale = ${SELU_SCALE};
return (x >= 0.0) ? scale * x : scaleAlpha * (exp(x) - 1.0);
`;
const selu$1 = unaryKernelFunc({ opSnippet: SELU });
const seluConfig$1 = {
kernelName: Selu$1,
backendName: 'webgl',
kernelFunc: selu$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SIGMOID = CHECK_NAN_SNIPPET_UNARY + `
return 1.0 / (1.0 + exp(-1.0 * x));
`;
const SIGMOID_PACKED = `
vec4 result = 1.0 / (1.0 + exp(-1.0 * x));
bvec4 isNaN = isnan(x);
result.r = isNaN.r ? x.r : result.r;
result.g = isNaN.g ? x.g : result.g;
result.b = isNaN.b ? x.b : result.b;
result.a = isNaN.a ? x.a : result.a;
return result;
`;
const sigmoid = unaryKernelFunc({
opSnippet: SIGMOID,
packedOpSnippet: SIGMOID_PACKED,
cpuKernelImpl: sigmoidImplCPU
});
const sigmoidConfig = {
kernelName: Sigmoid$1,
backendName: 'webgl',
kernelFunc: sigmoid,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Sign does not propagate NANs.
const SIGN = `
if (isnan(x)) { return 0.0; }
return sign(x);
`;
const sign$1 = unaryKernelFunc({ opSnippet: SIGN });
const signConfig$1 = {
kernelName: Sign,
backendName: 'webgl',
kernelFunc: sign$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SIN = CHECK_NAN_SNIPPET_UNARY + `
return sin(x);
`;
const SIN_PACKED = `
vec4 result = sin(x);
bvec4 isNaN = isnan(x);
${CHECK_NAN_SNIPPET_PACKED}
return result;
`;
const sin$1 = unaryKernelFunc({ opSnippet: SIN, packedOpSnippet: SIN_PACKED });
const sinConfig$1 = {
kernelName: Sin,
backendName: 'webgl',
kernelFunc: sin$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SINH = `
float e2x = exp(x);
return (e2x - 1.0 / e2x) / 2.0;
`;
const sinh$1 = unaryKernelFunc({ opSnippet: SINH });
const sinhConfig$1 = {
kernelName: Sinh,
backendName: 'webgl',
kernelFunc: sinh$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SOFTPLUS = `
float epsilon = 1.1920928955078125e-7;
float threshold = log(epsilon) + 2.0;
bool too_large = x > -threshold;
bool too_small = x < threshold;
float result;
float exp_x = exp(x);
if (too_large){
result = x;
}
else if (too_small){
result = exp_x;
}
else{
result = log(exp_x + 1.0);
}
return result;
`;
const softplus$1 = unaryKernelFunc({ opSnippet: SOFTPLUS });
const softplusConfig$1 = {
kernelName: Softplus$1,
backendName: 'webgl',
kernelFunc: softplus$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const spaceToBatchND$1 = (args) => {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { blockShape, paddings } = attrs;
assert$1(x.shape.length <= 4, () => 'spaceToBatchND for rank > 4 with a WebGL backend not ' +
'implemented yet');
const prod = blockShape.reduce((a, b) => a * b);
const completePaddings = [[0, 0]];
completePaddings.push(...paddings);
for (let i = 1 + blockShape.length; i < x.shape.length; ++i) {
completePaddings.push([0, 0]);
}
const toDispose = [];
const paddedX = padV2$1({
inputs: { x },
backend,
attrs: { paddings: completePaddings, constantValue: 0 }
});
const reshapedPaddedShape = getReshaped(paddedX.shape, blockShape, prod, false);
const permutedReshapedPaddedPermutation = getPermuted(reshapedPaddedShape.length, blockShape.length, false);
const flattenShape = getReshapedPermuted(paddedX.shape, blockShape, prod, false);
const reshapedPaddedX = reshape$1({ inputs: { x: paddedX }, backend, attrs: { shape: reshapedPaddedShape } });
const paddedXT = transpose({
inputs: { x: reshapedPaddedX },
backend,
attrs: { perm: permutedReshapedPaddedPermutation }
});
const result = reshape$1({ inputs: { x: paddedXT }, backend, attrs: { shape: flattenShape } });
toDispose.push(paddedX);
toDispose.push(reshapedPaddedX);
toDispose.push(paddedXT);
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return result;
};
const spaceToBatchNDConfig$1 = {
kernelName: SpaceToBatchND,
backendName: 'webgl',
kernelFunc: spaceToBatchND$1
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseFillEmptyRows$1(args) {
const { inputs, backend } = args;
const { indices, values, denseShape, defaultValue } = inputs;
if (denseShape.shape.length !== 1) {
throw new Error(`Dense shape must be a vector, saw:
${denseShape.shape}`);
}
if (indices.shape.length !== 2) {
throw new Error(`Indices must be a matrix, saw:
${indices.shape}`);
}
if (values.shape.length !== 1) {
throw new Error(`Values must be a vector, saw:
${values.shape}`);
}
if (defaultValue.shape.length !== 0) {
throw new Error(`Default value must be a scalar, saw:
${defaultValue.shape}`);
}
const $indices = backend.readSync(indices.dataId);
const $values = backend.readSync(values.dataId);
const $denseShape = backend.readSync(denseShape.dataId);
const $defaultValue = backend.readSync(defaultValue.dataId)[0];
const [outputIndices, outputIndicesShape, outputValues, emptyRowIndicator, reverseIndexMap] = sparseFillEmptyRowsImplCPU($indices, indices.shape, indices.dtype, $values, values.dtype, $denseShape, $defaultValue);
return [
backend.makeTensorInfo(outputIndicesShape, indices.dtype, outputIndices),
backend.makeTensorInfo([outputIndicesShape[0]], values.dtype, outputValues),
backend.makeTensorInfo([emptyRowIndicator.length], 'bool', new Uint8Array(emptyRowIndicator.map((value) => Number(value)))),
backend.makeTensorInfo([reverseIndexMap.length], indices.dtype, new Int32Array(reverseIndexMap)),
];
}
const sparseFillEmptyRowsConfig$1 = {
kernelName: SparseFillEmptyRows,
backendName: 'webgl',
kernelFunc: sparseFillEmptyRows$1,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseReshape$1(args) {
const { inputs, backend } = args;
const { inputIndices, inputShape, newShape } = inputs;
if (inputIndices.shape.length !== 2) {
throw new Error(`Input indices should be a matrix but received shape ${inputIndices.shape}`);
}
if (inputShape.shape.length !== 1) {
throw new Error(`Input shape should be a vector but received shape ${inputShape.shape}`);
}
if (newShape.shape.length !== 1) {
throw new Error(`Target shape should be a vector but received shape ${newShape.shape}`);
}
const $inputShape = Array.from(backend.readSync(inputShape.dataId));
const $inputIndices = backend.readSync(inputIndices.dataId);
const targetShape = Array.from(backend.readSync(newShape.dataId));
const [newIndices, indicesShape, outputShape] = sparseReshapeImplCPU($inputIndices, inputIndices.shape, inputIndices.dtype, $inputShape, targetShape);
return [
backend.makeTensorInfo(indicesShape, inputIndices.dtype, newIndices),
backend.makeTensorInfo([outputShape.length], newShape.dtype, new Int32Array(outputShape)),
];
}
const sparseReshapeConfig$1 = {
kernelName: SparseReshape,
backendName: 'webgl',
kernelFunc: sparseReshape$1,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseSegmentMean$1(args) {
const { inputs, backend } = args;
const { data, indices, segmentIds } = inputs;
if (data.shape.length < 1) {
throw new Error(`Data should be at least 1 dimensional but received scalar`);
}
if (indices.shape.length !== 1) {
throw new Error(`Indices should be a vector but received shape
${indices.shape}`);
}
if (segmentIds.shape.length !== 1) {
throw new Error(`Segment ids should be a vector but received shape
${segmentIds.shape}`);
}
const $data = backend.readSync(data.dataId);
const $indices = backend.readSync(indices.dataId);
const $segmentIds = backend.readSync(segmentIds.dataId);
const [outputData, outputDataShape] = sparseSegmentReductionImplCPU($data, data.shape, data.dtype, $indices, $segmentIds, true);
return backend.makeTensorInfo(outputDataShape, data.dtype, outputData);
}
const sparseSegmentMeanConfig$1 = {
kernelName: SparseSegmentMean,
backendName: 'webgl',
kernelFunc: sparseSegmentMean$1,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseSegmentSum$1(args) {
const { inputs, backend } = args;
const { data, indices, segmentIds } = inputs;
if (data.shape.length < 1) {
throw new Error(`Data should be at least 1 dimensional but received scalar`);
}
if (indices.shape.length !== 1) {
throw new Error(`Indices should be a vector but received shape
${indices.shape}`);
}
if (segmentIds.shape.length !== 1) {
throw new Error(`Segment ids should be a vector but received shape
${segmentIds.shape}`);
}
const $data = backend.readSync(data.dataId);
const $indices = backend.readSync(indices.dataId);
const $segmentIds = backend.readSync(segmentIds.dataId);
const [outputData, outputDataShape] = sparseSegmentReductionImplCPU($data, data.shape, data.dtype, $indices, $segmentIds);
return backend.makeTensorInfo(outputDataShape, data.dtype, outputData);
}
const sparseSegmentSumConfig$1 = {
kernelName: SparseSegmentSum,
backendName: 'webgl',
kernelFunc: sparseSegmentSum$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseToDense$1(args) {
const { inputs, backend, attrs } = args;
const { sparseIndices, sparseValues, defaultValue } = inputs;
const { outputShape } = attrs;
const { sliceRank, numUpdates, sliceSize, strides, outputSize } = calculateShapes(sparseValues, sparseIndices, outputShape);
const sumDupeIndices = false;
if (sparseValues.dtype === 'string') {
const indicesBuf = backend.bufferSync(sparseIndices);
const updatesBuf = backend.bufferSync(sparseValues);
const $defaultValue = decodeString(backend.readSync(defaultValue.dataId)[0]);
const outBuf = scatterImplCPU(indicesBuf, updatesBuf, outputShape, outputSize, sliceSize, numUpdates, sliceRank, strides, $defaultValue, sumDupeIndices);
return backend.makeTensorInfo(outputShape, outBuf.dtype, outBuf.values);
}
const program = new ScatterProgram(numUpdates, sliceRank, sparseIndices.shape.length, sparseValues.shape.length, strides, [outputSize, 1], sumDupeIndices);
const res = backend.runWebGLProgram(program, [sparseValues, sparseIndices, defaultValue], sparseValues.dtype);
const reshaped = reshape$1({ inputs: { x: res }, backend, attrs: { shape: outputShape } });
backend.disposeIntermediateTensorInfo(res);
return reshaped;
}
const sparseToDenseConfig$1 = {
kernelName: SparseToDense,
backendName: 'webgl',
kernelFunc: sparseToDense$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function splitV$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { numOrSizeSplits, axis } = attrs;
const $axis = parseAxisParam(axis, x.shape)[0];
const splitSizes = prepareSplitSize(x, numOrSizeSplits, $axis);
const xRank = x.shape.length;
const begin = new Array(xRank).fill(0);
const size = x.shape.slice();
return splitSizes.map(s => {
const sliceSize = [...size];
sliceSize[$axis] = s;
const sliceT = slice({ inputs: { x }, backend, attrs: { begin, size: sliceSize } });
begin[$axis] += s;
return sliceT;
});
}
const splitVConfig$1 = {
kernelName: SplitV,
backendName: 'webgl',
kernelFunc: splitV$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SQRT = `return sqrt(x);`;
const sqrt = unaryKernelFunc({ opSnippet: SQRT, packedOpSnippet: SQRT, cpuKernelImpl: sqrtImplCPU });
const sqrtConfig = {
kernelName: Sqrt,
backendName: 'webgl',
kernelFunc: sqrt
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SQUARE = `return x * x;`;
const square$1 = unaryKernelFunc({ opSnippet: SQUARE });
const squareConfig$1 = {
kernelName: Square,
backendName: 'webgl',
kernelFunc: square$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const SQUARED_DIFFERENCE = 'return (a - b) * (a - b);';
const squaredDifference = binaryKernelFunc({ opSnippet: SQUARED_DIFFERENCE, packedOpSnippet: SQUARED_DIFFERENCE });
const squaredDifferenceConfig = {
kernelName: SquaredDifference,
backendName: 'webgl',
kernelFunc: squaredDifference,
};
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function staticRegexReplace(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
if (x.dtype !== 'string') {
throw new Error('Input must be of datatype string');
}
const $x = backend.readSync(x.dataId);
const stringInput = fromUint8ToStringArray($x);
const output = staticRegexReplaceImplCPU(stringInput, 'string', attrs);
return backend.makeTensorInfo(x.shape, 'string', output);
}
const staticRegexReplaceConfig = {
kernelName: StaticRegexReplace,
backendName: 'webgl',
kernelFunc: staticRegexReplace,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function step$1({ inputs, attrs, backend }) {
const { x } = inputs;
const opSnippet = CHECK_NAN_SNIPPET$1 + `
return x > 0.0 ? 1.0 : float(${attrs.alpha});
`;
const program = new UnaryOpProgram(x.shape, opSnippet);
return backend.runWebGLProgram(program, [x], x.dtype);
}
const stepConfig$1 = {
kernelName: Step,
backendName: 'webgl',
kernelFunc: step$1,
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class StridedSliceProgram {
constructor(begin, strides, size) {
this.variableNames = ['x'];
this.outputShape = size;
const rank = size.length;
const inputDtype = getCoordsDataType(size.length);
const dtype = getCoordsDataType(size.length);
let newCoords = '';
if (rank === 1) {
newCoords = 'coords * strides + begin';
}
else {
let outputAxis = 0;
newCoords =
size.map((_, i) => {
outputAxis++;
return size.length === 1 ?
`coords * strides[${i}] + begin[${i}]` :
`coords[${outputAxis - 1}] * strides[${i}] + begin[${i}]`;
})
.join(',');
}
this.userCode = `
${inputDtype} begin = ${inputDtype}(${begin});
${inputDtype} strides = ${inputDtype}(${strides});
void main() {
${dtype} coords = getOutputCoords();
setOutput(getX(${newCoords}));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stridedSlice$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { begin, end, strides, beginMask, endMask, ellipsisMask, newAxisMask, shrinkAxisMask } = attrs;
const { finalShapeSparse, finalShape, isIdentity, sliceDim0, isSimpleSlice, begin: $begin, end: $end, strides: $strides } = sliceInfo(x.shape, begin, end, strides, beginMask, endMask, ellipsisMask, newAxisMask, shrinkAxisMask);
let result;
if (isIdentity) {
// Optimization #1, slice is a no-op plus reshape
result = reshape$1({ inputs: { x }, backend, attrs: { shape: finalShape } });
}
else if (sliceDim0 || isSimpleSlice) {
// Optimization #2, slice is memory contiguous (only occurs in dim 0)
assert$1(x.shape.length >= 1, () => `Input must have rank at least 1, got: ${x.shape.length}`);
const size = computeOutShape$2($begin, $end, $strides);
// To tolerate begin[0] > end[0] (a 0-output slice), we min(begin, end).
const sliced = slice({ inputs: { x }, backend, attrs: { begin: $begin, size } });
result =
reshape$1({ inputs: { x: sliced }, backend, attrs: { shape: finalShape } });
backend.disposeIntermediateTensorInfo(sliced);
}
else {
const shouldExecuteOnCPU = backend.shouldExecuteOnCPU([x]);
if (shouldExecuteOnCPU) {
// tslint:disable-next-line: no-unnecessary-type-assertion
const values = backend.readSync(x.dataId);
// tslint:disable-next-line: no-unnecessary-type-assertion
const xBuf = buffer(x.shape, x.dtype, values);
const resultValues = stridedSliceImplCPU(finalShapeSparse, xBuf, $strides, $begin);
result = backend.makeTensorInfo(finalShape, x.dtype, resultValues.values);
}
else {
const program = new StridedSliceProgram($begin, $strides, finalShapeSparse);
result = backend.runWebGLProgram(program, [x], x.dtype);
}
}
const resultReshaped = reshape$1({ inputs: { x: result }, backend, attrs: { shape: finalShape } });
backend.disposeIntermediateTensorInfo(result);
return resultReshaped;
}
const stridedSliceConfig$1 = {
kernelName: StridedSlice,
backendName: 'webgl',
kernelFunc: stridedSlice$1
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stringNGrams$1(args) {
const { inputs, backend, attrs } = args;
const { separator, nGramWidths, leftPad, rightPad, padWidth, preserveShortSequences } = attrs;
const { data, dataSplits } = inputs;
const $data = backend.readSync(data.dataId);
const $dataSplits = backend.readSync(dataSplits.dataId);
const [nGrams, nGramsSplits] = stringNGramsImplCPU($data, $dataSplits, separator, nGramWidths, leftPad, rightPad, padWidth, preserveShortSequences);
return [
backend.makeTensorInfo([nGrams.length], 'string', nGrams),
backend.makeTensorInfo(dataSplits.shape, 'int32', nGramsSplits),
];
}
const stringNGramsConfig$1 = {
kernelName: StringNGrams,
backendName: 'webgl',
kernelFunc: stringNGrams$1,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stringSplit$1(args) {
const { inputs, backend, attrs } = args;
const { skipEmpty } = attrs;
const { input, delimiter } = inputs;
if (input.dtype !== 'string') {
throw new Error('Input must be of datatype string');
}
if (input.shape.length !== 1) {
throw new Error(`Input must be a vector, got shape: ${input.shape}`);
}
if (delimiter.shape.length !== 0) {
throw new Error(`Delimiter must be a scalar, got shape: ${delimiter.shape}`);
}
const $input = backend.readSync(input.dataId);
const $delimiter = backend.readSync(delimiter.dataId)[0];
const [indices, values, shape] = stringSplitImplCPU($input, $delimiter, skipEmpty);
const outputSize = values.length;
return [
backend.makeTensorInfo([outputSize, 2], 'int32', indices),
backend.makeTensorInfo([outputSize], 'string', values),
backend.makeTensorInfo([2], 'int32', new Int32Array(shape))
];
}
const stringSplitConfig$1 = {
kernelName: StringSplit,
backendName: 'webgl',
kernelFunc: stringSplit$1,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stringToHashBucketFast$1(args) {
const { inputs, backend, attrs } = args;
const { numBuckets } = attrs;
const { input } = inputs;
if (input.dtype !== 'string') {
throw new Error('Input must be of datatype string');
}
if (numBuckets <= 0) {
throw new Error(`Number of buckets must be at least 1`);
}
const $input = backend.readSync(input.dataId);
const output = stringToHashBucketFastImplCPU($input, numBuckets);
return backend.makeTensorInfo(input.shape, 'int32', output);
}
const stringToHashBucketFastConfig$1 = {
kernelName: StringToHashBucketFast,
backendName: 'webgl',
kernelFunc: stringToHashBucketFast$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const TAN = `return tan(x);`;
const tan$1 = unaryKernelFunc({ opSnippet: TAN });
const tanConfig$1 = {
kernelName: Tan,
backendName: 'webgl',
kernelFunc: tan$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const TANH = `
float e2x = exp(-2.0 * abs(x));
return sign(x) * (1.0 - e2x) / (1.0 + e2x);
`;
const tanh$1 = unaryKernelFunc({ opSnippet: TANH });
const tanhConfig$1 = {
kernelName: Tanh$1,
backendName: 'webgl',
kernelFunc: tanh$1,
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function tensorScatterUpdate$1(args) {
const { inputs, backend} = args;
const { tensor, indices, updates } = inputs;
const { sliceRank, numUpdates, sliceSize, strides, outputSize } = calculateShapes(updates, indices, tensor.shape);
const flattenShape = [outputSize / sliceSize, sliceSize];
if (outputSize === 0) {
return backend.makeTensorInfo(tensor.shape, indices.dtype);
}
const flattenIndices = reshape$1({ inputs: { x: indices }, backend, attrs: { shape: [numUpdates, sliceRank] } });
const flattenX = reshape$1({ inputs: { x: updates }, backend, attrs: { shape: [numUpdates, sliceSize] } });
const flattenTensor = reshape$1({ inputs: { x: tensor }, backend, attrs: { shape: flattenShape } });
const program = new ScatterProgram(numUpdates, sliceRank, flattenIndices.shape.length, flattenX.shape.length, strides, flattenShape, false, true);
const res = backend.runWebGLProgram(program, [flattenX, flattenIndices, flattenTensor], flattenTensor.dtype);
const reshaped = reshape$1({ inputs: { x: res }, backend, attrs: { shape: tensor.shape } });
backend.disposeIntermediateTensorInfo(flattenIndices);
backend.disposeIntermediateTensorInfo(flattenX);
backend.disposeIntermediateTensorInfo(flattenTensor);
backend.disposeIntermediateTensorInfo(res);
return reshaped;
}
const tensorScatterUpdateConfig$1 = {
kernelName: TensorScatterUpdate,
backendName: 'webgl',
kernelFunc: tensorScatterUpdate$1
};
/**
* @license
* Copyright 2017 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class TileProgram {
constructor(aShape, reps) {
this.variableNames = ['A'];
const outputShape = new Array(aShape.length);
for (let i = 0; i < outputShape.length; i++) {
outputShape[i] = aShape[i] * reps[i];
}
this.outputShape = outputShape;
this.rank = outputShape.length;
const dtype = getCoordsDataType(this.rank);
const sourceCoords = getSourceCoords(aShape);
this.userCode = `
void main() {
${dtype} resRC = getOutputCoords();
setOutput(getA(${sourceCoords}));
}
`;
}
}
function getSourceCoords(aShape) {
const rank = aShape.length;
if (rank > 5) {
throw Error(`Tile for rank ${rank} is not yet supported`);
}
if (rank === 1) {
return `imod(resRC, ${aShape[0]})`;
}
const currentCoords = ['resRC.x', 'resRC.y', 'resRC.z', 'resRC.w', 'resRC.u'];
const sourceCoords = [];
for (let i = 0; i < aShape.length; i++) {
sourceCoords.push(`imod(${currentCoords[i]}, ${aShape[i]})`);
}
return sourceCoords.join();
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function tile$2(params) {
const { inputs, backend, attrs } = params;
const { x } = inputs;
const { reps } = attrs;
// tile gpu program cannot handle rank > 5 case.
if (x.dtype === 'string' || x.shape.length > 5) {
// Even thought string tensor is always on CPU, just to be consistent on how
// to access tensor data.
const data = backend.readSync(x.dataId);
const value = x.dtype === 'string' ?
data.map(d => decodeString(d)) :
data;
const buf = buffer(x.shape, x.dtype, value);
const outBuf = tileImplCPU(buf, reps);
return backend.makeTensorInfo(outBuf.shape, outBuf.dtype, outBuf.values);
}
const program = new TileProgram(x.shape, reps);
const output = backend.runWebGLProgram(program, [x], x.dtype);
return output;
}
const tileConfig$1 = {
kernelName: Tile,
backendName: 'webgl',
kernelFunc: tile$2,
};
// Based on Algorithm 2 of Bitonic Top K, ref:
// https://anilshanbhag.in/static/papers/gputopk_sigmod18.pdf
// The original algorithm is based on computing the top K only, however
// since for TFJS we require the indices of the top K values as well then the
// algorithm found here is a bit modified. Rather than producing the values
// at each step, the indices containing the top K are generated instead.
// The output values are not generated to reduce the number of outputs in the
// GPU, the values can easily be retrieved from the indices using a gather
// op.
class SwapProgram {
/**
* @param shape desired output shape (can be larger than input shape, output
* will be padded with -Infinity)
*/
constructor(shape) {
this.variableNames = ['x', 'indices'];
// |n| Size of the original input of TopK.
// |firstPass|indicates if this is the first time swap is being used which
// means no indices input containing the top K is present yet.
// |inc| Swaps pairs of indices (0, inc), (1, inc + 1), (2, inc + 2) ...
this.customUniforms = [
{ name: 'n', type: 'int' },
{ name: 'firstPass', type: 'int' },
{ name: 'negativeInf', type: 'float' },
{ name: 'dir', type: 'int' },
{ name: 'inc', type: 'int' }
];
this.outputShape = shape;
this.userCode = `
void main() {
ivec2 coords = getOutputCoords();
int batch = coords[0];
int elemIdx = coords[1];
// We compare elements pair-wise within a group of size 2 * inc.
// The comparing rule for each group alternates between ascending
// and descending. Within each group, we compare each pair at
// positions i and i+inc. To decide whether an element at position i
// is x0 or x1, we mod it by 2 * inc, if the result is smaller than
// inc, it is in the first half of the group, we denote it as x0,
// otherwise we denote it as x1.
// For example, as shown in the Bitonic top K paper referenced above,
// Figure5(a) shows that element[1] is in the
// second half of the group when group size is 2, but it is in the
// first half of the group when group size is 4.
bool isFirstInPair = imod(elemIdx, 2 * inc) < inc;
int i = isFirstInPair ? elemIdx : elemIdx - inc;
int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
int i1 = firstPass == 1 ? i + inc : int(getIndices(batch, i + inc));
float x0 = i0 < n ? getX(batch, i0) : negativeInf;
float x1 = i1 < n ? getX(batch, i1) : negativeInf;
// Denotes which direction indices are in (ascending or descending).
bool reverse = imod(elemIdx, 2 * dir) >= dir;
bool isGreater = x0 > x1 || (x0 == x1 && i1 > i0);
if (reverse == isGreater) { // Elements in opposite order of direction
int iTemp = i0;
i0 = i1;
i1 = iTemp;
}
if (isFirstInPair) {
setOutput(float(i0));
} else {
setOutput(float(i1));
}
}
`;
}
}
class MergeProgram {
/**
* @param shape desired output shape (must be half of the input size)
*/
constructor(shape) {
this.variableNames = ['x', 'indices'];
// |n| Size of the original input of TopK
// |firstPass| indicates if this is the first time swap is being used which
// means no indices input containing the top K is present yet.
// |k| Top k elements desired
this.customUniforms = [
{ name: 'n', type: 'int' },
{ name: 'firstPass', type: 'int' },
{ name: 'k', type: 'int' }
];
this.outputShape = shape;
this.userCode = `
void main() {
// Takes max of indices (0, k), (1, k + 1), (2, k + 2) ...
ivec2 coords = getOutputCoords();
int batch = coords[0];
int elemIdx = coords[1];
// The output size is half of the previous size.
// If the previous sequence is | | | | _ _ _ _ | | | | _ _ _ _ (k=4),
// we only need to output the indices at positions |, the indices at
// positions _ can be thrown away, see Figure5(b) After Phase 2
// (Merge phase) in the Bitonic Top K paper referenced above.
// For example, the paper shows we only need to output the orange bars.
// The output sequence should look like this | | | | | | | |.
// Because the sequence is halved, to map the output index back
// to the previous sequence to find the corresponding value,
// we need to double the index. When we double the index,
// we basically interpolate a position, so 2i looks like
// | _ | _ | _ | _ | _ | _ | _. We move the | to the first k position
// of each 2k positions by - elemIdx % k. E.g. for output at
// index 4,5,6,7, we want to get the corresponding element at
// original index 8,9,10,11, for output at index 8,9,10,11,
// we want to get the corresponding element at original index
// 16,17,18,19, so on and so forth.
int i = elemIdx < k ? elemIdx : (elemIdx * 2 - imod(elemIdx, k));
int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
int i1 = firstPass == 1 ? i + k : int(getIndices(batch, i + k));
float x0 = getX(batch, i0);
float x1 = i1 < n ? getX(batch, i1) : x0;
setOutput(x0 >= x1 ? float(i0) : float(i1));
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function disposeIntermediateTensorInfoOrNull(backend, tensorInfo) {
if (tensorInfo !== null) {
backend.disposeIntermediateTensorInfo(tensorInfo);
}
}
function roundUpToPow2(num) {
let pow2 = 1;
while (pow2 < num) {
pow2 *= 2;
}
return pow2;
}
// Based on Algorithm 2 of Bitonic Top K, ref:
// https://anilshanbhag.in/static/papers/gputopk_sigmod18.pdf
function topK$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { k, sorted } = attrs;
// Empirically determined constant used to determine last dim threshold for
// handing off execution to the CPU.
const TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD = env().getNumber('TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD');
// Empirically determined constant used to determine k threshold for handing
// off execution to the CPU.
const TOPK_K_CPU_HANDOFF_THRESHOLD = env().getNumber('TOPK_K_CPU_HANDOFF_THRESHOLD');
const xShape = x.shape;
const lastDim = xShape[xShape.length - 1];
if (backend.shouldExecuteOnCPU([x]) ||
lastDim < TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD ||
k > TOPK_K_CPU_HANDOFF_THRESHOLD) {
const xVals = backend.readSync(x.dataId);
const [allTopKVals, allTopKIndices] = topKImplCPU(xVals, xShape, x.dtype, k, sorted);
return [
backend.makeTensorInfo(allTopKVals.shape, allTopKVals.dtype, allTopKVals.values),
backend.makeTensorInfo(allTopKIndices.shape, allTopKIndices.dtype, allTopKIndices.values)
];
}
if (k === 0) {
xShape[xShape.length - 1] = 0;
return [
backend.makeTensorInfo(xShape, x.dtype, []),
backend.makeTensorInfo(xShape, 'int32', [])
];
}
if (lastDim === 1 /* firstPass */) {
return [
x, fill$1({ attrs: { shape: xShape, dtype: 'int32', value: 0 }, backend })
];
}
// Eagerly unpack x input since it is passed in to all the shaders which
// require unpacked inputs.
const xtexData = backend.texData.get(x.dataId);
const xIsPacked = xtexData !== null && xtexData.isPacked;
const xUnPacked = xIsPacked ? backend.unpackTensor(x) : x;
// Reshape into a 2d tensor [batch, lastDim] and compute topk along lastDim.
const xSize = sizeFromShape(xShape);
const batch = xSize / lastDim;
const x2D = reshape$1({ inputs: { x: xUnPacked }, attrs: { shape: [batch, lastDim] }, backend });
if (xIsPacked) {
disposeIntermediateTensorInfoOrNull(backend, xUnPacked);
}
const kPow2 = roundUpToPow2(k);
const lastDimPow2 = roundUpToPow2(lastDim);
// Only the indices containing the top K are kept at every step to reduce
// number of outputs in the GPU algorithms, so once the final set of indices
// is computed then gather is used to grab the corresponding values
// from the original input.
let indices = null;
// GPU algorithm always takes in an indices input but this input is not used
// on the first run of a GPU algorithm, therefore if indices is null we simply
// pass in x2D instead of it but the value will not actually be used
const getInputs = () => indices === null ? [x2D, x2D] : [x2D, indices];
const runSwap = (dir, inc, shape) => {
const inputs = getInputs();
const program = new SwapProgram(shape);
const fistPass = indices === null ? 1 : 0;
const customValues = [[lastDim], [fistPass], [Number.NEGATIVE_INFINITY], [dir], [inc]];
const prevIndices = indices;
indices = backend.runWebGLProgram(program, inputs, 'int32', customValues);
disposeIntermediateTensorInfoOrNull(backend, prevIndices);
};
// Step 1: local sort
for (let len = 1; len < kPow2; len *= 2) {
const dir = len * 2;
for (let inc = len; inc >= 1; inc /= 2) {
runSwap(dir, inc, [batch, lastDimPow2]);
}
}
// Step 2: merge
for (let indicesSize = lastDimPow2; indicesSize > kPow2; indicesSize /= 2) {
const inputs = getInputs();
const mergeProgram = new MergeProgram([batch, indicesSize / 2]);
const firstPass = indices === null ? 1 : 0;
const customValues = [[lastDim], [firstPass], [kPow2]];
const prevIndices = indices;
indices =
backend.runWebGLProgram(mergeProgram, inputs, 'int32', customValues);
disposeIntermediateTensorInfoOrNull(backend, prevIndices);
// Step 3: rebuild
const len = kPow2 / 2;
const dir = len * 2;
for (let inc = len; inc >= 1; inc /= 2) {
runSwap(dir, inc, indices.shape);
}
}
// Keep only the requested top K results instead of kPow2
let prevIndices = indices;
indices = slice({ inputs: { x: indices }, backend, attrs: { begin: 0, size: [batch, k] } });
disposeIntermediateTensorInfoOrNull(backend, prevIndices);
// Gather values on last dimension
let values = gatherV2$1({ inputs: { x: x2D, indices }, backend, attrs: { axis: 1, batchDims: 1 } });
disposeIntermediateTensorInfoOrNull(backend, x2D);
// Reshape back to the original input shape, except that the last
// dimension is k.
const newShape = xShape.slice(0, -1);
newShape.push(k);
prevIndices = indices;
indices = reshape$1({ inputs: { x: indices }, attrs: { shape: newShape }, backend });
disposeIntermediateTensorInfoOrNull(backend, prevIndices);
const prevValues = values;
values = reshape$1({ inputs: { x: values }, attrs: { shape: newShape }, backend });
disposeIntermediateTensorInfoOrNull(backend, prevValues);
return [values, indices];
}
const topKConfig$1 = {
kernelName: TopK,
backendName: 'webgl',
kernelFunc: topK$1
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class TransformProgram {
constructor(imageHeight, imageWidth, interpolation, fillMode, fillValue, outShape) {
this.variableNames = ['Image', 'Transforms'];
this.outputShape = outShape;
const interpolationModeId = interpolation === 'nearest' ? 1 : 2;
let fillModeId;
switch (fillMode) {
case 'constant':
fillModeId = 1;
break;
case 'reflect':
fillModeId = 2;
break;
case 'wrap':
fillModeId = 3;
break;
case 'nearest':
fillModeId = 4;
break;
default:
fillModeId = 1;
break;
}
this.userCode = `
float mapCoord(float outCoord, float len) {
float inCoord = outCoord;
if(${fillModeId} == 2) {
if (inCoord < 0.0) {
if (len <= 1.0) {
inCoord = 0.0;
} else {
float sz2 = 2.0 * len;
if (inCoord < sz2) {
inCoord = sz2 * float(int(float(-inCoord / sz2))) +
inCoord;
}
inCoord = inCoord < -len ? inCoord + sz2 : -inCoord - 1.0;
}
} else if (inCoord > len - 1.0) {
if (len <= 1.0) {
inCoord = 0.0;
} else {
float sz2 = 2.0 * len;
inCoord -= sz2 * float(int(float(inCoord / sz2)));
if (inCoord >= len) {
inCoord = sz2 - inCoord - 1.0;
}
}
}
return clamp(inCoord, 0.0, len - 1.0);
} else if (${fillModeId} == 3) {
if (inCoord < 0.0) {
if (len <= 1.0) {
inCoord = 0.0;
} else {
float sz = len - 1.0;
inCoord += len * (float(int(float(-inCoord / sz))) + 1.0);
}
} else if (inCoord > len - 1.0) {
if (len <= 1.0) {
inCoord = 0.0;
} else {
float sz = len - 1.0;
inCoord -= len * float(int(float(inCoord / sz)));
}
}
return clamp(inCoord, 0.0, len - 1.0);
} else if (${fillModeId} == 4) {
return clamp(outCoord, 0.0, len - 1.0);
} else {
return outCoord;
}
}
float readWithFillValue(int batch, int coordY, int coordX,
int channel) {
float outputValue;
if (0 <= coordY && coordY < ${imageHeight} && 0 <= coordX && coordX < ${imageWidth}) {
outputValue = getImage(batch, coordY, coordX, channel);
} else {
outputValue = float(${fillValue});
}
return outputValue;
}
void main() {
ivec4 coords = getOutputCoords();
float outputValue;
int batch = coords[0];
int x = coords[2];
int y = coords[1];
int channel = coords[3];
float xf = float(x);
float yf = float(y);
float a1 = getTransforms(batch, 0);
float a2 = getTransforms(batch, 1);
float a3 = getTransforms(batch, 2);
float b1 = getTransforms(batch, 3);
float b2 = getTransforms(batch, 4);
float b3 = getTransforms(batch, 5);
float c1 = getTransforms(batch, 6);
float c2 = getTransforms(batch, 7);
float projection = c1 * xf + c2 * yf + 1.0;
if (projection == 0.0) {
outputValue = float(${fillValue});
} else {
float inX = (a1 * xf + a2 * yf + a3) / projection;
float inY = (b1 * xf + b2 * yf + b3) / projection;
float mapX = mapCoord(inX, float(${imageWidth}));
float mapY = mapCoord(inY, float(${imageHeight}));
if (${interpolationModeId} == 1) {
int coordY = int(round(mapY));
int coordX = int(round(mapX));
outputValue = readWithFillValue(batch, coordY, coordX,
channel);
} else {
float yFloor = floor(mapY);
float xFloor = floor(mapX);
float yCeil = yFloor + 1.0;
float xCeil = xFloor + 1.0;
float valueYFloor = (xCeil - mapX) *
readWithFillValue(batch, int(yFloor), int(xFloor), channel) +
(mapX - xFloor) *
readWithFillValue(batch, int(yFloor), int(xCeil), channel);
float valueYCeil = (xCeil - mapX) *
readWithFillValue(batch, int(yCeil), int(xFloor), channel) +
(mapX - xFloor) *
readWithFillValue(batch, int(yCeil), int(xCeil), channel);
outputValue = (yCeil - mapY) * valueYFloor +
(mapY - yFloor) * valueYCeil;
}
}
setOutput(outputValue);
}
`;
}
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function transform$1(args) {
const { inputs, backend, attrs } = args;
const { image, transforms } = inputs;
const { interpolation, fillMode, fillValue, outputShape } = attrs;
const [batch, imageHeight, imageWidth, numChannels] = image.shape;
const [outHeight, outWidth] = outputShape != null ? outputShape : [imageHeight, imageWidth];
const outShape = [batch, outHeight, outWidth,
numChannels];
const program = new TransformProgram(imageHeight, imageWidth, interpolation, fillMode, fillValue, outShape);
return backend.runWebGLProgram(program, [image, transforms], 'float32');
}
const transformConfig$1 = {
kernelName: Transform,
backendName: 'webgl',
kernelFunc: transform$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function unique$2(args) {
const { inputs, attrs, backend } = args;
const { axis } = attrs;
const { x } = inputs;
assertNotComplex$1(x, 'unique');
// For now, always forward calculation to the CPU backend.
console.warn('WARNING: ', 'UI might be locked temporarily as data is being downloaded');
const values = backend.readSync(x.dataId);
const { outputValues, outputShape, indices } = uniqueImplCPU(values, axis, x.shape, x.dtype);
return [
backend.makeTensorInfo(outputShape, x.dtype, outputValues),
backend.makeTensorInfo([indices.length], 'int32', indices),
];
}
const uniqueConfig$1 = {
kernelName: Unique,
backendName: 'webgl',
kernelFunc: unique$2,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function unpack$1(args) {
const { inputs, backend, attrs } = args;
const { value } = inputs;
let { axis } = attrs;
if (axis < 0) {
axis += value.shape.length;
}
const x = value;
const xRank = x.shape.length;
const num = value.shape[axis];
const outShape = new Array(xRank - 1);
let outIndex = 0;
for (let i = 0; i < xRank; i++) {
if (i !== axis) {
outShape[outIndex++] = x.shape[i];
}
}
const toDispose = [];
const begin = new Array(xRank).fill(0);
const size = x.shape.slice();
size[axis] = 1;
const res = new Array(num);
for (let i = 0; i < res.length; i++) {
begin[axis] = i;
const sliced = slice({ inputs: { x }, backend, attrs: { begin, size } });
const reshaped = reshape$1({ inputs: { x: sliced }, backend, attrs: { shape: outShape } });
res[i] = reshaped;
toDispose.push(sliced);
}
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return res;
}
const unpackConfig$1 = {
kernelName: Unpack,
backendName: 'webgl',
kernelFunc: unpack$1
};
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
class SegmentOpProgram {
constructor(segOpInfo, segOpType) {
this.variableNames = ['x', 'segmentIds'];
const windowSize = segOpInfo.windowSize;
const batchSize = segOpInfo.batchSize;
const inSize = segOpInfo.inSize;
const numSegments = segOpInfo.numSegments;
const outSize = numSegments * Math.ceil(inSize / windowSize);
this.outputShape = [batchSize, outSize];
const initializationValue = '0.0';
const returnValue = `sumValue`;
const windowSizeNearestVec4 = Math.floor(windowSize / 4) * 4;
const windowSizeVec4Remainder = windowSize % 4;
const updateSnippet = `
sumValue += dot(values, segFilter);
`;
let checkValueOutOfBounds = '';
if (inSize % windowSize > 0) {
checkValueOutOfBounds = `
if (inIdx < 0 || inIdx >= ${inSize}) {
return initializationValue;
}
`;
}
let checkSegmentIdOutOfBounds = '';
if (inSize % windowSize > 0) {
checkSegmentIdOutOfBounds = `
if (inIdx < 0 || inIdx >= ${inSize}) {
return -1.0;
}
`;
}
this.userCode = `
const float initializationValue = ${initializationValue};
float getValue(int batch, int inIdx) {
${checkValueOutOfBounds}
return getX(batch, inIdx);
}
float getSegmentIdAtIndex(int inIdx) {
${checkSegmentIdOutOfBounds}
return getSegmentIds(inIdx);
}
void main() {
ivec2 coords = getOutputCoords();
int batch = coords[0];
int outIdx = coords[1];
int inOffset = int(floor(float(outIdx) / float(
${numSegments})) * float(${windowSize}));
int currentSeg = int(mod(float(outIdx), float(${numSegments})));
float sumValue = 0.0;
for (int i = 0; i < ${windowSizeNearestVec4}; i += 4) {
int inIdx = inOffset + i;
vec4 values = vec4(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
getValue(batch, inIdx + 2),
getValue(batch, inIdx + 3)
);
vec4 segFilter = vec4(
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
int(getSegmentIdAtIndex(inIdx + 3)) == currentSeg ? 1 : 0
);
${updateSnippet}
}
int inIdx = inOffset + ${windowSizeNearestVec4};
if (${windowSizeVec4Remainder === 1}) {
vec4 values = vec4(
getValue(batch, inIdx),
initializationValue,
initializationValue,
initializationValue
);
int inIdxSeg = int(getSegmentIdAtIndex(inIdx));
vec4 segFilter = vec4(
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
0,
0,
0
);
${updateSnippet}
} else if (${windowSizeVec4Remainder === 2}) {
vec4 values = vec4(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
initializationValue,
initializationValue
);
vec4 segFilter = vec4(
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
0,
0
);
${updateSnippet}
} else if (${windowSizeVec4Remainder === 3}) {
vec4 values = vec4(
getValue(batch, inIdx),
getValue(batch, inIdx + 1),
getValue(batch, inIdx + 2),
initializationValue
);
vec4 segFilter = vec4(
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
0
);
${updateSnippet}
}
setOutput(${returnValue});
}
`;
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function unsortedSegmentSum$1(args) {
const { inputs, backend, attrs } = args;
const { x, segmentIds } = inputs;
const { numSegments } = attrs;
const xRank = x.shape.length;
const toDispose = [];
let axis = 0;
const permutation = getAxesPermutation([axis], xRank);
let permutedX = x;
if (permutation != null) {
permutedX = transpose({ inputs: { x }, backend, attrs: { perm: permutation } });
toDispose.push(permutedX);
axis = getInnerMostAxes(1, xRank)[0];
}
const outShape = computeOutShape(permutedX.shape, axis, numSegments);
const inSize = sizeFromShape([permutedX.shape[axis]]);
const a2D = reshape$1({ inputs: { x: permutedX }, backend, attrs: { shape: [-1, inSize] } });
toDispose.push(a2D);
const outputDType = sumOutType(x.dtype);
const segOpCompute = (x, segOpType, segmentIds, dtype, numSegments) => {
const batchSize = x.shape[0];
const inSize = x.shape[1];
const windowSize = segOpComputeOptimalWindowSize(inSize, numSegments);
const segOpInfo = { windowSize, inSize, batchSize, numSegments };
const program = new SegmentOpProgram(segOpInfo, segOpType);
const output = backend.compileAndRun(program, [x, segmentIds], dtype);
toDispose.push(output);
// No need to run another GPGPU program.
if (output.shape[1] === numSegments) {
return output;
}
const rangeInfo = range$2({
backend,
attrs: { start: 0, stop: numSegments, step: 1, dtype: 'float32' }
});
const tileInfo = tile$2({
inputs: { x: rangeInfo },
backend,
attrs: { reps: [inSize / windowSize] }
});
toDispose.push(rangeInfo);
toDispose.push(tileInfo);
const result = segOpCompute(output, segOpType, tileInfo, dtype, numSegments);
return result;
};
const segOpResult = segOpCompute(a2D, 'unsortedSegmentSum', segmentIds, outputDType, numSegments);
const reshaped = reshape$1({ inputs: { x: segOpResult }, backend, attrs: { shape: outShape } });
let result = reshaped;
if (permutation != null) {
toDispose.push(reshaped);
const perm = getUndoAxesPermutation(permutation);
result = transpose({ inputs: { x: result }, backend, attrs: { perm } });
}
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return result;
}
const unsortedSegmentSumConfig$1 = {
kernelName: UnsortedSegmentSum,
backendName: 'webgl',
kernelFunc: unsortedSegmentSum$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// List all kernel configs here
const kernelConfigs$1 = [
_fusedMatMulConfig$1,
absConfig,
acosConfig$1,
acoshConfig$1,
addConfig,
addNConfig$1,
allConfig$1,
anyConfig$1,
argMaxConfig$1,
argMinConfig$1,
asinConfig$1,
asinhConfig$1,
atanConfig$1,
atan2Config$1,
atanhConfig$1,
avgPoolConfig$1,
avgPool3DConfig$1,
avgPool3DGradConfig$2,
avgPoolGradConfig$2,
batchMatMulConfig$1,
batchNormConfig$1,
batchToSpaceNDConfig$1,
bincountConfig$1,
bitwiseAndConfig,
broadcastArgsConfig$1,
castConfig,
ceilConfig,
clipByValueConfig$1,
complexConfig,
complexAbsConfig$1,
concatConfig$1,
conv2DConfig$1,
conv2DBackpropFilterConfig$1,
conv2DBackpropInputConfig$1,
conv3DConfig$1,
conv3DBackpropFilterV2Config$1,
conv3DBackpropInputConfig,
cosConfig$1,
coshConfig$1,
cropAndResizeConfig$1,
cumprodConfig$1,
cumsumConfig$1,
denseBincountConfig$1,
depthToSpaceConfig$1,
depthwiseConv2dNativeConfig$1,
depthwiseConv2dNativeBackpropFilterConfig$1,
depthwiseConv2dNativeBackpropInputConfig$1,
diagConfig$1,
dilation2DConfig$1,
einsumConfig$1,
eluConfig$1,
eluGradConfig$2,
equalConfig,
erfConfig$1,
expConfig,
expandDimsConfig$1,
expm1Config,
fftConfig$1,
fillConfig$1,
flipLeftRightConfig$1,
floorConfig,
floorDivConfig,
fromPixelsConfig,
fusedConv2DConfig$1,
fusedDepthwiseConv2DConfig$1,
gatherNdConfig$1,
gatherV2Config$1,
greaterConfig,
greaterEqualConfig,
identityConfig,
ifftConfig$1,
imagConfig$1,
isFiniteConfig$1,
isInfConfig$1,
isNaNConfig$1,
leakyReluConfig$1,
lessConfig,
lessEqualConfig,
linSpaceConfig$1,
logConfig,
log1pConfig$1,
logicalAndConfig$1,
logicalNotConfig$1,
logicalOrConfig$1,
LRNConfig$1,
LRNGradConfig$1,
maxConfig$1,
maximumConfig,
maxPoolConfig$1,
maxPool3DConfig$1,
maxPool3DGradConfig$2,
maxPoolGradConfig$2,
maxPoolWithArgmaxConfig$1,
meanConfig$1,
minConfig$1,
minimumConfig,
mirrorPadConfig$1,
modConfig$1,
multinomialConfig$1,
multiplyConfig,
negConfig,
nonMaxSuppressionV3Config$1,
nonMaxSuppressionV4Config$1,
nonMaxSuppressionV5Config$1,
notEqualConfig,
oneHotConfig$1,
onesLikeConfig$1,
packConfig$1,
padV2Config$1,
powConfig$1,
preluConfig$1,
prodConfig,
raggedGatherConfig$1,
raggedRangeConfig$1,
raggedTensorToTensorConfig$1,
rangeConfig$1,
realConfig,
realDivConfig$1,
reciprocalConfig$1,
reluConfig$1,
relu6Config$1,
reshapeConfig$1,
resizeBilinearConfig$1,
resizeBilinearGradConfig$2,
resizeNearestNeighborConfig$1,
resizeNearestNeighborGradConfig$2,
reverseConfig$1,
rotateWithOffsetConfig$1,
roundConfig$1,
rsqrtConfig,
scatterNdConfig$1,
searchSortedConfig$1,
selectConfig$1,
seluConfig$1,
sigmoidConfig,
signConfig$1,
sinConfig$1,
sinhConfig$1,
sliceConfig,
softmaxConfig$1,
softplusConfig$1,
spaceToBatchNDConfig$1,
sparseFillEmptyRowsConfig$1,
sparseReshapeConfig$1,
sparseSegmentMeanConfig$1,
sparseSegmentSumConfig$1,
sparseToDenseConfig$1,
splitVConfig$1,
sqrtConfig,
squareConfig$1,
squaredDifferenceConfig,
staticRegexReplaceConfig,
stepConfig$1,
stridedSliceConfig$1,
stringNGramsConfig$1,
stringSplitConfig$1,
stringToHashBucketFastConfig$1,
subConfig,
sumConfig$1,
tanConfig$1,
tanhConfig$1,
tensorScatterUpdateConfig$1,
tileConfig$1,
topKConfig$1,
transformConfig$1,
transposeConfig,
uniqueConfig$1,
unpackConfig$1,
unsortedSegmentSumConfig$1,
zerosLikeConfig$1
];
for (const kernelConfig of kernelConfigs$1) {
registerKernel(kernelConfig);
}
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const whereImpl = whereImpl$2;
class MathBackendCPU extends KernelBackend {
nextDataId() {
return MathBackendCPU.nextDataId++;
}
constructor() {
super();
this.blockSize = 48;
this.firstUse = true;
this.data = new DataStorage(this, engine());
}
write(values, shape, dtype) {
if (this.firstUse) {
this.firstUse = false;
if (env().get('IS_NODE')) {
warn('\n============================\n' +
'Hi, looks like you are running TensorFlow.js in ' +
'Node.js. To speed things up dramatically, install our node ' +
'backend, visit https://github.com/tensorflow/tfjs-node for more details. ' +
'\n============================');
}
}
const dataId = { id: this.nextDataId() };
this.data.set(dataId, { values, dtype, refCount: 1 });
return dataId;
}
/**
* Create a data bucket in cpu backend.
* @param shape Shape of the `TensorInfo`.
* @param dtype DType of the `TensorInfo`.
* @param values The value of the `TensorInfo` stored as a flattened array.
*/
makeTensorInfo(shape, dtype, values) {
let outId;
if (dtype === 'string' && values != null && values.length > 0 &&
isString(values[0])) {
const encodedValues = values.map(d => encodeString(d));
outId = this.write(encodedValues, shape, dtype);
}
else {
outId = this.write(values, shape, dtype);
}
return { dataId: outId, shape, dtype };
}
/** Return refCount of a `TensorData`. */
refCount(dataId) {
if (this.data.has(dataId)) {
const tensorData = this.data.get(dataId);
return tensorData.refCount;
}
return 0;
}
/** Increase refCount of a `TensorData`. */
incRef(dataId) {
const tensorData = this.data.get(dataId);
tensorData.refCount++;
}
/** Decrease refCount of a `TensorData`. */
decRef(dataId) {
if (this.data.has(dataId)) {
const tensorData = this.data.get(dataId);
tensorData.refCount--;
}
}
move(dataId, values, shape, dtype, refCount) {
this.data.set(dataId, { values, dtype, refCount });
}
numDataIds() {
return this.data.numDataIds();
}
async read(dataId) {
return this.readSync(dataId);
}
readSync(dataId) {
const { dtype, complexTensorInfos } = this.data.get(dataId);
if (dtype === 'complex64') {
const realValues = this.readSync(complexTensorInfos.real.dataId);
const imagValues = this.readSync(complexTensorInfos.imag.dataId);
return mergeRealAndImagArrays(realValues, imagValues);
}
return convertBackendValuesAndArrayBuffer(this.data.get(dataId).values, dtype);
}
bufferSync(t) {
const data = this.readSync(t.dataId);
if (t.dtype === 'string') {
try {
// Decode the bytes into string.
const strings = data.map(d => decodeString(d));
return buffer(t.shape, t.dtype, strings);
}
catch (_a) {
throw new Error('Failed to decode encoded string bytes into utf-8');
}
}
return buffer(t.shape, t.dtype, data);
}
makeOutput(values, shape, dtype) {
return engine().makeTensorFromTensorInfo(this.makeTensorInfo(shape, dtype, values), this);
}
/**
* Dispose the memory if the dataId has 0 refCount. Return true if the memory
* is released or memory is not managed in this backend, false if memory is
* not cleared.
* @param dataId
* @oaram force Optional, remove the data regardless of refCount
*/
disposeData(dataId, force = false) {
if (this.data.has(dataId)) {
this.data.get(dataId).refCount--;
if (!force && this.data.get(dataId).refCount > 0) {
return false;
}
const { complexTensorInfos } = this.data.get(dataId);
if (complexTensorInfos != null) {
this.disposeData(complexTensorInfos.real.dataId, true);
this.disposeData(complexTensorInfos.imag.dataId, true);
}
this.data.delete(dataId);
}
return true;
}
disposeIntermediateTensorInfo(tensorInfo) {
this.disposeData(tensorInfo.dataId);
}
async time(f) {
const start = now();
f();
const kernelMs = now() - start;
return { kernelMs };
}
memory() {
return {
// Unreliable due to automatic gc. The numbers above are cumulative.
unreliable: true,
reasons: ['The reported memory is an upper bound. Due to automatic garbage ' +
'collection, the true allocated memory may be less.']
};
}
where(condition) {
assertNotComplex([condition], 'where');
const condVals = this.readSync(condition.dataId);
return whereImpl(condition.shape, condVals);
}
dispose() { }
floatPrecision() {
return 32;
}
/** Returns the smallest representable number. */
epsilon() {
return super.epsilon();
}
}
MathBackendCPU.nextDataId = 0;
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/*
* base.ts contains all the exports from tfjs-backend-cpu
* without auto-kernel registration
*/
// Side effects for default initialization of MathBackendCPU
registerBackend('cpu', () => new MathBackendCPU(), 1 /* priority */);
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const elu$1 = unaryKernelFunc$1(Elu$1, (xi) => xi >= 0 ? xi : (Math.exp(xi) - 1));
const eluConfig = {
kernelName: Elu$1,
backendName: 'cpu',
kernelFunc: elu$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function leakyRelu(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { alpha } = attrs;
assertNotComplex([x], 'leakyRelu');
const xSize = sizeFromShape(x.shape);
const xVals = backend.data.get(x.dataId).values;
const outVals = getTypedArrayFromDType('float32', xSize);
for (let i = 0; i < xVals.length; i++) {
outVals[i] = xVals[i] < 0 ? alpha * xVals[i] : xVals[i];
}
return backend.makeTensorInfo(x.shape, 'float32', outVals);
}
const leakyReluConfig = {
kernelName: LeakyRelu,
backendName: 'cpu',
kernelFunc: leakyRelu
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const preluImpl = createSimpleBinaryKernelImpl((xValue, aValue) => xValue < 0 ? aValue * xValue : xValue);
function prelu(args) {
const { inputs, backend } = args;
const { x, alpha } = inputs;
assertNotComplex([x, alpha], 'prelu');
const aVals = backend.data.get(x.dataId).values;
const bVals = backend.data.get(alpha.dataId).values;
const [resultData, resultShape] = preluImpl(x.shape, alpha.shape, aVals, bVals, 'float32');
return backend.makeTensorInfo(resultShape, 'float32', resultData);
}
const preluConfig = {
kernelName: Prelu,
backendName: 'cpu',
kernelFunc: prelu,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const relu = unaryKernelFunc$1(Relu$1, (xi) => Math.max(0, xi));
const reluConfig = {
kernelName: Relu$1,
backendName: 'cpu',
kernelFunc: relu,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const relu6 = unaryKernelFunc$1(Relu6$1, (xi) => Math.min(Math.max(0, xi), 6));
const relu6Config = {
kernelName: Relu6$1,
backendName: 'cpu',
kernelFunc: relu6,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function applyActivation(backend, x, activation, preluActivationWeights, leakyreluAlpha) {
if (activation === 'linear') {
return identity$1({ inputs: { x }, backend });
}
else if (activation === 'relu') {
return relu({ inputs: { x }, backend });
}
else if (activation === 'elu') {
return elu$1({ inputs: { x }, backend });
}
else if (activation === 'relu6') {
return relu6({ inputs: { x }, backend });
}
else if (activation === 'prelu') {
return prelu({ inputs: { x, alpha: preluActivationWeights }, backend });
}
else if (activation === 'leakyrelu') {
return leakyRelu({ inputs: { x }, backend, attrs: { alpha: leakyreluAlpha } });
}
else if (activation === 'sigmoid') {
return sigmoid$1({ inputs: { x }, backend });
}
throw new Error(`Activation ${activation} has not been implemented for the CPU backend.`);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function reshape(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { shape } = attrs;
const xSize = sizeFromShape(x.shape);
const $shape = inferFromImplicitShape(shape, xSize);
const $xSize = sizeFromShape($shape);
assert$1(xSize === $xSize, () => `The new shape (${$shape}) has ${$xSize} elements and the old ` +
`shape (${x.shape}) has ${xSize} elements. The new shape and old ` +
`shape must have the same number of elements.`);
backend.incRef(x.dataId);
const xData = backend.data.get(x.dataId);
if (xData.complexTensorInfos != null) {
const real = xData.complexTensorInfos.real;
const imag = xData.complexTensorInfos.imag;
real.shape = $shape;
imag.shape = $shape;
}
return { dataId: x.dataId, shape: $shape, dtype: x.dtype };
}
const reshapeConfig = {
kernelName: Reshape$1,
backendName: 'cpu',
kernelFunc: reshape
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function batchMatMul(args) {
const { inputs, backend, attrs } = args;
const { a, b } = inputs;
const { transposeA, transposeB } = attrs;
assertNotComplex([a, b], 'matMul');
const aRank = a.shape.length;
const bRank = b.shape.length;
const innerShapeA = transposeA ? a.shape[aRank - 2] : a.shape[aRank - 1];
const innerShapeB = transposeB ? b.shape[bRank - 1] : b.shape[bRank - 2];
const outerShapeA = transposeA ? a.shape[aRank - 1] : a.shape[aRank - 2];
const outerShapeB = transposeB ? b.shape[bRank - 2] : b.shape[bRank - 1];
const outerDimsA = a.shape.slice(0, -2);
const outerDimsB = b.shape.slice(0, -2);
const batchDimA = sizeFromShape(outerDimsA);
const batchDimB = sizeFromShape(outerDimsB);
const outShapeOuterDims = assertAndGetBroadcastShape(a.shape.slice(0, -2), b.shape.slice(0, -2));
const outShape = outShapeOuterDims.concat([outerShapeA, outerShapeB]);
assert$1(innerShapeA === innerShapeB, () => `Error in matMul: inner shapes (${innerShapeA}) and (` +
`${innerShapeB}) of Tensors with shapes ${a.shape} and ` +
`${b.shape} and transposeA=${transposeA}` +
` and transposeB=${transposeB} must match.`);
const a3dShape = transposeA ? [batchDimA, innerShapeA, outerShapeA] :
[batchDimA, outerShapeA, innerShapeA];
const b3dShape = transposeB ? [batchDimB, outerShapeB, innerShapeB] :
[batchDimB, innerShapeB, outerShapeB];
// The rest of the implementation is designed to operate on rank-3 tensors
const a3d = reshape({ inputs: { x: a }, backend, attrs: { shape: a3dShape } });
const b3d = reshape({ inputs: { x: b }, backend, attrs: { shape: b3dShape } });
const sharedDim = transposeA ? a3d.shape[1] : a3d.shape[2];
const leftDim = transposeA ? a3d.shape[2] : a3d.shape[1];
const rightDim = transposeB ? b3d.shape[1] : b3d.shape[2];
const batchDim = Math.max(batchDimA, batchDimB);
const a3dValues = backend.data.get(a3d.dataId).values;
const b3dValues = backend.data.get(b3d.dataId).values;
const a3dStrides = computeStrides(a3d.shape);
const b3dStrides = computeStrides(b3d.shape);
const [aBatch, aOuterStep, aInnerStep] = transposeA ?
[a3dStrides[0], 1, a3dStrides[1]] :
[a3dStrides[0], a3dStrides[1], 1];
const [bInnerStep, bOuterStep, bBatch] = transposeB ?
[1, b3dStrides[1], b3dStrides[0]] :
[b3dStrides[1], 1, b3dStrides[0]];
const size = leftDim * rightDim;
const result = buffer([batchDim, leftDim, rightDim], a3d.dtype);
const resVals = result.values;
const blockSize = backend.blockSize;
for (let bi = 0; bi < batchDim; bi++) {
const batchIndexA = bi % batchDimA;
const batchIndexB = bi % batchDimB;
for (let i0 = 0; i0 < leftDim; i0 += blockSize) {
// for when blockSize doesn't evenly divide the input
const iBlock = Math.min(i0 + blockSize, leftDim);
for (let j0 = 0; j0 < rightDim; j0 += blockSize) {
const jBlock = Math.min(j0 + blockSize, rightDim);
for (let k0 = 0; k0 < sharedDim; k0 += blockSize) {
const kBlock = Math.min(k0 + blockSize, sharedDim);
for (let i = i0; i < iBlock; i++) {
for (let j = j0; j < jBlock; j++) {
let sum = 0.0;
for (let k = k0; k < kBlock; k++) {
const aVal =
// tslint:disable-next-line: max-line-length
a3dValues[batchIndexA * aBatch + i * aOuterStep + k * aInnerStep];
const bVal =
// tslint:disable-next-line: max-line-length
b3dValues[k * bInnerStep + j * bOuterStep + batchIndexB * bBatch];
sum += aVal * bVal;
}
resVals[bi * size + (i * rightDim + j)] += sum;
}
}
}
}
}
}
backend.disposeIntermediateTensorInfo(a3d);
backend.disposeIntermediateTensorInfo(b3d);
// set correct shape on output.
return backend.makeTensorInfo(outShape, result.dtype, result.values);
}
const batchMatMulConfig = {
kernelName: BatchMatMul,
backendName: 'cpu',
kernelFunc: batchMatMul,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function _fusedMatMul(args) {
const { inputs, backend, attrs } = args;
const { a, b, bias, preluActivationWeights } = inputs;
const { transposeA, transposeB, activation, leakyreluAlpha } = attrs;
let current;
let addRes;
let activationRes;
const intermediates = [];
const matMulRes = batchMatMul({ inputs: { a, b }, attrs: { transposeA, transposeB }, backend });
current = matMulRes;
if (bias) {
addRes = add({ inputs: { a: current, b: bias }, backend });
intermediates.push(current);
current = addRes;
}
if (activation) {
activationRes = applyActivation(backend, current, activation, preluActivationWeights, leakyreluAlpha);
intermediates.push(current);
current = activationRes;
}
for (const i of intermediates) {
backend.disposeIntermediateTensorInfo(i);
}
return current;
}
const _fusedMatMulConfig = {
kernelName: _FusedMatMul,
backendName: 'cpu',
kernelFunc: _fusedMatMul,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const acos = unaryKernelFunc$1(Acos, (xi) => Math.acos(xi));
const acosConfig = {
kernelName: Acos,
backendName: 'cpu',
kernelFunc: acos,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const acosh = unaryKernelFunc$1(Acosh, (xi) => Math.acosh(xi));
const acoshConfig = {
kernelName: Acosh,
backendName: 'cpu',
kernelFunc: acosh,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function addN(args) {
const { inputs, backend } = args;
const tensors = inputs;
assertNotComplex(inputs, 'addN');
const vals = tensors.map(t => backend.data.get(t.dataId).values);
const outBuf = buffer(tensors[0].shape, tensors[0].dtype);
const outVals = outBuf.values;
for (let i = 0; i < tensors.length; i++) {
const currVals = vals[i];
for (let j = 0; j < outVals.length; j++) {
outVals[j] += currVals[j];
}
}
return backend.makeTensorInfo(outBuf.shape, outBuf.dtype, outBuf.values);
}
const addNConfig = {
kernelName: AddN,
backendName: 'cpu',
kernelFunc: addN
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function all(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
assertNotComplex(x, 'all');
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, x.shape.length);
let $x = x;
if (permutedAxes != null) {
$x = transpose$1({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
axes = getInnerMostAxes(axes.length, x.shape.length);
}
assertAxesAreInnerMostDims('all', axes, $x.shape.length);
const [outShape, reduceShape] = computeOutAndReduceShapes($x.shape, axes);
const reduceSize = sizeFromShape(reduceShape);
const vals = makeZerosTypedArray(sizeFromShape(outShape), $x.dtype);
const aVals = backend.data.get($x.dataId).values;
for (let i = 0; i < vals.length; ++i) {
const offset = i * reduceSize;
let all = aVals[offset];
for (let j = 0; j < reduceSize; ++j) {
const value = aVals[offset + j];
all = all && value;
}
vals[i] = all;
}
if (permutedAxes != null) {
backend.disposeIntermediateTensorInfo($x);
}
const result = backend.makeTensorInfo(outShape, $x.dtype, vals);
if (keepDims) {
const expandedShape = expandShapeToKeepDim(outShape, origAxes);
const reshapedResult = reshape({ inputs: { x: result }, backend, attrs: { shape: expandedShape } });
backend.disposeIntermediateTensorInfo(result);
return reshapedResult;
}
return result;
}
const allConfig = {
kernelName: All,
backendName: 'cpu',
kernelFunc: all
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function any(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
assertNotComplex(x, 'any');
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, x.shape.length);
let $x = x;
if (permutedAxes != null) {
$x = transpose$1({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
axes = getInnerMostAxes(axes.length, x.shape.length);
}
assertAxesAreInnerMostDims('any', axes, $x.shape.length);
const [outShape, reduceShape] = computeOutAndReduceShapes($x.shape, axes);
const reduceSize = sizeFromShape(reduceShape);
const vals = makeZerosTypedArray(sizeFromShape(outShape), $x.dtype);
const aVals = backend.data.get($x.dataId).values;
for (let i = 0; i < vals.length; ++i) {
const offset = i * reduceSize;
let anyVal = aVals[offset];
for (let j = 0; j < reduceSize; ++j) {
const value = aVals[offset + j];
anyVal = anyVal || value;
}
vals[i] = anyVal;
}
if (permutedAxes != null) {
backend.disposeIntermediateTensorInfo($x);
}
const result = backend.makeTensorInfo(outShape, $x.dtype, vals);
if (keepDims) {
const expandedShape = expandShapeToKeepDim(outShape, origAxes);
const reshapedResult = reshape({ inputs: { x: result }, backend, attrs: { shape: expandedShape } });
backend.disposeIntermediateTensorInfo(result);
return reshapedResult;
}
return result;
}
const anyConfig = {
kernelName: Any,
backendName: 'cpu',
kernelFunc: any
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function argMax(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis } = attrs;
assertNotComplex(x, 'argMax');
let axes = parseAxisParam(axis, x.shape);
const permutedAxes = getAxesPermutation(axes, x.shape.length);
let $x = x;
const intermediateTensorInfos = [];
if (permutedAxes != null) {
$x = transpose$1({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
intermediateTensorInfos.push($x);
axes = getInnerMostAxes(axes.length, $x.shape.length);
}
axes = [axes[0]];
assertAxesAreInnerMostDims('argMax', axes, $x.shape.length);
const [outShape, reduceShape] = computeOutAndReduceShapes($x.shape, axes);
const outSize = sizeFromShape(outShape);
const vals = makeZerosTypedArray(outSize, 'int32');
const reduceSize = sizeFromShape(reduceShape);
const aVals = backend.data.get($x.dataId).values;
for (let i = 0; i < vals.length; ++i) {
const offset = i * reduceSize;
let max = aVals[offset];
let maxIndex = 0;
for (let j = 0; j < reduceSize; ++j) {
const value = aVals[offset + j];
if (value > max) {
max = value;
maxIndex = j;
}
}
vals[i] = maxIndex;
}
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return backend.makeTensorInfo(outShape, 'int32', vals);
}
const argMaxConfig = {
kernelName: ArgMax,
backendName: 'cpu',
kernelFunc: argMax
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function argMin(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis } = attrs;
assertNotComplex(x, 'argMin');
let axes = parseAxisParam(axis, x.shape);
const permutedAxes = getAxesPermutation(axes, x.shape.length);
let $x = x;
const intermediateTensorInfos = [];
if (permutedAxes != null) {
$x = transpose$1({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
intermediateTensorInfos.push($x);
axes = getInnerMostAxes(axes.length, $x.shape.length);
}
axes = [axes[0]];
assertAxesAreInnerMostDims('argMin', axes, $x.shape.length);
const [outShape, reduceShape] = computeOutAndReduceShapes($x.shape, axes);
const outSize = sizeFromShape(outShape);
const vals = makeZerosTypedArray(outSize, 'int32');
const reduceSize = sizeFromShape(reduceShape);
const aVals = backend.data.get($x.dataId).values;
for (let i = 0; i < vals.length; ++i) {
const offset = i * reduceSize;
let min = aVals[offset];
let minIndex = 0;
for (let j = 0; j < reduceSize; ++j) {
const value = aVals[offset + j];
if (value < min) {
min = value;
minIndex = j;
}
}
vals[i] = minIndex;
}
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return backend.makeTensorInfo(outShape, 'int32', vals);
}
const argMinConfig = {
kernelName: ArgMin,
backendName: 'cpu',
kernelFunc: argMin
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const asin = unaryKernelFunc$1(Asin, (xi) => Math.asin(xi));
const asinConfig = {
kernelName: Asin,
backendName: 'cpu',
kernelFunc: asin,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const asinh = unaryKernelFunc$1(Asinh, (xi) => Math.asinh(xi));
const asinhConfig = {
kernelName: Asinh,
backendName: 'cpu',
kernelFunc: asinh,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const atan = unaryKernelFunc$1(Atan, (xi) => Math.atan(xi));
const atanConfig = {
kernelName: Atan,
backendName: 'cpu',
kernelFunc: atan,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const atan2Impl = createSimpleBinaryKernelImpl((aValue, bValue) => Math.atan2(aValue, bValue));
const atan2 = binaryKernelFunc$1(Atan2, atan2Impl);
const atan2Config = {
kernelName: Atan2,
backendName: 'cpu',
kernelFunc: atan2,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const atanh = unaryKernelFunc$1(Atanh, (xi) => Math.atanh(xi));
const atanhConfig = {
kernelName: Atanh,
backendName: 'cpu',
kernelFunc: atanh,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function pool(xValues, xShape, dtype, strides, convInfo, poolType) {
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
const initialValue = (poolType === 'max' ? Number.NEGATIVE_INFINITY :
Number.POSITIVE_INFINITY);
const output = buffer(convInfo.outShape, dtype);
const outputVals = output.values;
const outputBatchStrides = convInfo.outShape[1] * convInfo.outShape[2] * convInfo.outShape[3];
const outputRowStrides = convInfo.outShape[2] * convInfo.outShape[3];
const outputColStrides = convInfo.outShape[3];
for (let b = 0; b < convInfo.batchSize; ++b) {
const outputBatchOffset = b * outputBatchStrides;
const inputBatchOffset = b * strides[0];
for (let d = 0; d < convInfo.inChannels; ++d) {
for (let yR = 0; yR < convInfo.outHeight; ++yR) {
const xRCorner = yR * strideHeight - padTop;
const xRMin = Math.max(0, xRCorner);
const xRMax = Math.min(convInfo.inHeight, effectiveFilterHeight + xRCorner);
const outputRowOffset = outputBatchOffset + yR * outputRowStrides;
for (let yC = 0; yC < convInfo.outWidth; ++yC) {
const xCCorner = yC * strideWidth - padLeft;
const xCMin = Math.max(0, xCCorner);
const xCMax = Math.min(convInfo.inWidth, effectiveFilterWidth + xCCorner);
let minMaxValue = initialValue;
let avgValue = 0;
let count = 0;
for (let xR = xRMin; xR < xRMax; xR += dilationHeight) {
const xROffset = inputBatchOffset + xR * strides[1];
for (let xC = xCMin; xC < xCMax; xC += dilationWidth) {
const xCOffset = xROffset + xC * strides[2];
const pixel = xValues[xCOffset + d];
if ((poolType === 'max' && pixel > minMaxValue)) {
minMaxValue = pixel;
}
else if (poolType === 'avg') {
avgValue += pixel;
count++;
}
}
if (isNaN(minMaxValue)) {
break;
}
}
const outputOffset = outputRowOffset + yC * outputColStrides + d;
outputVals[outputOffset] =
poolType === 'avg' ? avgValue / count : minMaxValue;
}
}
}
}
return output;
}
function maxPoolPositions(xValues, xShape, dtype, convInfo, flattenPositions = false, includeBatchInIndex = false) {
const maxPositions = buffer(convInfo.outShape, 'int32');
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
const xBuf = buffer(xShape, dtype, xValues);
for (let b = 0; b < convInfo.batchSize; ++b) {
for (let d = 0; d < convInfo.inChannels; ++d) {
for (let yR = 0; yR < convInfo.outHeight; ++yR) {
const xRCorner = yR * strideHeight - padTop;
let xRMin = xRCorner;
while (xRMin < 0) {
xRMin += dilationHeight;
}
// const xRMin = Math.max(0, xRCorner);
const xRMax = Math.min(convInfo.inHeight, effectiveFilterHeight + xRCorner);
for (let yC = 0; yC < convInfo.outWidth; ++yC) {
const xCCorner = yC * strideWidth - padLeft;
let xCMin = xCCorner;
while (xCMin < 0) {
xCMin += dilationWidth;
}
const xCMax = Math.min(convInfo.inWidth, effectiveFilterWidth + xCCorner);
let maxValue = Number.NEGATIVE_INFINITY;
let maxPosition = -1;
for (let xR = xRMin; xR < xRMax; xR += dilationHeight) {
const wR = xR - xRCorner;
for (let xC = xCMin; xC < xCMax; xC += dilationWidth) {
const wC = xC - xCCorner;
// For some reason, disable-next-line is not working
// TODO(mattsoulanille): Remove this when switching to TS5.
/* tslint:disable: no-unnecessary-type-assertion */
const pixel = xBuf.get(b, xR, xC, d);
if (pixel > maxValue) {
maxValue = pixel;
if (flattenPositions) {
maxPosition = includeBatchInIndex ?
((b * convInfo.inHeight + xR) * convInfo.inWidth + xC) *
convInfo.inChannels +
d :
(xR * convInfo.inWidth + xC) * convInfo.inChannels + d;
}
else {
maxPosition = wR * effectiveFilterWidth + wC;
}
}
}
}
maxPositions.set(maxPosition, b, yR, yC, d);
}
}
}
}
return maxPositions;
}
function pool3d(xValues, xShape, dtype, strides, convInfo, poolType) {
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterDepth = convInfo.effectiveFilterDepth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padFront = convInfo.padInfo.front;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
const initialValue = (poolType === 'max' ? Number.NEGATIVE_INFINITY :
Number.POSITIVE_INFINITY);
const output = buffer(convInfo.outShape, dtype);
const outputVals = output.values;
const outputBatchStrides = convInfo.outShape[1] * convInfo.outShape[2] *
convInfo.outShape[3] * convInfo.outShape[4];
const outputDepthStrides = convInfo.outShape[2] * convInfo.outShape[3] * convInfo.outShape[4];
const outputRowStrides = convInfo.outShape[3] * convInfo.outShape[4];
const outputColStrides = convInfo.outShape[4];
for (let batch = 0; batch < convInfo.batchSize; ++batch) {
const outputBatchOffset = batch * outputBatchStrides;
const inputBatchOffset = batch * strides[0];
for (let channel = 0; channel < convInfo.inChannels; ++channel) {
for (let yDepth = 0; yDepth < convInfo.outDepth; ++yDepth) {
const xDepthCorner = yDepth * strideDepth - padFront;
let xDepthMin = xDepthCorner;
while (xDepthMin < 0) {
xDepthMin += dilationDepth;
}
const xDepthMax = Math.min(convInfo.inDepth, effectiveFilterDepth + xDepthCorner);
const outputDepthOffset = outputBatchOffset + yDepth * outputDepthStrides;
for (let yRow = 0; yRow < convInfo.outHeight; ++yRow) {
const xRowCorner = yRow * strideHeight - padTop;
let xRowMin = xRowCorner;
while (xRowMin < 0) {
xRowMin += dilationHeight;
}
const xRowMax = Math.min(convInfo.inHeight, effectiveFilterHeight + xRowCorner);
const outputRowOffset = outputDepthOffset + yRow * outputRowStrides;
for (let yCol = 0; yCol < convInfo.outWidth; ++yCol) {
const xColCorner = yCol * strideWidth - padLeft;
let xColMin = xColCorner;
while (xColMin < 0) {
xColMin += dilationWidth;
}
const xColMax = Math.min(convInfo.inWidth, effectiveFilterWidth + xColCorner);
// Shader code begins
const outputColOffset = outputRowOffset + yCol * outputColStrides;
let minMaxValue = initialValue;
let avgValue = 0;
let count = 0;
for (let xDepth = xDepthMin; xDepth < xDepthMax; xDepth += dilationDepth) {
const xDepthOffset = inputBatchOffset + xDepth * strides[1];
for (let xRow = xRowMin; xRow < xRowMax; xRow += dilationHeight) {
const xRowOffset = xDepthOffset + xRow * strides[2];
for (let xCol = xColMin; xCol < xColMax; xCol += dilationWidth) {
const xColOffset = xRowOffset + xCol * strides[3];
const pixel = xValues[xColOffset + channel];
if ((poolType === 'max' && pixel > minMaxValue)) {
minMaxValue = pixel;
}
else if (poolType === 'avg') {
avgValue += pixel;
count++;
}
if (isNaN(minMaxValue)) {
break;
}
}
if (isNaN(minMaxValue)) {
break;
}
}
if (isNaN(minMaxValue)) {
break;
}
}
const outputOffset = outputColOffset + channel;
outputVals[outputOffset] = poolType === 'avg' ?
avgValue / Math.max(count, 1) :
minMaxValue;
}
}
}
}
}
return output;
}
function maxPool3dPositions(xBuf, convInfo) {
const maxPositions = buffer(convInfo.outShape, 'int32');
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterDepth = convInfo.effectiveFilterDepth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padFront = convInfo.padInfo.front;
const padTop = convInfo.padInfo.top;
const padLeft = convInfo.padInfo.left;
for (let batch = 0; batch < convInfo.batchSize; ++batch) {
for (let channel = 0; channel < convInfo.inChannels; ++channel) {
for (let yDepth = 0; yDepth < convInfo.outDepth; ++yDepth) {
const xDepthCorner = yDepth * strideDepth - padFront;
let xDepthMin = xDepthCorner;
while (xDepthMin < 0) {
xDepthMin += dilationDepth;
}
const xDepthMax = Math.min(convInfo.inDepth, effectiveFilterDepth + xDepthCorner);
for (let yRow = 0; yRow < convInfo.outHeight; ++yRow) {
const xRowCorner = yRow * strideHeight - padTop;
let xRowMin = xRowCorner;
while (xRowMin < 0) {
xRowMin += dilationHeight;
}
const xRowMax = Math.min(convInfo.inHeight, effectiveFilterHeight + xRowCorner);
for (let yCol = 0; yCol < convInfo.outWidth; ++yCol) {
const xColCorner = yCol * strideWidth - padLeft;
let xColMin = xColCorner;
while (xColMin < 0) {
xColMin += dilationWidth;
}
const xColMax = Math.min(convInfo.inWidth, effectiveFilterWidth + xColCorner);
// Shader code begins
let maxValue = Number.NEGATIVE_INFINITY;
let maxPosition = -1;
for (let xDepth = xDepthMin; xDepth < xDepthMax; xDepth += dilationDepth) {
const wDepth = xDepth - xDepthCorner;
for (let xRow = xRowMin; xRow < xRowMax; xRow += dilationHeight) {
const wRow = xRow - xRowCorner;
for (let xCol = xColMin; xCol < xColMax; xCol += dilationWidth) {
const wCol = xCol - xColCorner;
const pixel = xBuf.get(batch, xDepth, xRow, xCol, channel);
if (pixel >= maxValue) {
maxValue = pixel;
maxPosition =
wDepth * effectiveFilterHeight * effectiveFilterWidth +
wRow * effectiveFilterHeight + wCol;
}
}
}
}
maxPositions.set(maxPosition, batch, yDepth, yRow, yCol, channel);
}
}
}
}
}
return maxPositions;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPool(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
assertNotComplex(x, 'avgPool');
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const dilations = 1;
assert$1(eitherStridesOrDilationsAreOne(strides, dilations), () => 'Error in avgPool: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
const convInfo = computePool2DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode);
let res;
if (convInfo.filterWidth === 1 && convInfo.filterHeight === 1 &&
arraysEqual(convInfo.inShape, convInfo.outShape)) {
res = identity$1({ inputs: { x }, backend });
}
else {
const xValues = backend.data.get(x.dataId).values;
const strides = computeStrides(x.shape);
const buffer = pool(xValues, x.shape, x.dtype, strides, convInfo, 'avg');
res = backend.makeTensorInfo(convInfo.outShape, x.dtype, buffer.values);
}
return res;
}
const avgPoolConfig = {
kernelName: AvgPool,
backendName: 'cpu',
kernelFunc: avgPool
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPool3D(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { filterSize, strides, pad, dimRoundingMode, dataFormat } = attrs;
assertNotComplex(x, 'avgPool3d');
const convInfo = computePool3DInfo(x.shape, filterSize, strides, 1 /* dilations */, pad, dimRoundingMode, dataFormat);
const xValues = backend.data.get(x.dataId).values;
const outBuf = pool3d(xValues, x.shape, x.dtype, computeStrides(x.shape), convInfo, 'avg');
return backend.makeTensorInfo(outBuf.shape, 'float32', outBuf.values);
}
const avgPool3DConfig = {
kernelName: AvgPool3D,
backendName: 'cpu',
kernelFunc: avgPool3D
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPool3DGrad(args) {
const { inputs, backend, attrs } = args;
const { dy, input } = inputs;
const { filterSize, strides, pad, dimRoundingMode } = attrs;
assertNotComplex([dy, input], 'avgPool3DGrad');
const convInfo = computePool3DInfo(input.shape, filterSize, strides, 1 /* dilations */, pad, dimRoundingMode);
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const filterDepth = convInfo.filterDepth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterDepth = convInfo.effectiveFilterDepth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padFront = effectiveFilterDepth - 1 - convInfo.padInfo.front;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const dx = buffer(input.shape, 'float32');
const avgMultiplier = 1 / (filterDepth * filterHeight * filterWidth);
const dyBuf = backend.bufferSync(dy);
for (let batch = 0; batch < convInfo.batchSize; ++batch) {
for (let channel = 0; channel < convInfo.inChannels; ++channel) {
for (let dxDepth = 0; dxDepth < convInfo.inDepth; ++dxDepth) {
for (let dxRow = 0; dxRow < convInfo.inHeight; ++dxRow) {
for (let dxCol = 0; dxCol < convInfo.inWidth; ++dxCol) {
// Shader code begins.
const dyDepthCorner = dxDepth - padFront;
const dyRowCorner = dxRow - padTop;
const dyColCorner = dxCol - padLeft;
let dotProd = 0;
for (let wDepth = 0; wDepth < effectiveFilterDepth; wDepth += dilationDepth) {
const dyDepth = (dyDepthCorner + wDepth) / strideDepth;
if (dyDepth < 0 || dyDepth >= convInfo.outDepth ||
Math.floor(dyDepth) !== dyDepth) {
continue;
}
for (let wRow = 0; wRow < effectiveFilterHeight; wRow += dilationHeight) {
const dyRow = (dyRowCorner + wRow) / strideHeight;
if (dyRow < 0 || dyRow >= convInfo.outHeight ||
Math.floor(dyRow) !== dyRow) {
continue;
}
for (let wCol = 0; wCol < effectiveFilterWidth; wCol += dilationWidth) {
const dyCol = (dyColCorner + wCol) / strideWidth;
if (dyCol < 0 || dyCol >= convInfo.outWidth ||
Math.floor(dyCol) !== dyCol) {
continue;
}
const pixel = dyBuf.get(batch, dyDepth, dyRow, dyCol, channel);
dotProd += pixel;
}
}
}
dx.set(dotProd * avgMultiplier, batch, dxDepth, dxRow, dxCol, channel);
}
}
}
}
}
return backend.makeTensorInfo(dx.shape, dx.dtype, dx.values);
}
const avgPool3DGradConfig$1 = {
kernelName: AvgPool3DGrad,
backendName: 'cpu',
kernelFunc: avgPool3DGrad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function avgPoolGrad$1(args) {
const { inputs, backend, attrs } = args;
const { dy, input } = inputs;
const x = input;
assertNotComplex([dy, input], 'avgPoolGrad');
const { filterSize, strides, pad } = attrs;
const convInfo = computePool2DInfo(x.shape, filterSize, strides, 1 /* dilations */, pad);
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const dx = buffer(x.shape, 'float32');
const avgMultiplier = 1 / (filterHeight * filterWidth);
const dyData = backend.data.get(dy.dataId).values;
const dyBuf = buffer(dy.shape, 'float32', dyData);
for (let b = 0; b < convInfo.batchSize; ++b) {
for (let d = 0; d < convInfo.inChannels; ++d) {
for (let dxR = 0; dxR < convInfo.inHeight; ++dxR) {
for (let dxC = 0; dxC < convInfo.inWidth; ++dxC) {
// Shader code begins.
const dyRCorner = dxR - padTop;
const dyCCorner = dxC - padLeft;
let dotProd = 0;
for (let wR = 0; wR < effectiveFilterHeight; wR += dilationHeight) {
const dyR = (dyRCorner + wR) / strideHeight;
if (dyR < 0 || dyR >= convInfo.outHeight ||
Math.floor(dyR) !== dyR) {
continue;
}
for (let wC = 0; wC < effectiveFilterWidth; wC += dilationWidth) {
const dyC = (dyCCorner + wC) / strideWidth;
if (dyC < 0 || dyC >= convInfo.outWidth ||
Math.floor(dyC) !== dyC) {
continue;
}
const pixel = dyBuf.get(b, dyR, dyC, d);
dotProd += pixel;
}
}
dx.set(dotProd * avgMultiplier, b, dxR, dxC, d);
}
}
}
}
return backend.makeTensorInfo(dx.shape, dx.dtype, dx.values);
}
const avgPoolGradConfig$1 = {
kernelName: AvgPoolGrad,
backendName: 'cpu',
kernelFunc: avgPoolGrad$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function batchNorm(args) {
const { inputs, backend, attrs } = args;
const { x, scale, offset, mean, variance } = inputs;
assert$1(mean.shape.length === variance.shape.length, () => 'Batch normalization gradient requires mean and variance to have ' +
'equal ranks.');
assert$1(offset == null || mean.shape.length === offset.shape.length, () => 'Batch normalization gradient requires mean and offset to have ' +
'equal ranks.');
assert$1(scale == null || mean.shape.length === scale.shape.length, () => 'Batch normalization gradient requires mean and scale to have ' +
'equal ranks.');
assertNotComplex([x, mean, variance, scale, offset], 'batchNorm');
let { varianceEpsilon } = attrs;
if (varianceEpsilon == null) {
varianceEpsilon = 0.001;
}
const xVals = backend.data.get(x.dataId).values;
const mVals = backend.data.get(mean.dataId).values;
const varVals = backend.data.get(variance.dataId).values;
const sVals = scale ? backend.data.get(scale.dataId).values :
new Float32Array([1]);
const offVals = offset ?
backend.data.get(offset.dataId).values :
new Float32Array([0]);
const outVals = new Float32Array(xVals.length);
const offValsLength = offVals.length;
const sValsLength = sVals.length;
const varValsLength = varVals.length;
const mValsLength = mVals.length;
let offi = 0;
let mi = 0;
let si = 0;
let vi = 0;
for (let i = 0; i < xVals.length; ++i) {
outVals[i] = offVals[offi++] +
(xVals[i] - mVals[mi++]) * sVals[si++] /
Math.sqrt(varVals[vi++] + varianceEpsilon);
if (offi >= offValsLength) {
offi = 0;
}
if (mi >= mValsLength) {
mi = 0;
}
if (si >= sValsLength) {
si = 0;
}
if (vi >= varValsLength) {
vi = 0;
}
}
return backend.makeTensorInfo(x.shape, x.dtype, outVals);
}
const batchNormConfig = {
kernelName: FusedBatchNorm,
backendName: 'cpu',
kernelFunc: batchNorm,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function batchToSpaceND(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { blockShape, crops } = attrs;
assertNotComplex([x], 'batchToSpaceND');
const prod = blockShape.reduce((a, b) => a * b);
const reshaped = getReshaped(x.shape, blockShape, prod);
const permuted = getPermuted(reshaped.length, blockShape.length);
const reshapedPermuted = getReshapedPermuted(x.shape, blockShape, prod);
const sliceBeginCoords = getSliceBeginCoords(crops, blockShape.length);
const sliceSize = getSliceSize(reshapedPermuted, crops, blockShape.length);
const xReshaped = reshape({ inputs: { x }, backend, attrs: { shape: reshaped } });
const xTransposed = transpose$1({ inputs: { x: xReshaped }, backend, attrs: { perm: permuted } });
const xTransposedReshaped = reshape({ inputs: { x: xTransposed }, backend, attrs: { shape: reshapedPermuted } });
const result = slice$1({
inputs: { x: xTransposedReshaped },
backend,
attrs: { begin: sliceBeginCoords, size: sliceSize }
});
backend.disposeIntermediateTensorInfo(xReshaped);
backend.disposeIntermediateTensorInfo(xTransposed);
backend.disposeIntermediateTensorInfo(xTransposedReshaped);
return result;
}
const batchToSpaceNDConfig = {
kernelName: BatchToSpaceND,
backendName: 'cpu',
kernelFunc: batchToSpaceND
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function bincount(args) {
const { inputs, backend, attrs } = args;
const { x, weights } = inputs;
const { size } = attrs;
const xVals = backend.data.get(x.dataId).values;
const weightsVals = backend.data.get(weights.dataId).values;
const outVals = bincountImpl(xVals, weightsVals, weights.dtype, weights.shape, size);
return backend.makeTensorInfo([size], weights.dtype, outVals);
}
const bincountConfig = {
kernelName: Bincount,
backendName: 'cpu',
kernelFunc: bincount
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function broadcastArgs(args) {
const { inputs, backend } = args;
const { s0, s1 } = inputs;
const s0Vals = backend.data.get(s0.dataId).values;
const s1Vals = backend.data.get(s1.dataId).values;
const broadcastShape = assertAndGetBroadcastShape(Array.from(s0Vals), Array.from(s1Vals));
return backend.makeTensorInfo([broadcastShape.length], 'int32', Int32Array.from(broadcastShape));
}
const broadcastArgsConfig = {
kernelName: BroadcastArgs,
backendName: 'cpu',
kernelFunc: broadcastArgs
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const clipByValue = unaryKernelFunc$1(ClipByValue, (xi, attrs) => {
const clipAttrs = attrs;
if (xi > clipAttrs.clipValueMax) {
return clipAttrs.clipValueMax;
}
return xi < clipAttrs.clipValueMin ? clipAttrs.clipValueMin : xi;
});
const clipByValueConfig = {
kernelName: ClipByValue,
backendName: 'cpu',
kernelFunc: clipByValue,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const complexAbs = (args) => {
const { x } = args.inputs;
const cpuBackend = args.backend;
const resultValues = new Float32Array(sizeFromShape(x.shape));
const complexVals = cpuBackend.data.get(x.dataId);
const real = complexVals.complexTensorInfos.real;
const imag = complexVals.complexTensorInfos.imag;
const realVals = cpuBackend.data.get(real.dataId).values;
const imagVals = cpuBackend.data.get(imag.dataId).values;
for (let i = 0; i < realVals.length; i++) {
const real = realVals[i];
const imag = imagVals[i];
resultValues[i] = Math.hypot(real, imag);
}
return cpuBackend.makeOutput(resultValues, x.shape, 'float32');
};
const complexAbsConfig = {
kernelName: ComplexAbs,
backendName: 'cpu',
kernelFunc: complexAbs,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function imag(args) {
const { inputs, backend } = args;
const { input } = inputs;
const imag = backend.data.get(input.dataId).complexTensorInfos.imag;
const imagVal = backend.data.get(imag.dataId).values;
// When complex tensor is disposed, its underlying parts will be disposed too.
// Make new tensor out of the imag value of the complex. This makes sure the
// value is still accessible even if complex tensor is disposed.
return backend.makeTensorInfo(imag.shape, imag.dtype, imagVal);
}
const imagConfig = {
kernelName: Imag,
backendName: 'cpu',
kernelFunc: imag
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function concat(args) {
const { inputs, backend, attrs } = args;
const { axis } = attrs;
const $axis = parseAxisParam(axis, inputs[0].shape)[0];
const shapes = inputs.map(t => t.shape);
assertParamsConsistent(shapes, $axis);
let outShape = computeOutShape$1(inputs.map(t => t.shape), $axis);
if (sizeFromShape(outShape) === 0) {
return backend.makeTensorInfo(outShape, inputs[0].dtype, []);
}
// Keep only non-empty tensors (ignore tensors with 0 in their shape).
const $inputs = inputs.filter(t => sizeFromShape(t.shape) > 0);
if ($inputs.length === 1) {
return identity$1({ inputs: { x: $inputs[0] }, backend });
}
if ($inputs[0].dtype === 'complex64') {
const reals = $inputs.map((t) => real$1({ inputs: { input: t }, backend }));
const imags = $inputs.map((t) => imag({ inputs: { input: t }, backend }));
const realConcated = concat({ inputs: reals, backend, attrs: { axis: $axis } });
const imagConcated = concat({ inputs: imags, backend, attrs: { axis: $axis } });
const result = complex$1({ inputs: { real: realConcated, imag: imagConcated }, backend });
reals.forEach(r => backend.disposeIntermediateTensorInfo(r));
imags.forEach(i => backend.disposeIntermediateTensorInfo(i));
backend.disposeIntermediateTensorInfo(realConcated);
backend.disposeIntermediateTensorInfo(imagConcated);
return result;
}
// Any concat of n-dimensional tensors across any axis can be reduced to
// a concatenation of two-dimensional tensors across the axis 1 by first
// partitioning the axes of the original tensors into those less than the
// axis to be concatenated and the rest. Then reshape the tensors
// into a two-dimensional tensor by collapsing these two sets of axes and
// concatenate the resulting matrices across the axis 1, finally reshaping
// the result to have the proper shape.
const inputs2D = $inputs.map(t => {
const innerSize = sizeFromShape(t.shape.slice($axis));
const shape = [-1, innerSize];
return reshape({ inputs: { x: t }, backend, attrs: { shape } });
});
const inputsValShapes = inputs2D.map(t => {
return { vals: backend.data.get(t.dataId).values, shape: t.shape };
});
// Concats 2d tensors along axis=1.
outShape =
computeOutShape$1(inputs2D.map(t => t.shape), 1 /* axis */);
const simplyConcat = inputs2D[0].shape[0] === 1;
const outVals = concatImpl$1(inputsValShapes, outShape, inputs[0].dtype, simplyConcat);
const finalOutShape = computeOutShape$1($inputs.map(t => t.shape), $axis);
const outInfo = backend.makeTensorInfo(finalOutShape, inputs[0].dtype, outVals);
inputs2D.forEach(t => backend.disposeIntermediateTensorInfo(t));
return outInfo;
}
const concatConfig = {
kernelName: Concat,
backendName: 'cpu',
kernelFunc: concat
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv2D(args) {
const { inputs, backend, attrs } = args;
const { x, filter } = inputs;
const { strides, pad, dataFormat, dilations, dimRoundingMode } = attrs;
assertNotComplex([x, filter], 'conv2d');
const $dataFormat = convertConv2DDataFormat(dataFormat);
const convInfo = computeConv2DInfo(x.shape, filter.shape, strides, dilations, pad, dimRoundingMode, false /* depthwise */, $dataFormat);
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const padLeft = convInfo.padInfo.left;
const padTop = convInfo.padInfo.top;
const isChannelsLast = convInfo.dataFormat === 'channelsLast';
const y = new TensorBuffer(convInfo.outShape, x.dtype);
const xStrides = computeStrides(x.shape);
const filterStrides = computeStrides(filter.shape);
const xBatchStride = xStrides[0];
const xRowStride = isChannelsLast ? xStrides[1] : xStrides[2];
const xColStride = isChannelsLast ? xStrides[2] : 1;
const xChannelStride = isChannelsLast ? 1 : xStrides[1];
const yBatchStride = y.strides[0];
const yRowStride = isChannelsLast ? y.strides[1] : y.strides[2];
const yColStride = isChannelsLast ? y.strides[2] : 1;
const yChannelStride = isChannelsLast ? 1 : y.strides[1];
const xVals = backend.data.get(x.dataId).values;
const wVals = backend.data.get(filter.dataId).values;
const yVals = y.values;
for (let b = 0; b < convInfo.batchSize; ++b) {
const xOffset1 = b * xBatchStride;
const yOffset1 = b * yBatchStride;
for (let yR = 0; yR < convInfo.outHeight; ++yR) {
const yOffset2 = yOffset1 + yR * yRowStride;
const xRCorner = yR * convInfo.strideHeight - padTop;
for (let wR = 0; wR < filterHeight; ++wR) {
const xR = xRCorner + wR * dilationHeight;
if (xR < 0 || xR >= convInfo.inHeight) {
continue;
}
const wOffset1 = wR * filterStrides[0];
const xOffset2 = xOffset1 + xR * xRowStride;
for (let yC = 0; yC < convInfo.outWidth; ++yC) {
const yOffset3 = yOffset2 + yC * yColStride;
const xCCorner = yC * convInfo.strideWidth - padLeft;
for (let wC = 0; wC < filterWidth; ++wC) {
const xC = xCCorner + wC * dilationWidth;
if (xC < 0 || xC >= convInfo.inWidth) {
continue;
}
const wOffset2 = wOffset1 + wC * filterStrides[1];
const xOffset3 = xOffset2 + xC * xColStride;
let wOffset3 = wOffset2;
for (let d1 = 0; d1 < convInfo.inChannels; ++d1) {
const xVal = xVals[xOffset3 + d1 * xChannelStride];
for (let d2 = 0; d2 < convInfo.outChannels; ++d2) {
yVals[yOffset3 + d2 * yChannelStride] +=
xVal * wVals[wOffset3 + d2];
}
wOffset3 += convInfo.outChannels;
}
}
}
}
}
}
return backend.makeTensorInfo(y.shape, y.dtype, yVals);
}
const conv2DConfig = {
kernelName: Conv2D,
backendName: 'cpu',
kernelFunc: conv2D
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv2DBackpropFilter(args) {
const { inputs, backend, attrs } = args;
const { x, dy } = inputs;
const { strides, pad, dataFormat, dimRoundingMode, filterShape } = attrs;
assertNotComplex([x, dy], 'conv2dBackpropFilter');
const $dataFormat = convertConv2DDataFormat(dataFormat);
const convInfo = computeConv2DInfo(x.shape, filterShape, strides, 1 /* dilations */, pad, dimRoundingMode, false /* depthwise */, $dataFormat);
const { strideHeight, strideWidth, filterHeight, filterWidth } = convInfo;
const isChannelsLast = convInfo.dataFormat === 'channelsLast';
const dW = new TensorBuffer(convInfo.filterShape, 'float32');
const leftPad = convInfo.padInfo.left;
const topPad = convInfo.padInfo.top;
const xVals = backend.data.get(x.dataId).values;
const dyVals = backend.data.get(dy.dataId).values;
const xBuf = new TensorBuffer(x.shape, x.dtype, xVals);
const dyBuf = new TensorBuffer(dy.shape, dy.dtype, dyVals);
for (let wR = 0; wR < filterHeight; ++wR) {
const yRMin = Math.max(0, Math.ceil((topPad - wR) / strideHeight));
const yRMax = Math.min(convInfo.outHeight, (convInfo.inHeight + topPad - wR) / strideHeight);
for (let wC = 0; wC < filterWidth; ++wC) {
const yCMin = Math.max(0, Math.ceil((leftPad - wC) / strideWidth));
const yCMax = Math.min(convInfo.outWidth, (convInfo.inWidth + leftPad - wC) / strideWidth);
for (let d1 = 0; d1 < convInfo.inChannels; ++d1) {
for (let d2 = 0; d2 < convInfo.outChannels; ++d2) {
let dotProd = 0;
for (let b = 0; b < convInfo.batchSize; ++b) {
for (let yR = yRMin; yR < yRMax; ++yR) {
const xR = wR + yR * strideHeight - topPad;
for (let yC = yCMin; yC < yCMax; ++yC) {
const xC = wC + yC * strideWidth - leftPad;
if (isChannelsLast) {
dotProd += xBuf.get(b, xR, xC, d1) *
dyBuf.get(b, yR, yC, d2);
}
else {
dotProd += xBuf.get(b, d1, xR, xC) *
dyBuf.get(b, d2, yR, yC);
}
}
}
}
dW.set(dotProd, wR, wC, d1, d2);
}
}
}
}
return backend.makeTensorInfo(dW.shape, dW.dtype, dW.values);
}
const conv2DBackpropFilterConfig = {
kernelName: Conv2DBackpropFilter,
backendName: 'cpu',
kernelFunc: conv2DBackpropFilter
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv2DBackpropInput(args) {
const { inputs, backend, attrs } = args;
const { dy, filter } = inputs;
const { inputShape, strides, pad, dataFormat, dimRoundingMode } = attrs;
assertNotComplex([dy, filter], 'conv2dBackpropInput');
const filterStrides = computeStrides(filter.shape);
const dyStrides = computeStrides(dy.shape);
let $dataFormat = convertConv2DDataFormat(dataFormat);
const convInfo = computeConv2DInfo(inputShape, filter.shape, strides, 1 /* dilations */, pad, dimRoundingMode, false, $dataFormat);
const dx = new TensorBuffer(convInfo.inShape, 'float32');
const dxValues = dx.values;
const dyValues = backend.data.get(dy.dataId).values;
const fltValues = backend.data.get(filter.dataId).values;
const [fltS0, fltS1, fltS2] = filterStrides;
const { batchSize, filterHeight, filterWidth, inChannels, inHeight, inWidth, outChannels, outHeight, outWidth, strideHeight, strideWidth } = convInfo;
$dataFormat = convInfo.dataFormat;
const topPad = filterHeight - 1 - convInfo.padInfo.top;
const leftPad = filterWidth - 1 - convInfo.padInfo.left;
const isChannelsLast = $dataFormat === 'channelsLast';
const xBatchStride = dx.strides[0];
const xRowStride = isChannelsLast ? dx.strides[1] : dx.strides[2];
const xColStride = isChannelsLast ? dx.strides[2] : 1;
const xChannelStride = isChannelsLast ? 1 : dx.strides[1];
const yBatchStride = dyStrides[0];
const yRowStride = isChannelsLast ? dyStrides[1] : dyStrides[2];
const yColStride = isChannelsLast ? dyStrides[2] : 1;
const yChannelStride = isChannelsLast ? 1 : dyStrides[1];
for (let b = 0; b < batchSize; ++b) {
for (let d1 = 0; d1 < inChannels; ++d1) {
for (let xR = 0; xR < inHeight; ++xR) {
const xRCorner = xR - topPad;
const xRMin = Math.max(0, Math.ceil(xRCorner / strideHeight));
const yRMax = Math.min(outHeight, (filterHeight + xRCorner) / strideHeight);
for (let xC = 0; xC < inWidth; ++xC) {
const xCCorner = xC - leftPad;
const xCMin = Math.max(0, Math.ceil(xCCorner / strideWidth));
const yCMax = Math.min(outWidth, (filterWidth + xCCorner) / strideWidth);
let dotProd = 0;
for (let yR = xRMin; yR < yRMax; ++yR) {
const wR = yR * strideHeight - xRCorner;
for (let yC = xCMin; yC < yCMax; ++yC) {
const wC = yC * strideWidth - xCCorner;
const dyOffset = yBatchStride * b + yRowStride * yR + yColStride * yC;
const fltOffset = fltS0 * (filterHeight - 1 - wR) +
fltS1 * (filterWidth - 1 - wC) + fltS2 * d1;
for (let d2 = 0; d2 < outChannels; ++d2) {
const pixel = dyValues[dyOffset + yChannelStride * d2];
const weight = fltValues[fltOffset + d2];
dotProd += pixel * weight;
}
}
}
const dxOffset = xBatchStride * b + xRowStride * xR +
xColStride * xC + xChannelStride * d1;
dxValues[dxOffset] = dotProd;
}
}
}
}
return backend.makeTensorInfo(dx.shape, dx.dtype, dx.values);
}
const conv2DBackpropInputConfig = {
kernelName: Conv2DBackpropInput,
backendName: 'cpu',
kernelFunc: conv2DBackpropInput
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv3D(args) {
const { inputs, backend, attrs } = args;
const { x, filter } = inputs;
const { strides, pad, dilations } = attrs;
assertNotComplex([x, filter], 'conv3d');
const convInfo = computeConv3DInfo(x.shape, filter.shape, strides, dilations, pad);
const { filterDepth, filterHeight, filterWidth, dilationDepth, dilationHeight, dilationWidth, padInfo } = convInfo;
const padFront = padInfo.front;
const padLeft = padInfo.left;
const padTop = padInfo.top;
const y = new TensorBuffer(convInfo.outShape, x.dtype);
const xVals = backend.data.get(x.dataId).values;
const wVals = backend.data.get(filter.dataId).values;
const yVals = y.values;
const xStrides = computeStrides(x.shape);
const filterStrides = computeStrides(filter.shape);
for (let b = 0; b < convInfo.batchSize; ++b) {
const xOffset1 = b * xStrides[0];
const yOffset1 = b * y.strides[0];
for (let yF = 0; yF < convInfo.outDepth; ++yF) {
const yOffset2 = yOffset1 + yF * y.strides[1];
const xFCorner = yF * convInfo.strideDepth - padFront;
for (let wF = 0; wF < filterDepth; ++wF) {
const xF = xFCorner + wF * dilationDepth;
if (xF < 0 || xF >= convInfo.inDepth) {
continue;
}
const wOffset1 = wF * filterStrides[0];
const xOffset2 = xOffset1 + xF * xStrides[1];
for (let yR = 0; yR < convInfo.outHeight; ++yR) {
const yOffset3 = yOffset2 + yR * y.strides[2];
const xRCorner = yR * convInfo.strideHeight - padTop;
for (let wR = 0; wR < filterHeight; ++wR) {
const xR = xRCorner + wR * dilationHeight;
if (xR < 0 || xR >= convInfo.inHeight) {
continue;
}
const wOffset2 = wOffset1 + wR * filterStrides[1];
const xOffset3 = xOffset2 + xR * xStrides[2];
for (let yC = 0; yC < convInfo.outWidth; ++yC) {
const yOffset4 = yOffset3 + yC * convInfo.outChannels;
const xCCorner = yC * convInfo.strideWidth - padLeft;
for (let wC = 0; wC < filterWidth; ++wC) {
const xC = xCCorner + wC * dilationWidth;
if (xC < 0 || xC >= convInfo.inWidth) {
continue;
}
const wOffset3 = wOffset2 + wC * filterStrides[2];
const xOffset4 = xOffset3 + xC * convInfo.inChannels;
let wOffset4 = wOffset3;
for (let d1 = 0; d1 < convInfo.inChannels; ++d1) {
const xVal = xVals[xOffset4 + d1];
for (let d2 = 0; d2 < convInfo.outChannels; ++d2) {
yVals[yOffset4 + d2] += xVal * wVals[wOffset4 + d2];
}
wOffset4 += convInfo.outChannels;
}
}
}
}
}
}
}
}
return backend.makeTensorInfo(y.shape, y.dtype, y.values);
}
const conv3DConfig = {
kernelName: Conv3D,
backendName: 'cpu',
kernelFunc: conv3D
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv3DBackpropFilterV2(args) {
const { inputs, backend, attrs } = args;
const { x, dy } = inputs;
const { strides, pad, filterShape } = attrs;
assertNotComplex([x, dy], 'conv3dBackpropFilterV2');
const xStrides = computeStrides(x.shape);
const dyStrides = computeStrides(dy.shape);
const convInfo = computeConv3DInfo(x.shape, filterShape, strides, 1 /* dilations */, pad);
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const filterDepth = convInfo.filterDepth;
const filterHeight = convInfo.filterHeight;
const filterWidth = convInfo.filterWidth;
const dw = new TensorBuffer(convInfo.filterShape, 'float32');
const dwValues = dw.values;
const [dwS0, dwS1, dwS2, dwS3] = dw.strides;
const dyValues = backend.data.get(dy.dataId).values;
const [dyS0, dyS1, dyS2, dyS3] = dyStrides;
const xValues = backend.data.get(x.dataId).values;
const [xS0, xS1, xS2, xS3] = xStrides;
const frontPad = convInfo.padInfo.front;
const leftPad = convInfo.padInfo.left;
const topPad = convInfo.padInfo.top;
for (let wF = 0; wF < filterDepth; ++wF) {
const yFMin = Math.max(0, Math.ceil((frontPad - wF) / strideDepth));
const yFMax = Math.min(convInfo.outDepth, (convInfo.inDepth + frontPad - wF) / strideDepth);
const wOffset1 = wF * dwS0;
for (let wR = 0; wR < filterHeight; ++wR) {
const yRMin = Math.max(0, Math.ceil((topPad - wR) / strideHeight));
const yRMax = Math.min(convInfo.outHeight, (convInfo.inHeight + topPad - wR) / strideHeight);
const wOffset2 = wR * dwS1 + wOffset1;
for (let wC = 0; wC < filterWidth; ++wC) {
const yCMin = Math.max(0, Math.ceil((leftPad - wC) / strideWidth));
const yCMax = Math.min(convInfo.outWidth, (convInfo.inWidth + leftPad - wC) / strideWidth);
const wOffset3 = wC * dwS2 + wOffset2;
for (let d1 = 0; d1 < convInfo.inChannels; ++d1) {
const wOffset4 = d1 * dwS3 + wOffset3;
for (let d2 = 0; d2 < convInfo.outChannels; ++d2) {
let dotProd = 0;
for (let b = 0; b < convInfo.batchSize; ++b) {
const xOffset1 = b * xS0;
const yOffset1 = b * dyS0;
for (let yF = yFMin; yF < yFMax; ++yF) {
const xF = wF + yF * strideDepth - frontPad;
const xOffset2 = xF * xS1 + xOffset1;
const yOffset2 = yF * dyS1 + yOffset1;
for (let yR = yRMin; yR < yRMax; ++yR) {
const xR = wR + yR * strideHeight - topPad;
const xOffset3 = xR * xS2 + xOffset2;
const yOffset3 = yR * dyS2 + yOffset2;
for (let yC = yCMin; yC < yCMax; ++yC) {
const xC = wC + yC * strideWidth - leftPad;
const xOffset4 = xC * xS3 + xOffset3;
const yOffset4 = yC * dyS3 + yOffset3;
dotProd += xValues[xOffset4 + d1] * dyValues[yOffset4 + d2];
}
}
}
}
dwValues[wOffset4 + d2] = dotProd;
}
}
}
}
}
return backend.makeTensorInfo(dw.shape, dw.dtype, dw.values);
}
const conv3DBackpropFilterV2Config = {
kernelName: Conv3DBackpropFilterV2,
backendName: 'cpu',
kernelFunc: conv3DBackpropFilterV2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function conv3DBackpropInputV2(args) {
const { inputs, backend, attrs } = args;
const { dy, filter } = inputs;
const { pad, strides, inputShape } = attrs;
assertNotComplex([dy], 'conv3dBackpropInputV2');
const dyStrides = computeStrides(dy.shape);
const filterStrides = computeStrides(filter.shape);
const convInfo = computeConv3DInfo(inputShape, filter.shape, strides, 1 /* dilations */, pad);
const dx = new TensorBuffer(convInfo.inShape, 'float32');
const dxValues = dx.values;
const [dxS0, dxS1, dxS2, dxS3] = dx.strides;
const dyValues = backend.data.get(dy.dataId).values;
const [dyS0, dyS1, dyS2, dyS3] = dyStrides;
const fltValues = backend.data.get(filter.dataId).values;
const [fltS0, fltS1, fltS2, fltS3] = filterStrides;
const { batchSize, filterDepth, filterHeight, filterWidth, inChannels, inDepth, inHeight, inWidth, outChannels, outDepth, outHeight, outWidth, strideDepth, strideHeight, strideWidth } = convInfo;
const frontPad = filterDepth - 1 - convInfo.padInfo.front;
const topPad = filterHeight - 1 - convInfo.padInfo.top;
const leftPad = filterWidth - 1 - convInfo.padInfo.left;
for (let b = 0; b < batchSize; ++b) {
for (let d1 = 0; d1 < inChannels; ++d1) {
// Frames of depth
for (let xF = 0; xF < inDepth; ++xF) {
const xFCorner = xF - frontPad;
const xFMin = Math.max(0, Math.ceil(xFCorner / strideDepth));
const yFMax = Math.min(outDepth, (filterDepth + xFCorner) / strideDepth);
// Rows as per standard 2d matrix notation
for (let xR = 0; xR < inHeight; ++xR) {
const xRCorner = xR - topPad;
const xRMin = Math.max(0, Math.ceil(xRCorner / strideHeight));
const yRMax = Math.min(outHeight, (filterHeight + xRCorner) / strideHeight);
// Columns as per standard 2d matrix notation
for (let xC = 0; xC < inWidth; ++xC) {
const xCCorner = xC - leftPad;
const xCMin = Math.max(0, Math.ceil(xCCorner / strideWidth));
const yCMax = Math.min(outWidth, (filterWidth + xCCorner) / strideWidth);
let dotProd = 0;
for (let yF = xFMin; yF < yFMax; ++yF) {
const wF = yF * strideDepth - xFCorner;
for (let yR = xRMin; yR < yRMax; ++yR) {
const wR = yR * strideHeight - xRCorner;
for (let yC = xCMin; yC < yCMax; ++yC) {
const wC = yC * strideWidth - xCCorner;
const dyOffset = dyS0 * b + dyS1 * yF + dyS2 * yR + dyS3 * yC;
const fltOffset = fltS0 * (filterDepth - 1 - wF) +
fltS1 * (filterHeight - 1 - wR) +
fltS2 * (filterWidth - 1 - wC) + fltS3 * d1;
for (let d2 = 0; d2 < outChannels; ++d2) {
const pixel = dyValues[dyOffset + d2];
const weight = fltValues[fltOffset + d2];
dotProd += pixel * weight;
}
}
}
}
dxValues[dxS0 * b + dxS1 * xF + dxS2 * xR + dxS3 * xC + d1] =
dotProd;
}
}
}
}
}
return backend.makeTensorInfo(dx.shape, dx.dtype, dx.values);
}
const conv3DBackpropInputV2Config = {
kernelName: Conv3DBackpropInputV2,
backendName: 'cpu',
kernelFunc: conv3DBackpropInputV2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const cos = unaryKernelFunc$1(Cos, (xi) => Math.cos(xi));
const cosConfig = {
kernelName: Cos,
backendName: 'cpu',
kernelFunc: cos,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const cosh = unaryKernelFunc$1(Cosh, (xi) => Math.cosh(xi));
const coshConfig = {
kernelName: Cosh,
backendName: 'cpu',
kernelFunc: cosh,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function cropAndResize(args) {
const { inputs, backend, attrs } = args;
const { image, boxes, boxInd } = inputs;
const { cropSize, method, extrapolationValue } = attrs;
const [batch, imageHeight, imageWidth, numChannels] = image.shape;
const numBoxes = boxes.shape[0];
const [cropHeight, cropWidth] = cropSize;
const output = buffer([numBoxes, cropHeight, cropWidth, numChannels], 'float32');
const boxVals = backend.data.get(boxes.dataId).values;
const boxIndVals = backend.data.get(boxInd.dataId).values;
const imageVals = backend.data.get(image.dataId).values;
const inStride = computeStrides(image.shape); // to calculate flat indexes into image
const outStride = computeStrides(output.shape); // to calculate flat indexes into output
// Reference implementation
// tslint:disable-next-line:max-line-length
// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/crop_and_resize_op.cc
for (let b = 0; b < numBoxes; b++) {
const startInd = b * 4;
const y1 = boxVals[startInd];
const x1 = boxVals[startInd + 1];
const y2 = boxVals[startInd + 2];
const x2 = boxVals[startInd + 3];
const bInd = boxIndVals[b];
if (bInd >= batch) {
continue;
}
const heightScale = (cropHeight > 1) ? (y2 - y1) * (imageHeight - 1) / (cropHeight - 1) : 0;
const widthScale = (cropWidth > 1) ? (x2 - x1) * (imageWidth - 1) / (cropWidth - 1) : 0;
for (let y = 0; y < cropHeight; y++) {
const yInd = (cropHeight > 1) ?
y1 * (imageHeight - 1) + y * (heightScale) :
0.5 * (y1 + y2) * (imageHeight - 1);
if (yInd < 0 || yInd > imageHeight - 1) {
for (let x = 0; x < cropWidth; x++) {
for (let c = 0; c < numChannels; c++) {
const ind = c + x * outStride[2] + y * outStride[1] + b * outStride[0];
output.values[ind] = extrapolationValue;
}
}
continue;
}
if (method === 'bilinear') {
const topInd = Math.floor(yInd);
const bottomInd = Math.ceil(yInd);
const yLerp = yInd - topInd;
for (let x = 0; x < cropWidth; x++) {
const xInd = (cropWidth > 1) ?
x1 * (imageWidth - 1) + x * widthScale :
0.5 * (x1 + x2) * (imageWidth - 1);
if (xInd < 0 || xInd > imageWidth - 1) {
for (let c = 0; c < numChannels; c++) {
const ind = c + x * outStride[2] + y * outStride[1] + b * outStride[0];
output.values[ind] = extrapolationValue;
}
continue;
}
const leftInd = Math.floor(xInd);
const rightInd = Math.ceil(xInd);
const xLerp = xInd - leftInd;
for (let c = 0; c < numChannels; c++) {
let ind = c + leftInd * inStride[2] + topInd * inStride[1] +
bInd * inStride[0];
const topLeft = imageVals[ind];
ind = c + rightInd * inStride[2] + topInd * inStride[1] +
bInd * inStride[0];
const topRight = imageVals[ind];
ind = c + leftInd * inStride[2] + bottomInd * inStride[1] +
bInd * inStride[0];
const bottomLeft = imageVals[ind];
ind = c + rightInd * inStride[2] + bottomInd * inStride[1] +
bInd * inStride[0];
const bottomRight = imageVals[ind];
const top = topLeft + (topRight - topLeft) * xLerp;
const bottom = bottomLeft + (bottomRight - bottomLeft) * xLerp;
ind = c + x * outStride[2] + y * outStride[1] + b * outStride[0];
output.values[ind] = top + ((bottom - top) * yLerp);
}
}
}
else { // method == "nearest"
for (let x = 0; x < cropWidth; ++x) {
const xInd = (cropWidth > 1) ?
x1 * (imageWidth - 1) + x * widthScale :
0.5 * (x1 + x2) * (imageWidth - 1);
if (xInd < 0 || xInd > imageWidth - 1) {
for (let c = 0; c < numChannels; c++) {
const ind = c + x * outStride[2] + y * outStride[1] + b * outStride[0];
output.values[ind] = extrapolationValue;
}
continue;
}
const closestX = Math.round(xInd);
const closestY = Math.round(yInd);
for (let c = 0; c < numChannels; c++) {
const inInd = c + closestX * inStride[2] + closestY * inStride[1] +
bInd * inStride[0];
const outInd = c + x * outStride[2] + y * outStride[1] + b * outStride[0];
output.values[outInd] = imageVals[inInd];
}
}
}
}
}
return backend.makeTensorInfo(output.shape, output.dtype, output.values);
}
const cropAndResizeConfig = {
kernelName: CropAndResize,
backendName: 'cpu',
kernelFunc: cropAndResize
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function cumprod(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, exclusive, reverse } = attrs;
assertNotComplex(x, 'cumprod');
const permutation = getAxesPermutation([axis], x.shape.length);
let $x = x;
if (permutation != null) {
$x = transpose$1({ inputs: { x }, backend, attrs: { perm: permutation } });
}
const permutedAxis = getInnerMostAxes(1, x.shape.length)[0];
if (permutedAxis !== $x.shape.length - 1) {
throw new Error(`backend.cumprod in CPU expects an inner-most ` +
`axis=${$x.shape.length - 1} but got axis=${permutedAxis}`);
}
const resultDtype = upcastType($x.dtype, 'int32');
const vals = makeOnesTypedArray(sizeFromShape($x.shape), resultDtype);
const aVals = backend.data.get($x.dataId).values;
const finalDim = $x.shape[$x.shape.length - 1];
const indexAdjuster = reverse ?
(i, j) => i + finalDim - j - 1 :
(i, j) => i + j;
for (let i = 0; i < aVals.length; i += finalDim) {
for (let j = 0; j < finalDim; j++) {
const idx = indexAdjuster(i, j);
if (j === 0) {
vals[idx] = exclusive ? 1 : aVals[idx];
}
else {
const prevIdx = indexAdjuster(i, j - 1);
vals[idx] = exclusive ? aVals[prevIdx] * vals[prevIdx] :
aVals[idx] * vals[prevIdx];
}
}
}
const result = backend.makeTensorInfo($x.shape, resultDtype, vals);
if (permutation != null) {
const reversePermutation = getUndoAxesPermutation(permutation);
const reverseTransposedResult = transpose$1({ inputs: { x: result }, backend, attrs: { perm: reversePermutation } });
backend.disposeIntermediateTensorInfo(result);
backend.disposeIntermediateTensorInfo($x);
return reverseTransposedResult;
}
return result;
}
const cumprodConfig = {
kernelName: Cumprod,
backendName: 'cpu',
kernelFunc: cumprod
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function cumsum(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, exclusive, reverse } = attrs;
assertNotComplex(x, 'cumsum');
const permutation = getAxesPermutation([axis], x.shape.length);
let $x = x;
if (permutation != null) {
$x = transpose$1({ inputs: { x }, backend, attrs: { perm: permutation } });
}
const permutedAxis = getInnerMostAxes(1, x.shape.length)[0];
if (permutedAxis !== $x.shape.length - 1) {
throw new Error(`backend.cumsum in CPU expects an inner-most ` +
`axis=${$x.shape.length - 1} but got axis=${permutedAxis}`);
}
const resultDtype = upcastType($x.dtype, 'int32');
const vals = makeZerosTypedArray(sizeFromShape($x.shape), resultDtype);
const aVals = backend.data.get($x.dataId).values;
const finalDim = $x.shape[$x.shape.length - 1];
const indexAdjuster = reverse ?
(i, j) => i + finalDim - j - 1 :
(i, j) => i + j;
for (let i = 0; i < aVals.length; i += finalDim) {
for (let j = 0; j < finalDim; j++) {
const idx = indexAdjuster(i, j);
if (j === 0) {
vals[idx] = exclusive ? 0 : aVals[idx];
}
else {
const prevIdx = indexAdjuster(i, j - 1);
vals[idx] = exclusive ? aVals[prevIdx] + vals[prevIdx] :
aVals[idx] + vals[prevIdx];
}
}
}
const result = backend.makeTensorInfo($x.shape, resultDtype, vals);
if (permutation != null) {
const reversePermutation = getUndoAxesPermutation(permutation);
const reverseTransposedResult = transpose$1({ inputs: { x: result }, backend, attrs: { perm: reversePermutation } });
backend.disposeIntermediateTensorInfo(result);
backend.disposeIntermediateTensorInfo($x);
return reverseTransposedResult;
}
return result;
}
const cumsumConfig = {
kernelName: Cumsum,
backendName: 'cpu',
kernelFunc: cumsum
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function denseBincount(args) {
const { inputs, backend, attrs } = args;
const { x, weights } = inputs;
const { size, binaryOutput } = attrs;
if (x.shape.length === 1) {
const xVals = backend.data.get(x.dataId).values;
const weightsVals = backend.data.get(weights.dataId).values;
const outVals = bincountImpl(xVals, weightsVals, weights.dtype, weights.shape, size);
return backend.makeTensorInfo([size], weights.dtype, outVals);
}
else if (x.shape.length === 2) {
const xBuf = backend.bufferSync(x);
const weightsBuf = backend.bufferSync(weights);
const outBuf = bincountReduceImpl(xBuf, weightsBuf, size, binaryOutput);
return backend.makeTensorInfo(outBuf.shape, weights.dtype, outBuf.values);
}
throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank` +
`${x.shape.length}.`);
}
const denseBincountConfig = {
kernelName: DenseBincount,
backendName: 'cpu',
kernelFunc: denseBincount
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthToSpace(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { blockSize, dataFormat } = attrs;
assert$1(dataFormat === 'NHWC', () => `Only NHWC dataFormat supported on CPU for depthToSpace. Got ${dataFormat}`);
const batchSize = x.shape[0];
const inputHeight = x.shape[1];
const inputWidth = x.shape[2];
const inputDepth = x.shape[3];
const outputHeight = inputHeight * blockSize;
const outputWidth = inputWidth * blockSize;
const outputDepth = inputDepth / (blockSize * blockSize);
const xValues = backend.data.get(x.dataId).values;
const result = new Float32Array(batchSize * outputHeight * outputWidth * outputDepth);
let outputIdx = 0;
for (let b = 0; b < batchSize; ++b) {
for (let h = 0; h < outputHeight; ++h) {
const inH = Math.floor(h / blockSize);
const offsetH = (h % blockSize);
for (let w = 0; w < outputWidth; ++w) {
const inW = Math.floor(w / blockSize);
const offsetW = (w % blockSize);
const offsetD = (offsetH * blockSize + offsetW) * outputDepth;
for (let d = 0; d < outputDepth; ++d) {
const inD = d + offsetD;
const inputIdx = inD + inputDepth * (inW + inputWidth * (inH + inputHeight * b));
result[outputIdx++] = xValues[inputIdx];
}
}
}
}
return backend.makeTensorInfo([batchSize, outputHeight, outputWidth, outputDepth], x.dtype, result);
}
const depthToSpaceConfig = {
kernelName: DepthToSpace,
backendName: 'cpu',
kernelFunc: depthToSpace
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNative(args) {
const { inputs, backend, attrs } = args;
const { x, filter } = inputs;
const { strides, pad, dilations, dimRoundingMode } = attrs;
assertNotComplex([x, filter], 'depthwiseConv2DNative');
const xStrides = computeStrides(x.shape);
const filterStrides = computeStrides(filter.shape);
let $dilations = dilations;
if ($dilations == null) {
$dilations = [1, 1];
}
assert$1(eitherStridesOrDilationsAreOne(strides, $dilations), () => 'Error in depthwiseConv2d: Either strides or dilations must be ' +
`1. Got strides ${strides} and dilations '${$dilations}'`);
const convInfo = computeConv2DInfo(x.shape, filter.shape, strides, $dilations, pad, dimRoundingMode, true /* depthwise */);
const { filterHeight, filterWidth, dilationHeight, dilationWidth, padInfo } = convInfo;
const padLeft = padInfo.left;
const padTop = padInfo.top;
const chMul = convInfo.outChannels / convInfo.inChannels;
const y = new TensorBuffer(convInfo.outShape, x.dtype);
const xVals = backend.data.get(x.dataId).values;
const wVals = backend.data.get(filter.dataId).values;
const yVals = y.values;
for (let b = 0; b < convInfo.batchSize; ++b) {
const xOffset1 = b * xStrides[0];
const yOffset1 = b * y.strides[0];
for (let yR = 0; yR < convInfo.outHeight; ++yR) {
const yOffset2 = yOffset1 + yR * y.strides[1];
const xRCorner = yR * convInfo.strideHeight - padTop;
for (let wR = 0; wR < filterHeight; ++wR) {
const xR = xRCorner + wR * dilationHeight;
if (xR < 0 || xR >= convInfo.inHeight) {
continue;
}
const wOffset1 = wR * filterStrides[0];
const xOffset2 = xOffset1 + xR * xStrides[1];
for (let yC = 0; yC < convInfo.outWidth; ++yC) {
const yOffset3 = yOffset2 + yC * y.strides[2];
const xCCorner = yC * convInfo.strideWidth - padLeft;
for (let wC = 0; wC < filterWidth; ++wC) {
const xC = xCCorner + wC * dilationWidth;
if (xC < 0 || xC >= convInfo.inWidth) {
continue;
}
const wOffset2 = wOffset1 + wC * filterStrides[1];
const xOffset3 = xOffset2 + xC * convInfo.inChannels;
let yOffset4 = yOffset3;
let wOffset3 = wOffset2;
for (let d1 = 0; d1 < convInfo.inChannels; ++d1) {
const xVal = xVals[xOffset3 + d1];
for (let q = 0; q < chMul; ++q) {
yVals[yOffset4 + q] += xVal * wVals[wOffset3 + q];
}
yOffset4 += chMul;
wOffset3 += chMul;
}
}
}
}
}
}
return backend.makeTensorInfo(y.shape, y.dtype, y.values);
}
const depthwiseConv2dNativeConfig = {
kernelName: DepthwiseConv2dNative,
backendName: 'cpu',
kernelFunc: depthwiseConv2dNative
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNativeBackpropFilter(args) {
const { inputs, backend, attrs } = args;
const { x, dy } = inputs;
const { strides, dilations, pad, dimRoundingMode, filterShape } = attrs;
assertNotComplex([x, dy], 'depthwiseConv2dNativeBackpropFilter');
const convInfo = computeConv2DInfo(x.shape, filterShape, strides, dilations, pad, dimRoundingMode, true /* depthwise */);
const { strideHeight, strideWidth, filterHeight, filterWidth } = convInfo;
const dW = new TensorBuffer(convInfo.filterShape, 'float32');
const leftPad = convInfo.padInfo.left;
const topPad = convInfo.padInfo.top;
const chMul = convInfo.outChannels / convInfo.inChannels;
const xVals = backend.data.get(x.dataId).values;
const xBuf = new TensorBuffer(x.shape, x.dtype, xVals);
const dyVals = backend.data.get(dy.dataId).values;
const dyBuf = new TensorBuffer(dy.shape, dy.dtype, dyVals);
for (let wR = 0; wR < filterHeight; ++wR) {
const yRMin = Math.max(0, Math.ceil((topPad - wR) / strideHeight));
const yRMax = Math.min(convInfo.outHeight, (convInfo.inHeight + topPad - wR) / strideHeight);
for (let wC = 0; wC < filterWidth; ++wC) {
const yCMin = Math.max(0, Math.ceil((leftPad - wC) / strideWidth));
const yCMax = Math.min(convInfo.outWidth, (convInfo.inWidth + leftPad - wC) / strideWidth);
for (let d2 = 0; d2 < convInfo.outChannels; ++d2) {
const d1 = Math.trunc(d2 / chMul);
const dm = d2 % chMul;
let dotProd = 0;
for (let b = 0; b < convInfo.batchSize; ++b) {
for (let yR = yRMin; yR < yRMax; ++yR) {
const xR = wR + yR * strideHeight - topPad;
for (let yC = yCMin; yC < yCMax; ++yC) {
const xC = wC + yC * strideWidth - leftPad;
dotProd += xBuf.get(b, xR, xC, d1) *
dyBuf.get(b, yR, yC, d2);
}
}
}
dW.set(dotProd, wR, wC, d1, dm);
}
}
}
return backend.makeTensorInfo(dW.shape, dW.dtype, dW.values);
}
const depthwiseConv2dNativeBackpropFilterConfig = {
kernelName: DepthwiseConv2dNativeBackpropFilter,
backendName: 'cpu',
kernelFunc: depthwiseConv2dNativeBackpropFilter
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function depthwiseConv2dNativeBackpropInput(args) {
const { inputs, backend, attrs } = args;
const { dy, filter } = inputs;
const { strides, dilations, pad, dimRoundingMode, inputShape } = attrs;
assertNotComplex([dy, filter], 'depthwiseConv2DNativeBackpropInput');
const dyStrides = computeStrides(dy.shape);
const filterStrides = computeStrides(filter.shape);
const convInfo = computeConv2DInfo(inputShape, filter.shape, strides, dilations, pad, dimRoundingMode, true /* depthwise */);
const dx = new TensorBuffer(convInfo.inShape, 'float32');
const dxValues = dx.values;
const [dxS0, dxS1, dxS2] = dx.strides;
const dyValues = backend.data.get(dy.dataId).values;
const [dyS0, dyS1, dyS2] = dyStrides;
const fltValues = backend.data.get(filter.dataId).values;
const [fltS0, fltS1, fltS2] = filterStrides;
const { batchSize, filterHeight, filterWidth, inChannels, inHeight, inWidth, outChannels, outHeight, outWidth, strideHeight, strideWidth } = convInfo;
const topPad = filterHeight - 1 - convInfo.padInfo.top;
const leftPad = filterWidth - 1 - convInfo.padInfo.left;
const chMul = outChannels / inChannels;
for (let b = 0; b < batchSize; ++b) {
for (let d1 = 0; d1 < inChannels; ++d1) {
for (let xR = 0; xR < inHeight; ++xR) {
const xRCorner = xR - topPad;
const xRMin = Math.max(0, Math.ceil(xRCorner / strideHeight));
const yRMax = Math.min(outHeight, (filterHeight + xRCorner) / strideHeight);
for (let xC = 0; xC < inWidth; ++xC) {
const xCCorner = xC - leftPad;
const xCMin = Math.max(0, Math.ceil(xCCorner / strideWidth));
const yCMax = Math.min(outWidth, (filterWidth + xCCorner) / strideWidth);
let dotProd = 0;
for (let yR = xRMin; yR < yRMax; ++yR) {
const wR = yR * strideHeight - xRCorner;
for (let yC = xCMin; yC < yCMax; ++yC) {
const wC = yC * strideWidth - xCCorner;
const dyOffset = dyS0 * b + dyS1 * yR + dyS2 * yC;
const fltOffset = fltS0 * (filterHeight - 1 - wR) +
fltS1 * (filterWidth - 1 - wC) + fltS2 * d1;
for (let dm = 0; dm < chMul; ++dm) {
const d2 = d1 * chMul + dm;
const pixel = dyValues[dyOffset + d2];
const weight = fltValues[fltOffset + dm];
dotProd += pixel * weight;
}
}
}
dxValues[dxS0 * b + dxS1 * xR + dxS2 * xC + d1] = dotProd;
}
}
}
}
return backend.makeTensorInfo(dx.shape, dx.dtype, dx.values);
}
const depthwiseConv2dNativeBackpropInputConfig = {
kernelName: DepthwiseConv2dNativeBackpropInput,
backendName: 'cpu',
kernelFunc: depthwiseConv2dNativeBackpropInput
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function diag(args) {
const { inputs, backend } = args;
const { x } = inputs;
const xSize = sizeFromShape(x.shape);
const xVals = backend.data.get(x.dataId).values;
const outBuf = buffer([xSize, xSize], x.dtype);
const vals = outBuf.values;
for (let i = 0; i < xVals.length; i++) {
vals[i * xSize + i] = xVals[i];
}
const outShape = [...x.shape, ...x.shape];
return backend.makeTensorInfo(outShape, outBuf.dtype, outBuf.values);
}
const diagConfig = {
kernelName: Diag,
backendName: 'cpu',
kernelFunc: diag
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const dilation2DConfig = {
kernelName: Dilation2D,
backendName: 'cpu',
kernelFunc: ({ inputs, backend, attrs }) => {
const { x, filter } = inputs;
const { strides, pad, dilations } = attrs;
const cpuBackend = backend;
const xVals = cpuBackend.data.get(x.dataId).values;
const xRank = x.shape.length;
const filterVals = cpuBackend.data.get(filter.dataId).values;
const filterRank = filter.shape.length;
const { batchSize, inHeight, inWidth, inChannels, outHeight, outWidth, padInfo, strideHeight, strideWidth, filterHeight, filterWidth, dilationHeight, dilationWidth, outShape } = computeDilation2DInfo(x.shape, filter.shape, strides, pad, 'NHWC' /* dataFormat */, dilations);
const outSize = sizeFromShape(outShape);
const outRank = outShape.length;
const outputVals = getArrayFromDType(x.dtype, outSize);
// Upsampling the input by fill in `dilation size - 1` values between each
// input value.
// This implementation follows the TF c++ implementation:
// https://github.com/tensorflow/tensorflow/blob/d9a3a849edc198e90172bc58eb293de457f9d986/tensorflow/core/kernels/dilation_ops.cc
for (let b = 0; b < batchSize; ++b) {
for (let hOut = 0; hOut < outHeight; ++hOut) {
const hBeg = hOut * strideHeight - padInfo.top;
for (let wOut = 0; wOut < outWidth; ++wOut) {
const wBeg = wOut * strideWidth - padInfo.left;
for (let d = 0; d < inChannels; ++d) {
let curVal = Number.MIN_SAFE_INTEGER;
for (let h = 0; h < filterHeight; ++h) {
const hIn = hBeg + h * dilationHeight;
if (hIn >= 0 && hIn < inHeight) {
for (let w = 0; w < filterWidth; ++w) {
const wIn = wBeg + w * dilationWidth;
if (wIn >= 0 && wIn < inWidth) {
const xIndex = locToIndex([b, hIn, wIn, d], xRank, computeStrides(x.shape));
const filterIndex = locToIndex([h, w, d], filterRank, computeStrides(filter.shape));
const val = xVals[xIndex] + filterVals[filterIndex];
if (val > curVal) {
curVal = val;
}
}
}
}
}
const outputIndex = locToIndex([b, hOut, wOut, d], outRank, computeStrides(outShape));
outputVals[outputIndex] = curVal;
}
}
}
}
const dataId = cpuBackend.write(toTypedArray(outputVals, x.dtype), outShape, x.dtype);
return { dataId, shape: outShape, dtype: x.dtype };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const dilation2DBackpropFilterConfig = {
kernelName: Dilation2DBackpropFilter,
backendName: 'cpu',
kernelFunc: ({ inputs, backend, attrs }) => {
const { x, filter, dy } = inputs;
const { strides, pad, dilations } = attrs;
const cpuBackend = backend;
const $x = toNestedArray(x.shape, cpuBackend.data.get(x.dataId).values);
const $filter = toNestedArray(filter.shape, cpuBackend.data.get(filter.dataId).values);
const { batchSize, inHeight, inWidth, inChannels, outHeight, outWidth, padInfo, strideHeight, strideWidth, filterHeight, filterWidth, dilationHeight, dilationWidth, outShape } = computeDilation2DInfo(x.shape, filter.shape, strides, pad, 'NHWC' /* dataFormat */, dilations);
assert$1(dy.rank === outShape.length, () => `Error in ${Dilation2DBackpropFilter}, dy ` +
`must have the same rank as output ${outShape.length}, but got ` +
`${dy.rank}`);
const $dy = toNestedArray(outShape, cpuBackend.data.get(dy.dataId).values);
// The computed filter gradients has the same dimensions as the filter:
// [filterHeight, filterWidth, depth]
const gradients = makeZerosNestedTypedArray(filter.shape, filter.dtype);
// In the case of multiple argmax branches, we only back-propagate along the
// last branch, i.e., the one with largest value of `h * filter_cols + w`,
// similarly to the max-pooling backward routines.
// This implementation follows the TF c++ implementation:
// https://github.com/tensorflow/tensorflow/blob/d9a3a849edc198e90172bc58eb293de457f9d986/tensorflow/core/kernels/dilation_ops.cc
for (let b = 0; b < batchSize; ++b) {
for (let hOut = 0; hOut < outHeight; ++hOut) {
const hBeg = hOut * strideHeight - padInfo.top;
for (let wOut = 0; wOut < outWidth; ++wOut) {
const wBeg = wOut * strideWidth - padInfo.left;
for (let d = 0; d < inChannels; ++d) {
let curVal = Number.MIN_SAFE_INTEGER;
let hMax = 0;
let wMax = 0;
for (let h = 0; h < filterHeight; ++h) {
const hIn = hBeg + h * dilationHeight;
if (hIn >= 0 && hIn < inHeight) {
for (let w = 0; w < filterWidth; ++w) {
const wIn = wBeg + w * dilationWidth;
if (wIn >= 0 && wIn < inWidth) {
const val = $x[b][hIn][wIn][d] + $filter[h][w][d];
if (val > curVal) {
curVal = val;
hMax = h;
wMax = w;
}
}
}
}
}
gradients[hMax][wMax][d] += $dy[b][hOut][wOut][d];
}
}
}
}
const dataId = cpuBackend.write(toTypedArray(gradients, x.dtype), filter.shape, filter.dtype);
return { dataId, shape: filter.shape, dtype: filter.dtype };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const dilation2DBackpropInputConfig = {
kernelName: Dilation2DBackpropInput,
backendName: 'cpu',
kernelFunc: ({ inputs, backend, attrs }) => {
const { x, filter, dy } = inputs;
const { strides, pad, dilations } = attrs;
const cpuBackend = backend;
const $x = toNestedArray(x.shape, cpuBackend.data.get(x.dataId).values);
const $filter = toNestedArray(filter.shape, cpuBackend.data.get(filter.dataId).values);
const { batchSize, inHeight, inWidth, inChannels, outHeight, outWidth, padInfo, strideHeight, strideWidth, filterHeight, filterWidth, dilationHeight, dilationWidth, outShape } = computeDilation2DInfo(x.shape, filter.shape, strides, pad, 'NHWC' /* dataFormat */, dilations);
assert$1(dy.rank === outShape.length, () => `Error in ${Dilation2DBackpropInput}, dy ` +
`must have the same rank as output ${outShape.length}, but got ` +
`${dy.rank}`);
const $dy = toNestedArray(outShape, cpuBackend.data.get(dy.dataId).values);
// The computed gradients has the same dimensions as the input:
// [batch, inputHeight, inputCols, inChannel]
const gradients = makeZerosNestedTypedArray(x.shape, x.dtype);
// In the case of multiple argmax branches, we only back-propagate along the
// last branch, i.e., the one with largest value of `h * filter_cols + w`,
// similarly to the max-pooling backward routines.
// This implementation follows the TF c++ implementation:
// https://github.com/tensorflow/tensorflow/blob/d9a3a849edc198e90172bc58eb293de457f9d986/tensorflow/core/kernels/dilation_ops.cc
for (let b = 0; b < batchSize; ++b) {
for (let hOut = 0; hOut < outHeight; ++hOut) {
const hBeg = hOut * strideHeight - padInfo.top;
for (let wOut = 0; wOut < outWidth; ++wOut) {
const wBeg = wOut * strideWidth - padInfo.left;
for (let d = 0; d < inChannels; ++d) {
let curVal = Number.MIN_SAFE_INTEGER;
let hInMax = (hBeg < 0) ? 0 : hBeg;
let wInMax = (wBeg < 0) ? 0 : wBeg;
for (let h = 0; h < filterHeight; ++h) {
const hIn = hBeg + h * dilationHeight;
if (hIn >= 0 && hIn < inHeight) {
for (let w = 0; w < filterWidth; ++w) {
const wIn = wBeg + w * dilationWidth;
if (wIn >= 0 && wIn < inWidth) {
const val = $x[b][hIn][wIn][d] + $filter[h][w][d];
if (val > curVal) {
curVal = val;
hInMax = hIn;
wInMax = wIn;
}
}
}
}
}
gradients[b][hInMax][wInMax][d] += $dy[b][hOut][wOut][d];
}
}
}
}
const dataId = cpuBackend.write(toTypedArray(gradients, x.dtype), x.shape, x.dtype);
return { dataId, shape: x.shape, dtype: x.dtype };
}
};
/**
* @license
* Copyright 2023 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function draw(args) {
const { inputs, backend, attrs } = args;
const { image } = inputs;
const { canvas, options } = attrs;
const { contextOptions, imageOptions } = options || {};
const alpha = (imageOptions === null || imageOptions === void 0 ? void 0 : imageOptions.alpha) || 1;
const contextType = (contextOptions === null || contextOptions === void 0 ? void 0 : contextOptions.contextType) || '2d';
if (contextType !== '2d') {
throw new Error(`Context type ${contextOptions.contextType} is not supported by the CPU backend.`);
}
const ctx = canvas.getContext(contextType, (contextOptions === null || contextOptions === void 0 ? void 0 : contextOptions.contextAttributes) || {});
if (ctx == null) {
throw new Error(`Could not get the context with ${contextType} type.`);
}
const [height, width] = image.shape.slice(0, 2);
const depth = image.shape.length === 2 ? 1 : image.shape[2];
const data = backend.data.get(image.dataId).values;
const multiplier = image.dtype === 'float32' ? 255 : 1;
const bytes = new Uint8ClampedArray(width * height * 4);
for (let i = 0; i < height * width; ++i) {
const rgba = [0, 0, 0, 255 * alpha];
for (let d = 0; d < depth; d++) {
const value = data[i * depth + d];
if (image.dtype === 'float32') {
if (value < 0 || value > 1) {
throw new Error(`Tensor values for a float32 Tensor must be in the ` +
`range [0 - 1] but encountered ${value}.`);
}
}
else if (image.dtype === 'int32') {
if (value < 0 || value > 255) {
throw new Error(`Tensor values for a int32 Tensor must be in the ` +
`range [0 - 255] but encountered ${value}.`);
}
}
if (depth === 1) {
rgba[0] = value * multiplier;
rgba[1] = value * multiplier;
rgba[2] = value * multiplier;
}
else {
rgba[d] = value * multiplier;
}
}
const j = i * 4;
bytes[j + 0] = Math.round(rgba[0]);
bytes[j + 1] = Math.round(rgba[1]);
bytes[j + 2] = Math.round(rgba[2]);
bytes[j + 3] = Math.round(rgba[3]);
}
canvas.width = width;
canvas.height = height;
const imageData = new ImageData(bytes, width, height);
ctx.putImageData(imageData, 0, 0);
return image;
}
const drawConfig = {
kernelName: Draw,
backendName: 'cpu',
kernelFunc: draw
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sum(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
assertNotComplex(x, 'sum');
let $x;
if (x.dtype === 'bool') {
$x = cast$2({ inputs: { x }, backend, attrs: { dtype: 'int32' } });
}
else {
$x = identity$1({ inputs: { x }, backend });
}
const xRank = $x.shape.length;
const axes = parseAxisParam(axis, $x.shape);
const permutation = getAxesPermutation(axes, xRank);
let reductionAxes = axes;
let permutedX = $x;
if (permutation != null) {
permutedX =
transpose$1({ inputs: { x: $x }, backend, attrs: { perm: permutation } });
reductionAxes = getInnerMostAxes(reductionAxes.length, xRank);
}
assertAxesAreInnerMostDims('sum', reductionAxes, permutedX.shape.length);
const [outShape, reduceShape] = computeOutAndReduceShapes(permutedX.shape, reductionAxes);
const resultDtype = upcastType(permutedX.dtype, 'int32');
let result = zeros(backend, outShape, resultDtype);
const reduceSize = sizeFromShape(reduceShape);
const vals = backend.data.get(result.dataId).values;
const aVals = backend.data.get(permutedX.dataId).values;
for (let i = 0; i < vals.length; ++i) {
const offset = i * reduceSize;
let sum = 0;
for (let j = 0; j < reduceSize; ++j) {
sum += aVals[offset + j];
}
vals[i] = sum;
}
if (keepDims) {
const newShape = expandShapeToKeepDim(result.shape, axes);
const oldResult = result;
result = reshape({ inputs: { x: result }, backend, attrs: { shape: newShape } });
backend.disposeIntermediateTensorInfo(oldResult);
}
backend.disposeIntermediateTensorInfo($x);
if (permutation != null) {
backend.disposeIntermediateTensorInfo(permutedX);
}
return result;
}
const sumConfig = {
kernelName: Sum,
backendName: 'cpu',
kernelFunc: sum
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function einsum(args) {
const { inputs, backend, attrs } = args;
const { equation } = attrs;
const tensors = inputs;
const { allDims, summedDims, idDims } = decodeEinsumEquation(equation, tensors.length);
checkEinsumDimSizes(allDims.length, idDims, tensors);
const { path, steps } = getEinsumComputePath(summedDims, idDims);
const nSteps = steps.length;
let out = null;
let numDimsRemaining = allDims.length;
const tensorsToDispose = [];
for (let i = 0; i < nSteps; ++i) {
for (const idTerm of steps[i]) {
const { permutationIndices: perm, expandDims: dimsToExpand } = getEinsumPermutation(numDimsRemaining, idDims[idTerm]);
let x;
if (isIdentityPermutation(perm)) {
x = tensors[idTerm];
}
else {
x = transpose$1({ inputs: { x: tensors[idTerm] }, backend, attrs: { perm } });
tensorsToDispose.push(x);
}
const targetShape = x.shape.slice();
for (let k = 0; k < dimsToExpand.length; ++k) {
targetShape.splice(dimsToExpand[k], 0, 1);
}
if (!arraysEqual(x.shape, targetShape)) {
x = reshape({ inputs: { x }, backend, attrs: { shape: targetShape } });
tensorsToDispose.push(x);
}
if (out === null) {
out = x;
}
else {
// tslint:disable-next-line: no-unnecessary-type-assertion
out = multiply$1({ inputs: { a: x, b: out }, backend });
tensorsToDispose.push(out);
}
}
if (i < nSteps - 1) {
if (path[i] >= 0) {
out = sum({
inputs: { x: out },
backend,
attrs: {
axis: path[i] - (allDims.length - numDimsRemaining),
keepDims: false
}
});
tensorsToDispose.push(out);
}
numDimsRemaining--;
}
}
// Clean up intermediate tensors.
for (const tensorInfo of tensorsToDispose) {
if (tensorInfo === out) {
continue;
}
backend.disposeIntermediateTensorInfo(tensorInfo);
}
return out;
}
const einsumConfig = {
kernelName: Einsum,
backendName: 'cpu',
kernelFunc: einsum
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function eluGrad(args) {
const { inputs, backend } = args;
const { dy, y } = inputs;
assertNotComplex([dy, y], 'eluGrad');
const resultValues = new Float32Array(sizeFromShape(y.shape));
const values = backend.data.get(y.dataId).values;
const dyValues = backend.data.get(dy.dataId).values;
for (let i = 0; i < values.length; ++i) {
const v = values[i];
if (v >= 0) {
resultValues[i] = dyValues[i];
}
else {
resultValues[i] = dyValues[i] * (v + 1);
}
}
return backend.makeTensorInfo(y.shape, 'float32', resultValues);
}
const eluGradConfig$1 = {
kernelName: EluGrad,
backendName: 'cpu',
kernelFunc: eluGrad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const p = ERF_P;
const a1 = ERF_A1;
const a2 = ERF_A2;
const a3 = ERF_A3;
const a4 = ERF_A4;
const a5 = ERF_A5;
const erf = unaryKernelFunc$1(Erf, (xi) => {
const sign = Math.sign(xi);
const v = Math.abs(xi);
const t = 1.0 / (1.0 + p * v);
return sign *
(1.0 -
(((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t *
Math.exp(-v * v));
});
const erfConfig = {
kernelName: Erf,
backendName: 'cpu',
kernelFunc: erf,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function expandDims$1(args) {
const { inputs, backend, attrs } = args;
const { input } = inputs;
const { dim } = attrs;
const inputRank = input.shape.length;
const newShape = input.shape.slice();
let $dim = dim;
if (dim < 0) {
// Negative value is counted from the tail of rank.
assert$1(-(inputRank + 1) <= dim, () => `Axis must be in the interval [${-(inputRank + 1)}, ${inputRank}]`);
$dim = inputRank + dim + 1;
}
newShape.splice($dim, 0, 1);
return reshape({ inputs: { x: input }, backend, attrs: { shape: newShape } });
}
const expandDimsConfig = {
kernelName: ExpandDims,
backendName: 'cpu',
kernelFunc: expandDims$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const realDivImpl = createSimpleBinaryKernelImpl((a, b) => a / b);
const div = binaryKernelFunc$1(RealDiv, realDivImpl);
const realDivConfig = {
kernelName: RealDiv,
backendName: 'cpu',
kernelFunc: div
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Calculate FFT of inner most elements of batch tensor.
*/
function fftBatch(input, inverse, cpuBackend) {
const inputShape = input.shape;
const batch = inputShape[0];
const innerDim = inputShape[1];
const inputVals = cpuBackend.data.get(input.dataId);
const real2D = inputVals.complexTensorInfos.real;
const imag2D = inputVals.complexTensorInfos.imag;
// Collects real and imaginary values separately.
const resultShape = [batch, innerDim];
const resultSize = sizeFromShape(resultShape);
const resultReal = getTypedArrayFromDType('float32', resultSize);
const resultImag = getTypedArrayFromDType('float32', resultSize);
for (let b = 0; b < batch; b++) {
// TODO: Support slice ops for complex type.
const r = slice$1({
inputs: { x: real2D },
backend: cpuBackend,
attrs: { begin: [b, 0], size: [1, innerDim] }
});
const i = slice$1({
inputs: { x: imag2D },
backend: cpuBackend,
attrs: { begin: [b, 0], size: [1, innerDim] }
});
const input = complex$1({ inputs: { real: r, imag: i }, backend: cpuBackend });
// Run FFT by batch element.
const { real, imag } = fftImpl(input, inverse, cpuBackend);
const res = mergeRealAndImagArrays(real, imag);
for (let d = 0; d < innerDim; d++) {
const c = getComplexWithIndex(res, d);
resultReal[b * innerDim + d] = c.real;
resultImag[b * innerDim + d] = c.imag;
}
cpuBackend.disposeIntermediateTensorInfo(r);
cpuBackend.disposeIntermediateTensorInfo(i);
cpuBackend.disposeIntermediateTensorInfo(input);
}
const $realInfo = cpuBackend.makeTensorInfo(resultShape, 'float32', resultReal);
const $imagInfo = cpuBackend.makeTensorInfo(resultShape, 'float32', resultImag);
const result = complex$1({ inputs: { real: $realInfo, imag: $imagInfo }, backend: cpuBackend });
cpuBackend.disposeIntermediateTensorInfo($realInfo);
cpuBackend.disposeIntermediateTensorInfo($imagInfo);
return result;
}
function fftImpl(input, inverse, cpuBackend) {
const inputSize = sizeFromShape(input.shape);
const inputVals = cpuBackend.data.get(input.dataId);
const realVals = cpuBackend.data.get(inputVals.complexTensorInfos.real.dataId).values;
const imagVals = cpuBackend.data.get(inputVals.complexTensorInfos.imag.dataId).values;
if (isExponentOf2(inputSize)) {
const result = fftRadix2(realVals, imagVals, inputSize, inverse, cpuBackend);
const resultShape = [input.shape[0], input.shape[1]];
if (inverse) {
const realInfo = cpuBackend.makeTensorInfo(resultShape, 'float32', result.real);
const imagInfo = cpuBackend.makeTensorInfo(resultShape, 'float32', result.imag);
const sizeInfo = cpuBackend.makeTensorInfo([], 'float32', createScalarValue(inputSize, 'float32'));
const sizeInfoCopy = identity$1({ inputs: { x: sizeInfo }, backend: cpuBackend });
const divRealInfo = realDivConfig.kernelFunc({ inputs: { a: realInfo, b: sizeInfo }, backend: cpuBackend });
const divImagInfo = realDivConfig.kernelFunc({ inputs: { a: imagInfo, b: sizeInfoCopy }, backend: cpuBackend });
const divRealVals = cpuBackend.data.get(divRealInfo.dataId).values;
const divImagVals = cpuBackend.data.get(divImagInfo.dataId).values;
cpuBackend.disposeIntermediateTensorInfo(realInfo);
cpuBackend.disposeIntermediateTensorInfo(imagInfo);
cpuBackend.disposeIntermediateTensorInfo(sizeInfo);
cpuBackend.disposeIntermediateTensorInfo(sizeInfoCopy);
cpuBackend.disposeIntermediateTensorInfo(divRealInfo);
cpuBackend.disposeIntermediateTensorInfo(divImagInfo);
return { real: divRealVals, imag: divImagVals };
}
return result;
}
else {
const data = mergeRealAndImagArrays(realVals, imagVals);
const rawOutput = fourierTransformByMatmul(data, inputSize, inverse);
return splitRealAndImagArrays(rawOutput);
}
}
function isExponentOf2(size) {
return (size & size - 1) === 0;
}
// FFT using Cooley-Tukey algorithm on radix 2 dimensional input.
function fftRadix2(realVals, imagVals, size, inverse, cpuBackend) {
if (size === 1) {
return { real: realVals, imag: imagVals };
}
const data = mergeRealAndImagArrays(realVals, imagVals);
const half = size / 2;
const evenComplex = complexWithEvenIndex(data);
const evenRealVals = evenComplex.real;
const evenImagVals = evenComplex.imag;
const evenShape = [evenRealVals.length];
const evenRealInfo = cpuBackend.makeTensorInfo(evenShape, 'float32', evenRealVals);
const evenImagInfo = cpuBackend.makeTensorInfo(evenShape, 'float32', evenImagVals);
const evenTensorInfo = complex$1({ inputs: { real: evenRealInfo, imag: evenImagInfo }, backend: cpuBackend });
const oddComplex = complexWithOddIndex(data);
const oddRealVals = oddComplex.real;
const oddImagVals = oddComplex.imag;
const oddShape = [oddRealVals.length];
const oddRealInfo = cpuBackend.makeTensorInfo(oddShape, 'float32', oddRealVals);
const oddImagInfo = cpuBackend.makeTensorInfo(oddShape, 'float32', oddImagVals);
const oddTensorInfo = complex$1({ inputs: { real: oddRealInfo, imag: oddImagInfo }, backend: cpuBackend });
// Recursive call for half part of original input.
const $evenComplex = fftRadix2(evenRealVals, evenImagVals, half, inverse, cpuBackend);
const $evenRealVals = $evenComplex.real;
const $evenImagVals = $evenComplex.imag;
const $evenShape = [$evenRealVals.length];
const $evenRealInfo = cpuBackend.makeTensorInfo($evenShape, 'float32', $evenRealVals);
const $evenImagInfo = cpuBackend.makeTensorInfo($evenShape, 'float32', $evenImagVals);
const $evenTensorInfo = complex$1({
inputs: { real: $evenRealInfo, imag: $evenImagInfo },
backend: cpuBackend
});
const $oddComplex = fftRadix2(oddRealVals, oddImagVals, half, inverse, cpuBackend);
const $oddRealVals = $oddComplex.real;
const $oddImagVals = $oddComplex.imag;
const $oddShape = [$oddRealVals.length];
const $oddRealInfo = cpuBackend.makeTensorInfo($oddShape, 'float32', $oddRealVals);
const $oddImagInfo = cpuBackend.makeTensorInfo($oddShape, 'float32', $oddImagVals);
const $oddTensorInfo = complex$1({ inputs: { real: $oddRealInfo, imag: $oddImagInfo }, backend: cpuBackend });
const e = exponents(size, inverse);
const eShape = [e.real.length];
const eRealInfo = cpuBackend.makeTensorInfo(eShape, 'float32', e.real);
const eImagInfo = cpuBackend.makeTensorInfo(eShape, 'float32', e.imag);
const complexInfo = complex$1({ inputs: { real: eRealInfo, imag: eImagInfo }, backend: cpuBackend });
const exponentInfo = multiply$1({ inputs: { a: complexInfo, b: $oddTensorInfo }, backend: cpuBackend });
const addPart = add({
inputs: { a: $evenTensorInfo, b: exponentInfo },
backend: cpuBackend
});
const subPart = sub$1({
inputs: { a: $evenTensorInfo, b: exponentInfo },
backend: cpuBackend
});
const addPartReal = real$1({ inputs: { input: addPart }, backend: cpuBackend });
const subPartReal = real$1({ inputs: { input: subPart }, backend: cpuBackend });
const addPartImag = imag({ inputs: { input: addPart }, backend: cpuBackend });
const subPartImag = imag({ inputs: { input: subPart }, backend: cpuBackend });
const $real = concat({
inputs: [addPartReal, subPartReal],
backend: cpuBackend,
attrs: { axis: 0 }
});
const $imag = concat({
inputs: [addPartImag, subPartImag],
backend: cpuBackend,
attrs: { axis: 0 }
});
const $realVals = cpuBackend.data.get($real.dataId).values;
const $imagVals = cpuBackend.data.get($imag.dataId).values;
cpuBackend.disposeIntermediateTensorInfo(evenRealInfo);
cpuBackend.disposeIntermediateTensorInfo(evenImagInfo);
cpuBackend.disposeIntermediateTensorInfo(evenTensorInfo);
cpuBackend.disposeIntermediateTensorInfo(oddRealInfo);
cpuBackend.disposeIntermediateTensorInfo(oddImagInfo);
cpuBackend.disposeIntermediateTensorInfo(oddTensorInfo);
cpuBackend.disposeIntermediateTensorInfo($evenRealInfo);
cpuBackend.disposeIntermediateTensorInfo($evenImagInfo);
cpuBackend.disposeIntermediateTensorInfo($evenTensorInfo);
cpuBackend.disposeIntermediateTensorInfo($oddRealInfo);
cpuBackend.disposeIntermediateTensorInfo($oddImagInfo);
cpuBackend.disposeIntermediateTensorInfo($oddTensorInfo);
cpuBackend.disposeIntermediateTensorInfo(eRealInfo);
cpuBackend.disposeIntermediateTensorInfo(eImagInfo);
cpuBackend.disposeIntermediateTensorInfo(complexInfo);
cpuBackend.disposeIntermediateTensorInfo(exponentInfo);
cpuBackend.disposeIntermediateTensorInfo(addPart);
cpuBackend.disposeIntermediateTensorInfo(subPart);
cpuBackend.disposeIntermediateTensorInfo(addPartReal);
cpuBackend.disposeIntermediateTensorInfo(addPartImag);
cpuBackend.disposeIntermediateTensorInfo(subPartReal);
cpuBackend.disposeIntermediateTensorInfo(subPartImag);
cpuBackend.disposeIntermediateTensorInfo($real);
cpuBackend.disposeIntermediateTensorInfo($imag);
return { real: $realVals, imag: $imagVals };
}
// Calculate fourier transform by multplying sinusoid matrix.
function fourierTransformByMatmul(data, size, inverse) {
const ret = new Float32Array(size * 2);
// TODO: Use matmul instead once it supports complex64 type.
for (let r = 0; r < size; r++) {
let real = 0.0;
let imag = 0.0;
for (let c = 0; c < size; c++) {
const e = exponent(r * c, size, inverse);
const term = getComplexWithIndex(data, c);
real += term.real * e.real - term.imag * e.imag;
imag += term.real * e.imag + term.imag * e.real;
}
if (inverse) {
real /= size;
imag /= size;
}
assignToTypedArray(ret, real, imag, r);
}
return ret;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fft(args) {
const { inputs, backend } = args;
const { input } = inputs;
const inputSize = sizeFromShape(input.shape);
// Collapse all outer dimensions to a single batch dimension.
const innerDimensionSize = input.shape[input.shape.length - 1];
const batch = inputSize / innerDimensionSize;
const input2D = reshape({
inputs: { x: input },
backend,
attrs: { shape: [batch, innerDimensionSize] }
});
const result = fftBatch(input2D, false, backend);
const resultReshaped = reshape({ inputs: { x: result }, backend, attrs: { shape: input.shape } });
backend.disposeIntermediateTensorInfo(input2D);
backend.disposeIntermediateTensorInfo(result);
return resultReshaped;
}
const fftConfig = {
kernelName: FFT,
backendName: 'cpu',
kernelFunc: fft
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fill(args) {
const { backend, attrs } = args;
const { shape, value, dtype } = attrs;
const $dtype = dtype || inferDtype(value);
const values = getArrayFromDType($dtype, sizeFromShape(shape));
fillValues(values, value, $dtype);
return backend.makeTensorInfo(shape, $dtype, values);
}
const fillConfig = {
kernelName: Fill,
backendName: 'cpu',
kernelFunc: fill
};
function fillValues(values, value, dtype) {
if (dtype === 'string') {
values.fill(value);
}
else {
values.fill(value);
}
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const flipLeftRightConfig = {
kernelName: FlipLeftRight,
backendName: 'cpu',
kernelFunc: ({ inputs, attrs, backend }) => {
const { image } = inputs;
const cpuBackend = backend;
const output = getTypedArrayFromDType(image.dtype, sizeFromShape(image.shape));
const [batch, imageHeight, imageWidth, numChannels] = image.shape;
const imageVals = cpuBackend.data.get(image.dataId).values;
for (let batchIdx = 0; batchIdx < batch; batchIdx++) {
const batchOffset = batchIdx * imageWidth * imageHeight * numChannels;
for (let row = 0; row < imageHeight; row++) {
const rowOffset = row * (imageWidth * numChannels);
for (let col = 0; col < imageWidth; col++) {
const colOffset = col * numChannels;
for (let channel = 0; channel < numChannels; channel++) {
const coordX = Math.round(imageWidth - col - 1);
const outIdx = batchOffset + rowOffset + colOffset + channel;
let outputValue = imageVals[outIdx];
// If the coordinate position falls within the image boundaries...
if (coordX >= 0 && coordX < imageWidth) {
// set the output to the image value at the coordinate position.
const rotatedColOffset = coordX * numChannels;
const imageIdx = batchOffset + rowOffset + rotatedColOffset + channel;
outputValue = imageVals[imageIdx];
}
output[outIdx] = outputValue;
}
}
}
}
const dataId = cpuBackend.write(output, image.shape, image.dtype);
return { dataId, shape: image.shape, dtype: image.dtype };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fusedConv2D(args) {
const { inputs, backend, attrs } = args;
const { x, filter, bias, preluActivationWeights } = inputs;
const { strides, pad, dataFormat, dilations, dimRoundingMode, activation, leakyreluAlpha } = attrs;
let result = conv2D({
inputs: { x, filter },
backend,
attrs: { strides, pad, dataFormat, dilations, dimRoundingMode }
});
if (bias) {
const resultOld = result;
// For NCHW format, if bias is a 1-D tensor, it is supposed to be aligned
// to the channel of the conv2d's result; if the bias is a scalar, the
// bias_add is computed as if the bias was broadcasted to the shape of the
// conv2d's result.
if (dataFormat === 'NCHW' && bias.shape.length === 1 &&
bias.shape[0] !== 1) {
const reshapedBias = reshape({ inputs: { x: bias }, backend, attrs: { shape: [bias.shape[0], 1, 1] } });
result =
add({ inputs: { a: result, b: reshapedBias }, backend });
backend.disposeIntermediateTensorInfo(reshapedBias);
}
else {
// This condition handles NHWC and NCHW (scalar case). The only other case
// for NCHW (1D case) is handled above.
result = add({ inputs: { a: result, b: bias }, backend });
}
backend.disposeIntermediateTensorInfo(resultOld);
}
if (activation) {
const resultOld = result;
// For NCHW format, if PReLu activation weights is a 1-D tensor, it is
// supposed to be aligned with the channel of the conv2d's result. For other
// cases, whether NCHW or NHWC data format, the conv2d result is
// already aligned with the activation weights.
if (dataFormat === 'NCHW' && activation === 'prelu' &&
preluActivationWeights.shape.length === 1 &&
preluActivationWeights.shape[0] !== 1) {
const reshapedAlpha = reshape({
inputs: { x: preluActivationWeights },
backend,
attrs: { shape: [preluActivationWeights.shape[0], 1, 1] }
});
result = applyActivation(backend, result, activation, reshapedAlpha, leakyreluAlpha);
backend.disposeIntermediateTensorInfo(reshapedAlpha);
}
else {
result = applyActivation(backend, result, activation, preluActivationWeights, leakyreluAlpha);
}
backend.disposeIntermediateTensorInfo(resultOld);
}
return result;
}
const fusedConv2DConfig = {
kernelName: FusedConv2D,
backendName: 'cpu',
kernelFunc: fusedConv2D
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function fusedDepthwiseConv2D(args) {
const { inputs, backend, attrs } = args;
const { x, filter, bias, preluActivationWeights } = inputs;
const { strides, pad, dataFormat, dilations, dimRoundingMode, activation, leakyreluAlpha } = attrs;
let result = depthwiseConv2dNative({
inputs: { x, filter },
backend,
attrs: { strides, pad, dataFormat, dilations, dimRoundingMode }
});
if (bias) {
const oldResult = result;
result = add({ inputs: { a: result, b: bias }, backend });
backend.disposeIntermediateTensorInfo(oldResult);
}
if (activation) {
const oldResult = result;
result = applyActivation(backend, result, activation, preluActivationWeights, leakyreluAlpha);
backend.disposeIntermediateTensorInfo(oldResult);
}
return result;
}
const fusedDepthwiseConv2DConfig = {
kernelName: FusedDepthwiseConv2D,
backendName: 'cpu',
kernelFunc: fusedDepthwiseConv2D
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function gatherNd(args) {
const { inputs, backend } = args;
const { params, indices } = inputs;
const paramsSize = sizeFromShape(params.shape);
const indicesShape = indices.shape;
const sliceRank = indicesShape[indicesShape.length - 1];
const [resultShape, numSlices, sliceSize, strides] = prepareAndValidate(params, indices);
if (numSlices === 0) {
return backend.makeTensorInfo(resultShape, params.dtype, []);
}
const indicesData = backend.data.get(indices.dataId).values;
const paramsBuf = backend.bufferSync(params);
const outBuf = gatherNdImpl(indicesData, paramsBuf, params.dtype, numSlices, sliceRank, sliceSize, strides, params.shape, paramsSize);
return backend.makeTensorInfo(resultShape, params.dtype, outBuf.values);
}
const gatherNdConfig = {
kernelName: GatherNd,
backendName: 'cpu',
kernelFunc: gatherNd
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function gatherV2(args) {
const { inputs, backend, attrs } = args;
const { x, indices } = inputs;
const { axis, batchDims } = attrs;
assertNotComplex([x, indices], 'gatherV2');
// Throw error when any index is out of bound.
const parsedAxis = parseAxisParam(axis, x.shape)[0];
const indicesVals = backend.data.get(indices.dataId).values;
const axisDim = x.shape[parsedAxis];
for (let i = 0; i < indicesVals.length; ++i) {
const index = indicesVals[i];
assert$1(index <= axisDim - 1 && index >= 0, () => `GatherV2: the index value ${index} is not in [0, ${axisDim - 1}]`);
}
let $batchDims = batchDims;
if (batchDims == null) {
$batchDims = 0;
}
const indicesSize = sizeFromShape(indices.shape);
const shapeInfo = collectGatherOpShapeInfo(x, indices, parsedAxis, $batchDims);
const flattenX = reshape({
inputs: { x },
backend,
attrs: {
shape: [
shapeInfo.batchSize, shapeInfo.outerSize, shapeInfo.dimSize,
shapeInfo.sliceSize
]
}
});
const flattenIndex = reshape({
inputs: { x: indices },
backend,
attrs: { shape: [shapeInfo.batchSize, indicesSize / shapeInfo.batchSize] }
});
const flattenOutputShape = [
shapeInfo.batchSize, shapeInfo.outerSize, indicesSize / shapeInfo.batchSize,
shapeInfo.sliceSize
];
const indicesBuf = backend.bufferSync(flattenIndex);
const xBuf = backend.bufferSync(flattenX);
const outBuf = gatherV2Impl(xBuf, indicesBuf, flattenOutputShape);
backend.disposeIntermediateTensorInfo(flattenX);
backend.disposeIntermediateTensorInfo(flattenIndex);
return backend.makeTensorInfo(shapeInfo.outputShape, outBuf.dtype, outBuf.values);
}
const gatherV2Config = {
kernelName: GatherV2,
backendName: 'cpu',
kernelFunc: gatherV2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function ifft(args) {
const { inputs, backend } = args;
const { input } = inputs;
const inputSize = sizeFromShape(input.shape);
// Collapse all outer dimensions to a single batch dimension.
const innerDimensionSize = input.shape[input.shape.length - 1];
const batch = inputSize / innerDimensionSize;
const input2D = reshape({
inputs: { x: input },
backend,
attrs: { shape: [batch, innerDimensionSize] }
});
const result = fftBatch(input2D, true, backend);
const resultReshaped = reshape({ inputs: { x: result }, backend, attrs: { shape: input.shape } });
backend.disposeIntermediateTensorInfo(input2D);
backend.disposeIntermediateTensorInfo(result);
return resultReshaped;
}
const ifftConfig = {
kernelName: IFFT,
backendName: 'cpu',
kernelFunc: ifft
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const isFinite$1 = unaryKernelFunc$1(IsFinite, (xi) => Number.isFinite(xi) ? 1 : 0, 'bool');
const isFiniteConfig = {
kernelName: IsFinite,
backendName: 'cpu',
kernelFunc: isFinite$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const isInf = unaryKernelFunc$1(IsInf, (xi) => Math.abs(xi) === Infinity ? 1 : 0, 'bool');
const isInfConfig = {
kernelName: IsInf,
backendName: 'cpu',
kernelFunc: isInf,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const isNaN$1 = unaryKernelFunc$1(IsNan, (xi) => Number.isNaN(xi) ? 1 : 0, 'bool');
const isNaNConfig = {
kernelName: IsNan,
backendName: 'cpu',
kernelFunc: isNaN$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function linSpace(args) {
const { backend, attrs } = args;
const { start, stop, num } = attrs;
const outVals = linSpaceImpl(start, stop, num);
return backend.makeTensorInfo([outVals.length], 'float32', outVals);
}
const linSpaceConfig = {
kernelName: LinSpace,
backendName: 'cpu',
kernelFunc: linSpace
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const log1p = unaryKernelFunc$1(Log1p, (xi) => Math.log1p(xi));
const log1pConfig = {
kernelName: Log1p,
backendName: 'cpu',
kernelFunc: log1p,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const logicalAndImpl = createSimpleBinaryKernelImpl((a, b) => a && b);
const logicalAnd = binaryKernelFunc$1(LogicalAnd, logicalAndImpl, null /* complexImpl */, 'bool');
const logicalAndConfig = {
kernelName: LogicalAnd,
backendName: 'cpu',
kernelFunc: logicalAnd
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const logicalNot = unaryKernelFunc$1(LogicalNot, (xi) => xi ? 0 : 1, 'bool');
const logicalNotConfig = {
kernelName: LogicalNot,
backendName: 'cpu',
kernelFunc: logicalNot,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const logicalOrImpl = createSimpleBinaryKernelImpl((a, b) => a || b);
const logicalOr = binaryKernelFunc$1(LogicalOr, logicalOrImpl, null /* complexImpl */, 'bool');
const logicalOrConfig = {
kernelName: LogicalOr,
backendName: 'cpu',
kernelFunc: logicalOr
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function lRN(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { depthRadius, bias, alpha, beta } = attrs;
assertNotComplex(x, 'LRN');
const channels = x.shape[3];
const maxD = channels - 1;
const xValues = backend.data.get(x.dataId).values;
const size = sizeFromShape(x.shape);
const result = new Float32Array(size);
function sumAcrossChannels(offset) {
const currentChannel = offset % channels;
let beginSumOffset = offset - currentChannel + Math.max(0, currentChannel - depthRadius);
const endSumOffset = offset - currentChannel + Math.min(currentChannel + depthRadius, maxD);
let sum = 0.0;
for (; beginSumOffset <= endSumOffset; beginSumOffset++) {
const z = xValues[beginSumOffset];
sum += z * z;
}
return sum;
}
for (let offset = 0; offset < size; offset++) {
const sum = sumAcrossChannels(offset);
const val = xValues[offset] * Math.pow(bias + alpha * sum, -beta);
result[offset] = val;
}
return backend.makeTensorInfo(x.shape, x.dtype, result);
}
// tslint:disable-next-line: variable-name
const LRNConfig = {
kernelName: LRN,
backendName: 'cpu',
kernelFunc: lRN
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function lRNGrad(args) {
const { inputs, backend, attrs } = args;
const { x, y, dy } = inputs;
const { depthRadius, bias, alpha, beta } = attrs;
assertNotComplex(dy, 'LRNGrad');
const dySize = sizeFromShape(dy.shape);
const channels = dy.shape[3];
const dyValues = backend.data.get(dy.dataId).values;
const xValues = backend.data.get(x.dataId).values;
const yValues = backend.data.get(y.dataId).values;
const result = new Float32Array(dySize);
const size = dySize;
for (let offset = 0; offset < size; offset++) {
const currentChannel = offset % channels;
const depthBegin = (offset - currentChannel) + Math.max(0, currentChannel - depthRadius);
const depthEnd = (offset - currentChannel) +
Math.min(channels, currentChannel + depthRadius + 1);
let norm = 0;
for (let k = depthBegin; k < depthEnd; k++) {
norm += Math.pow(xValues[k], 2);
}
norm = alpha * norm + bias;
for (let k = depthBegin; k < depthEnd; k++) {
let dyi = -2 * alpha * beta * xValues[k] * yValues[offset] / norm;
if (offset === k) {
dyi += Math.pow(norm, -beta);
}
dyi *= dyValues[offset];
result[k] += dyi;
}
}
return backend.makeTensorInfo(dy.shape, x.dtype, result);
}
// tslint:disable-next-line: variable-name
const LRNGradConfig = {
kernelName: LRNGrad,
backendName: 'cpu',
kernelFunc: lRNGrad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function max(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { reductionIndices, keepDims } = attrs;
const cpuBackend = backend;
let xShape = x.shape;
const xRank = xShape.length;
const origAxes = parseAxisParam(reductionIndices, xShape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, xRank);
let xVals = cpuBackend.data.get(x.dataId).values;
if (permutedAxes != null) {
const newShape = new Array(xRank);
for (let i = 0; i < newShape.length; i++) {
newShape[i] = xShape[permutedAxes[i]];
}
xVals = transposeImpl$1(xVals, xShape, x.dtype, permutedAxes, newShape);
axes = getInnerMostAxes(axes.length, xRank);
xShape = newShape;
}
assertNotComplex(x, 'max');
assertAxesAreInnerMostDims('max', axes, xRank);
const [maxOutShape, reduceShape] = computeOutAndReduceShapes(xShape, axes);
const reduceSize = sizeFromShape(reduceShape);
const result = maxImpl$1(xVals, reduceSize, maxOutShape, x.dtype);
const dataId = cpuBackend.write(result, maxOutShape, x.dtype);
let outShape = maxOutShape;
if (keepDims) {
// reshape
const newShape = expandShapeToKeepDim(maxOutShape, origAxes);
outShape = newShape;
}
return { dataId, shape: outShape, dtype: x.dtype };
}
const maxConfig = {
kernelName: Max,
backendName: 'cpu',
kernelFunc: max
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPool(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
assertNotComplex(x, 'maxPool');
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const dilations = 1;
assert$1(eitherStridesOrDilationsAreOne(strides, dilations), () => 'Error in maxPool: Either strides or dilations must be 1. ' +
`Got strides ${strides} and dilations '${dilations}'`);
const convInfo = computePool2DInfo(x.shape, filterSize, strides, dilations, pad, dimRoundingMode);
let res;
if (convInfo.filterWidth === 1 && convInfo.filterHeight === 1 &&
arraysEqual(convInfo.inShape, convInfo.outShape)) {
res = identity$1({ inputs: { x }, backend });
}
else {
const xValues = backend.data.get(x.dataId).values;
const strides = computeStrides(x.shape);
const buffer = pool(xValues, x.shape, x.dtype, strides, convInfo, 'max');
res = backend.makeTensorInfo(convInfo.outShape, x.dtype, buffer.values);
}
return res;
}
const maxPoolConfig = {
kernelName: MaxPool,
backendName: 'cpu',
kernelFunc: maxPool
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPool3D(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { filterSize, strides, pad, dimRoundingMode, dataFormat } = attrs;
assertNotComplex(x, 'maxPool3d');
const convInfo = computePool3DInfo(x.shape, filterSize, strides, 1 /* dilations */, pad, dimRoundingMode, dataFormat);
const xValues = backend.data.get(x.dataId).values;
const outBuf = pool3d(xValues, x.shape, x.dtype, computeStrides(x.shape), convInfo, 'max');
return backend.makeTensorInfo(outBuf.shape, 'float32', outBuf.values);
}
const maxPool3DConfig = {
kernelName: MaxPool3D,
backendName: 'cpu',
kernelFunc: maxPool3D
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPool3DGrad(args) {
const { inputs, backend, attrs } = args;
const { dy, input } = inputs;
const { filterSize, strides, pad, dimRoundingMode } = attrs;
assertNotComplex([dy, input], 'maxPool3DGrad');
const convInfo = computePool3DInfo(input.shape, filterSize, strides, 1 /* dilations */, pad, dimRoundingMode);
const inputBuf = backend.bufferSync(input);
const maxPosBuf = maxPool3dPositions(inputBuf, convInfo);
const strideDepth = convInfo.strideDepth;
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationDepth = convInfo.dilationDepth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterDepth = convInfo.effectiveFilterDepth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padFront = effectiveFilterDepth - 1 - convInfo.padInfo.front;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const dx = buffer(input.shape, 'float32');
const dyBuf = backend.bufferSync(dy);
for (let batch = 0; batch < convInfo.batchSize; ++batch) {
for (let channel = 0; channel < convInfo.inChannels; ++channel) {
for (let dxDepth = 0; dxDepth < convInfo.inDepth; ++dxDepth) {
for (let dxRow = 0; dxRow < convInfo.inHeight; ++dxRow) {
for (let dxCol = 0; dxCol < convInfo.inWidth; ++dxCol) {
// Shader code begins
const dyDepthCorner = dxDepth - padFront;
const dyRowCorner = dxRow - padTop;
const dyColCorner = dxCol - padLeft;
let dotProd = 0;
for (let wDepth = 0; wDepth < effectiveFilterDepth; wDepth += dilationDepth) {
const dyDepth = (dyDepthCorner + wDepth) / strideDepth;
if (dyDepth < 0 || dyDepth >= convInfo.outDepth ||
Math.floor(dyDepth) !== dyDepth) {
continue;
}
for (let wRow = 0; wRow < effectiveFilterHeight; wRow += dilationHeight) {
const dyRow = (dyRowCorner + wRow) / strideHeight;
if (dyRow < 0 || dyRow >= convInfo.outHeight ||
Math.floor(dyRow) !== dyRow) {
continue;
}
for (let wCol = 0; wCol < effectiveFilterWidth; wCol += dilationWidth) {
const dyCol = (dyColCorner + wCol) / strideWidth;
if (dyCol < 0 || dyCol >= convInfo.outWidth ||
Math.floor(dyCol) !== dyCol) {
continue;
}
const maxPos = effectiveFilterDepth * effectiveFilterHeight *
effectiveFilterWidth -
1 -
maxPosBuf.get(batch, dyDepth, dyRow, dyCol, channel);
const curPos = wDepth * effectiveFilterHeight * effectiveFilterWidth +
wRow * effectiveFilterWidth + wCol;
const mask = maxPos === curPos ? 1 : 0;
if (mask === 0) {
continue;
}
const pixel = dyBuf.get(batch, dyDepth, dyRow, dyCol, channel);
dotProd += pixel * mask;
}
}
}
dx.set(dotProd, batch, dxDepth, dxRow, dxCol, channel);
}
}
}
}
}
return backend.makeTensorInfo(dx.shape, dx.dtype, dx.values);
}
const maxPool3DGradConfig$1 = {
kernelName: MaxPool3DGrad,
backendName: 'cpu',
kernelFunc: maxPool3DGrad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPoolGrad$1(args) {
const { inputs, backend, attrs } = args;
const { dy, input, output } = inputs;
const x = input;
assertNotComplex([input, output], 'maxPoolGrad');
const { filterSize, strides, pad, dimRoundingMode } = attrs;
const convInfo = computePool2DInfo(x.shape, filterSize, strides, 1 /* dilations */, pad, dimRoundingMode);
const xValues = backend.data.get(x.dataId).values;
const maxPosBuf = buffer(convInfo.outShape, x.dtype, maxPoolPositions(xValues, x.shape, x.dtype, convInfo).values);
const strideHeight = convInfo.strideHeight;
const strideWidth = convInfo.strideWidth;
const dilationHeight = convInfo.dilationHeight;
const dilationWidth = convInfo.dilationWidth;
const effectiveFilterHeight = convInfo.effectiveFilterHeight;
const effectiveFilterWidth = convInfo.effectiveFilterWidth;
const padLeft = effectiveFilterWidth - 1 - convInfo.padInfo.left;
const padTop = effectiveFilterHeight - 1 - convInfo.padInfo.top;
const dx = buffer(x.shape, 'float32');
const dyData = backend.data.get(dy.dataId).values;
const dyBuf = buffer(dy.shape, 'float32', dyData);
for (let b = 0; b < convInfo.batchSize; ++b) {
for (let d = 0; d < convInfo.inChannels; ++d) {
for (let dxR = 0; dxR < convInfo.inHeight; ++dxR) {
for (let dxC = 0; dxC < convInfo.inWidth; ++dxC) {
// Shader code begins.
const dyRCorner = dxR - padTop;
const dyCCorner = dxC - padLeft;
let dotProd = 0;
for (let wR = 0; wR < effectiveFilterHeight; wR += dilationHeight) {
const dyR = (dyRCorner + wR) / strideHeight;
if (dyR < 0 || dyR >= convInfo.outHeight ||
Math.floor(dyR) !== dyR) {
continue;
}
for (let wC = 0; wC < effectiveFilterWidth; wC += dilationWidth) {
const dyC = (dyCCorner + wC) / strideWidth;
if (dyC < 0 || dyC >= convInfo.outWidth ||
Math.floor(dyC) !== dyC) {
continue;
}
const maxPos = effectiveFilterHeight * effectiveFilterWidth - 1 -
maxPosBuf.get(b, dyR, dyC, d);
const curPos = wR * effectiveFilterWidth + wC;
const mask = maxPos === curPos ? 1 : 0;
if (mask === 0) {
continue;
}
const pixel = dyBuf.get(b, dyR, dyC, d);
dotProd += pixel * mask;
}
}
dx.set(dotProd, b, dxR, dxC, d);
}
}
}
}
return backend.makeTensorInfo(dx.shape, dx.dtype, dx.values);
}
const maxPoolGradConfig$1 = {
kernelName: MaxPoolGrad,
backendName: 'cpu',
kernelFunc: maxPoolGrad$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function maxPoolWithArgmaxImpl(xValues, xShape, dtype, includeBatchInIndex, convInfo) {
const strides = computeStrides(xShape);
const maxPools = pool(xValues, xShape, dtype, strides, convInfo, 'max');
const maxPositions = maxPoolPositions(xValues, xShape, dtype, convInfo, true, includeBatchInIndex);
return [maxPools.values, maxPositions.values];
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const maxPoolWithArgmaxConfig = {
kernelName: MaxPoolWithArgmax,
backendName: 'cpu',
kernelFunc: ({ inputs, attrs, backend }) => {
const { x } = inputs;
const { filterSize, strides, pad, includeBatchInIndex } = attrs;
const cpuBackend = backend;
assertNotComplex(x, 'MaxPoolWithArgmax');
const values = cpuBackend.data.get(x.dataId).values;
const convInfo = computePool2DInfo(x.shape, filterSize, strides, [1, 1], pad);
const [pooled, indexes] = maxPoolWithArgmaxImpl(values, x.shape, x.dtype, includeBatchInIndex, convInfo);
const pooledDataId = cpuBackend.write(pooled, convInfo.outShape, x.dtype);
const indexesDataId = cpuBackend.write(indexes, convInfo.outShape, x.dtype);
return [
{ dataId: pooledDataId, shape: convInfo.outShape, dtype: x.dtype },
{ dataId: indexesDataId, shape: convInfo.outShape, dtype: 'int32' }
];
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function mean(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
const axes = parseAxisParam(axis, x.shape);
const shapes = computeOutAndReduceShapes(x.shape, axes);
const reduceShape = shapes[1];
const reduceSize = sizeFromShape(reduceShape);
const toDispose = [];
const reduceSizeScalar = backend.makeTensorInfo([], 'float32', new Float32Array([reduceSize]));
toDispose.push(reduceSizeScalar);
const $x = cast$2({ inputs: { x }, backend, attrs: { dtype: 'float32' } });
toDispose.push($x);
const res = div({ inputs: { a: $x, b: reduceSizeScalar }, backend });
toDispose.push(res);
const result = sum({ inputs: { x: res }, backend, attrs: { axis, keepDims } });
toDispose.forEach(t => backend.disposeIntermediateTensorInfo(t));
return result;
}
const meanConfig = {
kernelName: Mean,
backendName: 'cpu',
kernelFunc: mean
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function min(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { axis, keepDims } = attrs;
assertNotComplex(x, 'min');
const origAxes = parseAxisParam(axis, x.shape);
let axes = origAxes;
const permutedAxes = getAxesPermutation(axes, x.shape.length);
let $x = x;
if (permutedAxes != null) {
$x = transpose$1({ inputs: { x }, backend, attrs: { perm: permutedAxes } });
axes = getInnerMostAxes(axes.length, x.shape.length);
}
assertAxesAreInnerMostDims('min', axes, $x.shape.length);
const [outShape, reduceShape] = computeOutAndReduceShapes($x.shape, axes);
const reduceSize = sizeFromShape(reduceShape);
const vals = makeZerosTypedArray(sizeFromShape(outShape), $x.dtype);
const aVals = backend.data.get($x.dataId).values;
for (let i = 0; i < vals.length; ++i) {
const offset = i * reduceSize;
let min = aVals[offset];
for (let j = 0; j < reduceSize; ++j) {
const value = aVals[offset + j];
if (Number.isNaN(value) ||
value < min) { // comparison with NaN always return false
min = value;
}
}
vals[i] = min;
}
if (permutedAxes != null) {
backend.disposeIntermediateTensorInfo($x);
}
const result = backend.makeTensorInfo(outShape, $x.dtype, vals);
if (keepDims) {
const expandedShape = expandShapeToKeepDim(outShape, origAxes);
const reshapedResult = reshape({ inputs: { x: result }, backend, attrs: { shape: expandedShape } });
backend.disposeIntermediateTensorInfo(result);
return reshapedResult;
}
return result;
}
const minConfig = {
kernelName: Min,
backendName: 'cpu',
kernelFunc: min
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function mirrorPad(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { paddings, mode } = attrs;
assertNotComplex(x, 'mirrorPad');
const outShape = paddings.map((p, i) => p[0] /* beforePad */ + x.shape[i] + p[1] /* afterPad */);
const start = paddings.map(p => p[0]);
const end = paddings.map((p, i) => p[0] + x.shape[i]);
const offset = mode === 'reflect' ? 0 : 1;
const xVals = backend.data.get(x.dataId).values;
const xRank = x.shape.length;
const xStrides = computeStrides(x.shape);
const resultSize = sizeFromShape(outShape);
const resultRank = outShape.length;
const resultStrides = computeStrides(outShape);
const resVals = getTypedArrayFromDType(x.dtype, resultSize);
for (let i = 0; i < resultSize; i++) {
let coords = indexToLoc(i, resultRank, resultStrides);
for (let i = 0; i < resultRank; i++) {
if (coords[i] < start[i]) {
coords[i] = start[i] * 2 - coords[i] - offset;
}
else if (coords[i] >= end[i]) {
coords[i] = (end[i] - 1) * 2 - coords[i] + offset;
}
}
coords = coords.map((c, i) => c - start[i]);
const inIndex = locToIndex(coords, xRank, xStrides);
resVals[i] = xVals[inIndex];
}
const outId = backend.write(resVals, outShape, x.dtype);
return { dataId: outId, shape: outShape, dtype: x.dtype };
}
const mirrorPadConfig = {
kernelName: MirrorPad,
backendName: 'cpu',
kernelFunc: mirrorPad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const modImpl = createSimpleBinaryKernelImpl(((aValue, bValue) => {
const rem = aValue % bValue;
if ((aValue < 0 && bValue < 0) || (aValue >= 0 && bValue >= 0)) {
return rem;
}
else {
return (rem + bValue) % bValue;
}
}));
const mod = binaryKernelFunc$1(Mod, modImpl);
const modConfig = {
kernelName: Mod,
backendName: 'cpu',
kernelFunc: mod
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function softmax(args) {
const { inputs, backend, attrs } = args;
const { logits } = inputs;
const { dim } = attrs;
const logitsRank = logits.shape.length;
let $dim = dim;
if ($dim === -1) {
$dim = logitsRank - 1;
}
if ($dim !== logitsRank - 1) {
throw Error('Softmax along a non-last dimension is not yet supported. ' +
`Logits was rank ${logitsRank} and dim was ${$dim}`);
}
const axes = parseAxisParam([$dim], logits.shape);
const maxLogit = max({
inputs: { x: logits },
backend,
attrs: { reductionIndices: axes, keepDims: false }
});
const expandedShape = expandShapeToKeepDim(maxLogit.shape, axes);
const maxLogitReshaped = reshape({ inputs: { x: maxLogit }, backend, attrs: { shape: expandedShape } });
const a = sub$1({ inputs: { a: logits, b: maxLogitReshaped }, backend });
const b = exp$1({ inputs: { x: a }, backend });
const sumExp = sum({ inputs: { x: b }, backend, attrs: { axis: axes, keepDims: false } });
const sumReshaped = reshape({ inputs: { x: sumExp }, backend, attrs: { shape: expandedShape } });
const result = div({ inputs: { a: b, b: sumReshaped }, backend });
backend.disposeIntermediateTensorInfo(maxLogit);
backend.disposeIntermediateTensorInfo(maxLogitReshaped);
backend.disposeIntermediateTensorInfo(a);
backend.disposeIntermediateTensorInfo(b);
backend.disposeIntermediateTensorInfo(sumExp);
backend.disposeIntermediateTensorInfo(sumReshaped);
return result;
}
const softmaxConfig = {
kernelName: Softmax$1,
backendName: 'cpu',
kernelFunc: softmax
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function multinomial(args) {
const { inputs, backend, attrs } = args;
const { logits } = inputs;
const { numSamples, seed, normalized } = attrs;
assertNotComplex(logits, 'multinomial');
const probabilities = normalized ?
logits :
softmax({ inputs: { logits }, backend, attrs: { dim: -1 } });
const batchSize = probabilities.shape[0];
const numEvents = probabilities.shape[1];
const probVals = backend.data.get(probabilities.dataId).values;
const resShape = [batchSize, numSamples];
const resVals = makeZerosTypedArray(sizeFromShape(resShape), 'int32');
for (let b = 0; b < batchSize; ++b) {
const offset = b * numEvents;
// The cdf won't include the last event. It will be implicit if no other
// event happened.
const cdf = new Float32Array(numEvents - 1);
cdf[0] = probVals[offset];
for (let event = 1; event < cdf.length; ++event) {
cdf[event] = cdf[event - 1] + probVals[offset + event];
}
const random = seedrandom.alea(seed.toString());
const outOffset = b * numSamples;
for (let sampleId = 0; sampleId < numSamples; ++sampleId) {
const r = random();
// Assume last event happened by default.
resVals[outOffset + sampleId] = cdf.length;
for (let event = 0; event < cdf.length; event++) {
if (r < cdf[event]) {
resVals[outOffset + sampleId] = event;
break;
}
}
}
}
if (!normalized) {
backend.disposeIntermediateTensorInfo(probabilities);
}
return backend.makeTensorInfo(resShape, 'int32', resVals);
}
const multinomialConfig = {
kernelName: Multinomial,
backendName: 'cpu',
kernelFunc: multinomial
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const nonMaxSuppressionV3Impl = nonMaxSuppressionV3Impl$2;
function nonMaxSuppressionV3(args) {
const { inputs, backend, attrs } = args;
const { boxes, scores } = inputs;
const { maxOutputSize, iouThreshold, scoreThreshold } = attrs;
assertNotComplex(boxes, 'NonMaxSuppression');
const boxesVals = backend.data.get(boxes.dataId).values;
const scoresVals = backend.data.get(scores.dataId).values;
const { selectedIndices } = nonMaxSuppressionV3Impl(boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold);
return backend.makeTensorInfo([selectedIndices.length], 'int32', new Int32Array(selectedIndices));
}
const nonMaxSuppressionV3Config = {
kernelName: NonMaxSuppressionV3,
backendName: 'cpu',
kernelFunc: nonMaxSuppressionV3
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const nonMaxSuppressionV4Impl = nonMaxSuppressionV4Impl$2;
function nonMaxSuppressionV4(args) {
const { inputs, backend, attrs } = args;
const { boxes, scores } = inputs;
const { maxOutputSize, iouThreshold, scoreThreshold, padToMaxOutputSize } = attrs;
assertNotComplex(boxes, 'NonMaxSuppressionPadded');
const boxesVals = backend.data.get(boxes.dataId).values;
const scoresVals = backend.data.get(scores.dataId).values;
const { selectedIndices, validOutputs } = nonMaxSuppressionV4Impl(boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold, padToMaxOutputSize);
return [
backend.makeTensorInfo([selectedIndices.length], 'int32', new Int32Array(selectedIndices)),
backend.makeTensorInfo([], 'int32', new Int32Array([validOutputs]))
];
}
const nonMaxSuppressionV4Config = {
kernelName: NonMaxSuppressionV4,
backendName: 'cpu',
kernelFunc: nonMaxSuppressionV4
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const nonMaxSuppressionV5Impl = nonMaxSuppressionV5Impl$2;
function nonMaxSuppressionV5(args) {
const { inputs, backend, attrs } = args;
const { boxes, scores } = inputs;
const { maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma } = attrs;
assertNotComplex(boxes, 'NonMaxSuppressionWithScore');
const boxesVals = backend.data.get(boxes.dataId).values;
const scoresVals = backend.data.get(scores.dataId).values;
const maxOutputSizeVal = maxOutputSize;
const iouThresholdVal = iouThreshold;
const scoreThresholdVal = scoreThreshold;
const softNmsSigmaVal = softNmsSigma;
const { selectedIndices, selectedScores } = nonMaxSuppressionV5Impl(boxesVals, scoresVals, maxOutputSizeVal, iouThresholdVal, scoreThresholdVal, softNmsSigmaVal);
return [
backend.makeTensorInfo([selectedIndices.length], 'int32', new Int32Array(selectedIndices)),
backend.makeTensorInfo([selectedScores.length], 'float32', new Float32Array(selectedScores))
];
}
const nonMaxSuppressionV5Config = {
kernelName: NonMaxSuppressionV5,
backendName: 'cpu',
kernelFunc: nonMaxSuppressionV5
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function oneHot(args) {
const { inputs, backend, attrs } = args;
const { indices } = inputs;
const { dtype, depth, onValue, offValue } = attrs;
assertNotComplex(indices, 'oneHot');
const indicesSize = sizeFromShape(indices.shape);
const res = new Float32Array(indicesSize * depth);
res.fill(offValue);
const indicesVal = backend.data.get(indices.dataId).values;
for (let event = 0; event < indicesSize; ++event) {
if (indicesVal[event] >= 0 && indicesVal[event] < depth) {
res[event * depth + indicesVal[event]] = onValue;
}
}
return backend.makeTensorInfo([...indices.shape, depth], dtype, res);
}
const oneHotConfig = {
kernelName: OneHot,
backendName: 'cpu',
kernelFunc: oneHot
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function zerosLike(args) {
const { inputs, backend } = args;
const { x } = inputs;
if (x.dtype === 'string') {
throw new Error('zerosLike is not supported for string tensors');
}
else if (x.dtype === 'complex64') {
const realPart = real$1({ inputs: { input: x }, backend });
const r = zerosLike({ inputs: { x: realPart }, backend });
const imagPart = imag({ inputs: { input: x }, backend });
const i = zerosLike({ inputs: { x: imagPart }, backend });
const result = complex$1({ inputs: { real: r, imag: i }, backend });
backend.disposeIntermediateTensorInfo(realPart);
backend.disposeIntermediateTensorInfo(r);
backend.disposeIntermediateTensorInfo(imagPart);
backend.disposeIntermediateTensorInfo(i);
return result;
}
else {
return fill({ backend, attrs: { shape: x.shape, value: 0, dtype: x.dtype } });
}
}
const zerosLikeConfig = {
kernelName: ZerosLike,
backendName: 'cpu',
kernelFunc: zerosLike
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function onesLike(args) {
const { inputs, backend } = args;
const { x } = inputs;
if (x.dtype === 'string') {
throw new Error('onesLike is not supported for string tensors');
}
else if (x.dtype === 'complex64') {
const realPart = real$1({ inputs: { input: x }, backend });
const r = onesLike({ inputs: { x: realPart }, backend });
const imagPart = imag({ inputs: { input: x }, backend });
const i = zerosLike({ inputs: { x: imagPart }, backend });
const result = complex$1({ inputs: { real: r, imag: i }, backend });
backend.disposeIntermediateTensorInfo(realPart);
backend.disposeIntermediateTensorInfo(r);
backend.disposeIntermediateTensorInfo(imagPart);
backend.disposeIntermediateTensorInfo(i);
return result;
}
else {
return fill({ backend, attrs: { shape: x.shape, value: 1, dtype: x.dtype } });
}
}
const onesLikeConfig = {
kernelName: OnesLike,
backendName: 'cpu',
kernelFunc: onesLike
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function pack(args) {
const { inputs, backend, attrs } = args;
const { axis } = attrs;
if (inputs.length === 1) {
return expandDims$1({ inputs: { input: inputs[0] }, backend, attrs: { dim: axis } });
}
const shape = inputs[0].shape;
const dtype = inputs[0].dtype;
inputs.forEach(t => {
assertShapesMatch(shape, t.shape, 'All tensors passed to stack must have matching shapes');
assert$1(dtype === t.dtype, () => 'All tensors passed to stack must have matching dtypes');
});
const intermediateTensorInfos = [];
const expandedTensors = inputs.map(t => {
const expandedT = expandDims$1({ inputs: { input: t }, backend, attrs: { dim: axis } });
intermediateTensorInfos.push(expandedT);
return expandedT;
});
const result = concat({ inputs: expandedTensors, backend, attrs: { axis } });
intermediateTensorInfos.forEach(t => backend.disposeIntermediateTensorInfo(t));
return result;
}
const packConfig = {
kernelName: Pack,
backendName: 'cpu',
kernelFunc: pack
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function padV2(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { paddings, constantValue } = attrs;
assertNotComplex(x, 'pad');
const outShape = paddings.map((p, i) => p[0] /* beforePad */ + x.shape[i] + p[1] /* afterPad */);
const start = paddings.map(p => p[0]);
const xVals = backend.data.get(x.dataId).values;
const xSize = sizeFromShape(x.shape);
const xRank = x.shape.length;
const xStrides = computeStrides(x.shape);
const resultSize = sizeFromShape(outShape);
const resultRank = outShape.length;
const resultStrides = computeStrides(outShape);
const resVals = getTypedArrayFromDType(x.dtype, resultSize);
if (constantValue !== 0) {
resVals.fill(constantValue);
}
for (let i = 0; i < xSize; i++) {
const coords = indexToLoc(i, xRank, xStrides);
const outCoords = coords.map((c, i) => c + start[i]);
const outIndex = locToIndex(outCoords, resultRank, resultStrides);
resVals[outIndex] = xVals[i];
}
const outId = backend.write(resVals, outShape, x.dtype);
return { dataId: outId, shape: outShape, dtype: x.dtype };
}
const padV2Config = {
kernelName: PadV2,
backendName: 'cpu',
kernelFunc: padV2
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const powImpl = createSimpleBinaryKernelImpl((a, b) => Math.pow(a, b));
const pow = binaryKernelFunc$1(Pow, powImpl);
const powConfig = {
kernelName: Pow,
backendName: 'cpu',
kernelFunc: pow
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function raggedGather(args) {
const { inputs, backend} = args;
const { paramsNestedSplits, paramsDenseValues, indices } = inputs;
const $paramsNestedSplits = paramsNestedSplits.map(t => backend.data.get(t.dataId).values);
const $paramsNestedSplitsShapes = paramsNestedSplits.map(t => t.shape);
const $paramsDenseValues = backend.data.get(paramsDenseValues.dataId).values;
const $indices = backend.data.get(indices.dataId).values;
const [outputNestedSplits, outputDenseValues, outputDenseValuesShape] = raggedGatherImpl($paramsNestedSplits, $paramsNestedSplitsShapes, $paramsDenseValues, paramsDenseValues.shape, paramsDenseValues.dtype, $indices, indices.shape);
const outputNestedSplitsTensors = outputNestedSplits.map((splits) => backend.makeTensorInfo([splits.length], 'int32', splits));
const outputDenseValuesTensor = backend.makeTensorInfo(outputDenseValuesShape, paramsDenseValues.dtype, outputDenseValues);
return outputNestedSplitsTensors.concat([outputDenseValuesTensor]);
}
const raggedGatherConfig = {
kernelName: RaggedGather,
backendName: 'cpu',
kernelFunc: raggedGather,
};
/**
* @license
* Copyright 2022 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function raggedRange(args) {
const { inputs, backend } = args;
const { starts, limits, deltas } = inputs;
const $starts = backend.data.get(starts.dataId).values;
const $limits = backend.data.get(limits.dataId).values;
const $deltas = backend.data.get(deltas.dataId).values;
const [rtNestedSplitsData, rtDenseValuesData] = raggedRangeImpl($starts, starts.shape, starts.dtype, $limits, limits.shape, $deltas, deltas.shape);
const rtNestedSplits = backend.makeTensorInfo([rtNestedSplitsData.length], 'int32', rtNestedSplitsData);
const rtDenseValues = backend.makeTensorInfo([rtDenseValuesData.length], starts.dtype, rtDenseValuesData);
return [rtNestedSplits, rtDenseValues];
}
const raggedRangeConfig = {
kernelName: RaggedRange,
backendName: 'cpu',
kernelFunc: raggedRange,
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function raggedTensorToTensor(args) {
const { inputs, backend, attrs } = args;
const { shape, values, defaultValue, rowPartitionTensors } = inputs;
const { rowPartitionTypes } = attrs;
const $shape = backend.data.get(shape.dataId).values;
const $values = backend.data.get(values.dataId).values;
const $defaultValue = backend.data.get(defaultValue.dataId).values;
const $rowPartitionValues = rowPartitionTensors.map(t => backend.data.get(t.dataId).values);
const rowPartitionValuesShapes = rowPartitionTensors.map(t => t.shape);
const [outputShape, output] = raggedTensorToTensorImpl($shape, shape.shape, $values, values.shape, values.dtype, $defaultValue, defaultValue.shape, $rowPartitionValues, rowPartitionValuesShapes, rowPartitionTypes);
return backend.makeTensorInfo(outputShape, values.dtype, output);
}
const raggedTensorToTensorConfig = {
kernelName: RaggedTensorToTensor,
backendName: 'cpu',
kernelFunc: raggedTensorToTensor,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function range$1(args) {
const { backend, attrs } = args;
const { start, stop, dtype, step } = attrs;
const values = rangeImpl(start, stop, step, dtype);
return backend.makeTensorInfo([values.length], dtype, values);
}
const rangeConfig = {
kernelName: Range,
backendName: 'cpu',
kernelFunc: range$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const reciprocal = unaryKernelFunc$1(Reciprocal, (xi) => 1 / xi);
const reciprocalConfig = {
kernelName: Reciprocal,
backendName: 'cpu',
kernelFunc: reciprocal,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeBilinear(args) {
const { inputs, backend, attrs } = args;
const { images } = inputs;
const { alignCorners, halfPixelCenters, size } = attrs;
assertNotComplex(images, 'resizeBilinear');
const imagesStrides = computeStrides(images.shape);
const [newHeight, newWidth] = size;
const [batch, oldHeight, oldWidth, numChannels] = images.shape;
const xValues = backend.data.get(images.dataId).values;
const result = new Float32Array(sizeFromShape([batch, newHeight, newWidth, numChannels]));
const effectiveInputSize = [
(alignCorners && newHeight > 1) ? oldHeight - 1 : oldHeight,
(alignCorners && newWidth > 1) ? oldWidth - 1 : oldWidth
];
const effectiveOutputSize = [
(alignCorners && newHeight > 1) ? newHeight - 1 : newHeight,
(alignCorners && newWidth > 1) ? newWidth - 1 : newWidth
];
let outputIdx = 0;
const effectiveRowSizeRatio = effectiveInputSize[0] / effectiveOutputSize[0];
const effectiveColSizeRatio = effectiveInputSize[1] / effectiveOutputSize[1];
for (let b = 0; b < batch; b++) {
for (let r = 0; r < newHeight; r++) {
let sourceFracRow;
if (halfPixelCenters) {
sourceFracRow = effectiveRowSizeRatio * (r + 0.5) - 0.5;
}
else {
sourceFracRow = effectiveRowSizeRatio * r;
}
const sourceRowFloor = Math.max(0, Math.floor(sourceFracRow));
const rowFrac = sourceFracRow - sourceRowFloor;
const sourceRowCeil = Math.min(oldHeight - 1, Math.ceil(sourceFracRow));
const topRowOffset = b * imagesStrides[0] + sourceRowFloor * imagesStrides[1];
const botRowOffset = b * imagesStrides[0] + sourceRowCeil * imagesStrides[1];
for (let c = 0; c < newWidth; c++) {
let sourceFracCol;
if (halfPixelCenters) {
sourceFracCol = effectiveColSizeRatio * (c + 0.5) - 0.5;
}
else {
sourceFracCol = effectiveColSizeRatio * c;
}
const sourceColFloor = Math.max(0, Math.floor(sourceFracCol));
const colFrac = sourceFracCol - sourceColFloor;
const sourceColCeil = Math.min(oldWidth - 1, Math.ceil(sourceFracCol));
const topLeftOffest = topRowOffset + sourceColFloor * imagesStrides[2];
const botLeftOffset = botRowOffset + sourceColFloor * imagesStrides[2];
const topRightOffset = topRowOffset + sourceColCeil * imagesStrides[2];
const botRightOffest = botRowOffset + sourceColCeil * imagesStrides[2];
for (let d = 0; d < numChannels; d++) {
// Begin shader.
// Compute the fractional index of the source.
const topLeft = xValues[topLeftOffest + d];
const bottomLeft = xValues[botLeftOffset + d];
const topRight = xValues[topRightOffset + d];
const bottomRight = xValues[botRightOffest + d];
const top = topLeft + (topRight - topLeft) * colFrac;
const bottom = bottomLeft + (bottomRight - bottomLeft) * colFrac;
const newValue = top + (bottom - top) * rowFrac;
result[outputIdx++] = newValue;
}
}
}
}
return backend.makeTensorInfo([batch, newHeight, newWidth, numChannels], 'float32', result);
}
const resizeBilinearConfig = {
kernelName: ResizeBilinear,
backendName: 'cpu',
kernelFunc: resizeBilinear
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeBilinearGrad(args) {
const { inputs, backend, attrs } = args;
const { images, dy } = inputs;
const { alignCorners } = attrs;
assertNotComplex([dy, images], 'resizeBilinearGrad');
const imagesStrides = computeStrides(images.shape);
const [batch, xHeight, xWidth, depth] = images.shape;
const [, yHeight, yWidth] = dy.shape;
const output = new Float32Array(batch * xHeight * xWidth * depth);
// In the backwards pass, we want to find the pixels that were generated
// for each pixel in the input image the forward pass and add the
// corresponding coefficient from dy to the gradient (with some
// interpolation).
const effectiveXSize = [
(alignCorners && yHeight > 1) ? xHeight - 1 : xHeight,
(alignCorners && yWidth > 1) ? xWidth - 1 : xWidth
];
const effectiveYSize = [
(alignCorners && yHeight > 1) ? yHeight - 1 : yHeight,
(alignCorners && yWidth > 1) ? yWidth - 1 : yWidth
];
const heightScale = effectiveXSize[0] / effectiveYSize[0];
const widthScale = effectiveXSize[1] / effectiveYSize[1];
// Reference implementation
// tslint:disable-next-line:max-line-length
// https://github.com/tensorflow/tensorflow/blob/3039375c86a5bbc9610c7725dcaa95d635f87ba2/tensorflow/core/kernels/resize_bilinear_op.cc#L275
const dyValues = backend.data.get(dy.dataId).values;
let offset = 0;
for (let b = 0; b < batch; b++) {
const bOffset = b * imagesStrides[0];
for (let r = 0; r < yHeight; r++) {
const dxR = r * heightScale;
const topDxRIndex = Math.floor(dxR);
const bottomDxRIndex = Math.min(Math.ceil(dxR), xHeight - 1);
const topDxROffset = bOffset + topDxRIndex * imagesStrides[1];
const bottomDxROffset = bOffset + bottomDxRIndex * imagesStrides[1];
const dxRLerp = dxR - topDxRIndex;
const inverseDxRLerp = 1.0 - dxRLerp;
for (let c = 0; c < yWidth; c++) {
const dxC = c * widthScale;
const leftDxCIndex = Math.floor(dxC);
const rightDxCIndex = Math.min(Math.ceil(dxC), xWidth - 1);
const dxCLerp = dxC - leftDxCIndex;
const inverseDxCLerp = 1.0 - dxCLerp;
const topLeftRCOffset = topDxROffset + leftDxCIndex * imagesStrides[2];
const topRightRCOffset = topDxROffset + rightDxCIndex * imagesStrides[2];
const bottomLeftRCOffset = bottomDxROffset + leftDxCIndex * imagesStrides[2];
const bottomRightRCOffset = bottomDxROffset + rightDxCIndex * imagesStrides[2];
const inverseDxRLerpTimesInverseDxCLerp = inverseDxRLerp * inverseDxCLerp;
const inverseDxRLerpTimesDxCLerp = inverseDxRLerp * dxCLerp;
const dxRLerpTimesInverseDxCLerp = dxRLerp * inverseDxCLerp;
const dxRLerpTimesDxCLerp = dxRLerp * dxCLerp;
for (let d = 0; d < depth; d++) {
const dyVal = dyValues[offset++];
output[topLeftRCOffset + d] +=
dyVal * inverseDxRLerpTimesInverseDxCLerp;
output[topRightRCOffset + d] += dyVal * inverseDxRLerpTimesDxCLerp;
output[bottomLeftRCOffset + d] += dyVal * dxRLerpTimesInverseDxCLerp;
output[bottomRightRCOffset + d] += dyVal * dxRLerpTimesDxCLerp;
}
}
}
}
return backend.makeTensorInfo([batch, xWidth, xHeight, depth], 'float32', output);
}
const resizeBilinearGradConfig$1 = {
kernelName: ResizeBilinearGrad,
backendName: 'cpu',
kernelFunc: resizeBilinearGrad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeNearestNeighbor(args) {
const { inputs, backend, attrs } = args;
const { images } = inputs;
const { alignCorners, halfPixelCenters, size } = attrs;
assertNotComplex(images, 'resizeNearestNeighbor');
const imagesStrides = computeStrides(images.shape);
const [newHeight, newWidth] = size;
const [batch, oldHeight, oldWidth, numChannels] = images.shape;
const xValues = backend.data.get(images.dataId).values;
const output = new Float32Array(batch * newHeight * newWidth * numChannels);
const effectiveInputSize = [
(alignCorners && newHeight > 1) ? oldHeight - 1 : oldHeight,
(alignCorners && newWidth > 1) ? oldWidth - 1 : oldWidth
];
const effectiveOutputSize = [
(alignCorners && newHeight > 1) ? newHeight - 1 : newHeight,
(alignCorners && newWidth > 1) ? newWidth - 1 : newWidth
];
const effectiveRowSizeRatio = effectiveInputSize[0] / effectiveOutputSize[0];
const effectiveColSizeRatio = effectiveInputSize[1] / effectiveOutputSize[1];
let outputOffset = 0;
for (let b = 0; b < batch; b++) {
const batchOffset = b * imagesStrides[0];
for (let r = 0; r < newHeight; r++) {
const sourceFracRow = halfPixelCenters ?
effectiveRowSizeRatio * (r + 0.5) :
effectiveRowSizeRatio * r;
let sourceNearestRow = Math.min(oldHeight - 1, alignCorners ? Math.round(sourceFracRow) : Math.floor(sourceFracRow));
if (halfPixelCenters) {
sourceNearestRow = Math.max(0, sourceNearestRow);
}
const rowOffset = batchOffset + sourceNearestRow * imagesStrides[1];
for (let c = 0; c < newWidth; c++) {
const sourceFracCol = halfPixelCenters ?
effectiveColSizeRatio * (c + 0.5) :
effectiveColSizeRatio * c;
let sourceNearestCol = Math.min(oldWidth - 1, alignCorners ? Math.round(sourceFracCol) :
Math.floor(sourceFracCol));
if (halfPixelCenters) {
sourceNearestCol = Math.max(0, sourceNearestCol);
}
const colOffset = rowOffset + sourceNearestCol * imagesStrides[2];
for (let d = 0; d < numChannels; d++) {
// Begin shader.
// Compute the fractional index of the source.
const newVal = xValues[colOffset + d];
output[outputOffset++] = newVal;
}
}
}
}
return backend.makeTensorInfo([batch, newHeight, newWidth, numChannels], images.dtype, output);
}
const resizeNearestNeighborConfig = {
kernelName: ResizeNearestNeighbor,
backendName: 'cpu',
kernelFunc: resizeNearestNeighbor
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function resizeNearestNeighborGrad(args) {
const { inputs, backend, attrs } = args;
const { images, dy } = inputs;
const { alignCorners } = attrs;
assertNotComplex([dy, images], 'resizeNearestNeighborGrad');
const imagesStrides = computeStrides(images.shape);
const dyStrides = computeStrides(dy.shape);
const [batch, xHeight, xWidth, depth] = images.shape;
const [, yHeight, yWidth] = dy.shape;
const output = new Float32Array(batch * xHeight * xWidth * depth);
const dyValues = backend.data.get(dy.dataId).values;
// In the backwards pass, we want to find the pixels that were generated
// for each pixel in the input image the forward pass
const effectiveXSize = [
(alignCorners && yHeight > 1) ? xHeight - 1 : xHeight,
(alignCorners && yWidth > 1) ? xWidth - 1 : xWidth
];
const effectiveYSize = [
(alignCorners && yHeight > 1) ? yHeight - 1 : yHeight,
(alignCorners && yWidth > 1) ? yWidth - 1 : yWidth
];
const heightScale = effectiveXSize[0] / effectiveYSize[0];
const widthScale = effectiveXSize[1] / effectiveYSize[1];
const invHeightScale = 1 / heightScale;
const invWidthScale = 1 / widthScale;
// This defines the size of the window of values around a particular
// index in dy that we want to search for contributions to dx.
const winHeight = (Math.ceil(invHeightScale) * 2) + 2;
const winWidth = (Math.ceil(invWidthScale) * 2) + 2;
// Loop over the output space.
for (let b = 0; b < batch; b++) {
const batchOffset = b * imagesStrides[0];
for (let r = 0; r < xHeight; r++) {
const rowOffset = batchOffset + r * imagesStrides[1];
// Compute bounds for where in dy we will look
const startRLerp = Math.floor(r * invHeightScale);
const startDyR = Math.floor(startRLerp - (winHeight / 2));
for (let c = 0; c < xWidth; c++) {
const colOffset = rowOffset + c * imagesStrides[2];
// Compute bounds for where in dy we will look
const startCLerp = Math.floor(c * invWidthScale);
const startDyC = Math.floor(startCLerp - (winWidth / 2));
for (let d = 0; d < depth; d++) {
let accum = 0;
// loop over dy
for (let dyRIndex = 0; dyRIndex < winHeight; dyRIndex++) {
const dyR = dyRIndex + startDyR;
// Guard against the window exceeding the bounds of dy
if (dyR < 0 || dyR >= yHeight) {
continue;
}
const dyROffset = batchOffset + dyR * dyStrides[1];
const sourceFracRow = dyR * heightScale;
const sourceNearestRow = Math.min(xHeight - 1, alignCorners ? Math.round(sourceFracRow) :
Math.floor(sourceFracRow));
if (r !== sourceNearestRow) {
continue;
}
for (let dyCIndex = 0; dyCIndex < winWidth; dyCIndex++) {
const dyC = dyCIndex + startDyC;
// Guard against the window exceeding the bounds of dy
if (dyC < 0 || dyC >= yWidth) {
continue;
}
const dyCOffset = dyROffset + dyC * dyStrides[2];
const sourceFracCol = dyC * widthScale;
const sourceNearestCol = Math.min(xWidth - 1, alignCorners ? Math.round(sourceFracCol) :
Math.floor(sourceFracCol));
if (c === sourceNearestCol) {
accum += dyValues[dyCOffset + d];
}
}
}
output[colOffset + d] = accum;
}
}
}
}
return backend.makeTensorInfo(images.shape, images.dtype, output);
}
const resizeNearestNeighborGradConfig$1 = {
kernelName: ResizeNearestNeighborGrad,
backendName: 'cpu',
kernelFunc: resizeNearestNeighborGrad
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function reverse(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { dims } = attrs;
assertNotComplex(x, 'reverse');
const xRank = x.shape.length;
const $dims = parseAxisParam(dims, x.shape);
if (xRank === 0) {
return identity$1({ inputs: { x }, backend });
}
const outBuf = new TensorBuffer(x.shape, x.dtype);
const xBuf = backend.bufferSync(x);
for (let i = 0; i < outBuf.size; i++) {
const outLoc = outBuf.indexToLoc(i);
const inLoc = outLoc.slice();
$dims.forEach(d => inLoc[d] = x.shape[d] - 1 - inLoc[d]);
outBuf.set(xBuf.get(...inLoc), ...outLoc);
}
return backend.makeTensorInfo(outBuf.shape, outBuf.dtype, outBuf.values);
}
const reverseConfig = {
kernelName: Reverse,
backendName: 'cpu',
kernelFunc: reverse
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const rotateWithOffsetConfig = {
kernelName: RotateWithOffset,
backendName: 'cpu',
kernelFunc: ({ inputs, attrs, backend }) => {
const { image } = inputs;
const { radians, fillValue, center } = attrs;
const cpuBackend = backend;
const output = getTypedArrayFromDType(image.dtype, sizeFromShape(image.shape));
const [batch, imageHeight, imageWidth, numChannels] = image.shape;
const [centerX, centerY] = getImageCenter(center, imageHeight, imageWidth);
const fullOpacityValue = 255;
const sinFactor = Math.sin(radians);
const cosFactor = Math.cos(radians);
const imageVals = cpuBackend.data.get(image.dataId).values;
for (let batchIdx = 0; batchIdx < batch; batchIdx++) {
const batchOffset = batchIdx * imageWidth * imageHeight * numChannels;
for (let row = 0; row < imageHeight; row++) {
const rowOffset = row * (imageWidth * numChannels);
for (let col = 0; col < imageWidth; col++) {
const colOffset = col * numChannels;
for (let channel = 0; channel < numChannels; channel++) {
const coords = [batch, row, col, channel];
const x = coords[2];
const y = coords[1];
// coordX/coordY are the result of rotating and translating x/y.
let coordX = (x - centerX) * cosFactor - (y - centerY) * sinFactor;
let coordY = (x - centerX) * sinFactor + (y - centerY) * cosFactor;
coordX = Math.round(coordX + centerX);
coordY = Math.round(coordY + centerY);
let outputValue = fillValue;
if (typeof fillValue !== 'number') {
if (channel === 3) {
outputValue = fullOpacityValue;
}
else {
outputValue = fillValue[channel];
}
}
// If the coordinate position falls within the image boundaries...
if (coordX >= 0 && coordX < imageWidth && coordY >= 0 &&
coordY < imageHeight) {
// set the output to the image value at the coordinate position.
const rotatedRowOffset = coordY * (imageWidth * numChannels);
const rotatedColOffset = coordX * numChannels;
const imageIdx = batchOffset + rotatedRowOffset + rotatedColOffset + channel;
outputValue = imageVals[imageIdx];
}
const outIdx = batchOffset + rowOffset + colOffset + channel;
output[outIdx] = outputValue;
}
}
}
}
const dataId = cpuBackend.write(output, image.shape, image.dtype);
return { dataId, shape: image.shape, dtype: image.dtype };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const round = unaryKernelFunc$1(Round, (xi) => {
// The algorithm is based on banker's rounding.
const base = Math.floor(xi);
if (xi - base < 0.5) {
return Math.floor(xi);
}
else if (xi - base > 0.5) {
return Math.ceil(xi);
}
else {
if (base % 2.0 === 0.0) {
return base;
}
else {
return base + 1.0;
}
}
});
const roundConfig = {
kernelName: Round,
backendName: 'cpu',
kernelFunc: round,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function scatterNd(args) {
const { inputs, backend, attrs } = args;
const { indices, updates } = inputs;
const { shape } = attrs;
const { sliceRank, numUpdates, sliceSize, strides, outputSize } = calculateShapes(updates, indices, shape);
const sumDupeIndices = true;
const indicesBuf = backend.bufferSync(indices);
const updatesBuf = backend.bufferSync(updates);
const outBuf = scatterImpl(indicesBuf, updatesBuf, shape, outputSize, sliceSize, numUpdates, sliceRank, strides, 0 /* defaultValue */, sumDupeIndices);
return backend.makeTensorInfo(shape, outBuf.dtype, outBuf.values);
}
const scatterNdConfig = {
kernelName: ScatterNd,
backendName: 'cpu',
kernelFunc: scatterNd
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function lowerBound(array, value) {
let left = 0;
let right = array.length;
let mid = 0;
while (left < right) {
mid = Math.floor((left + right) / 2);
if (array[mid] < value) {
left = mid + 1;
}
else {
right = mid;
}
}
return right;
}
function upperBound(array, value) {
let left = 0;
let right = array.length;
let mid = 0;
while (left < right) {
mid = Math.floor((left + right) / 2);
if (array[mid] <= value) {
left = mid + 1;
}
else {
right = mid;
}
}
return right;
}
function searchSortedImpl(sortedInputs, values, batchSize, numInputs, numValues, side) {
const output = getArrayFromDType('int32', batchSize * numValues);
for (let b = 0; b < batchSize; ++b) {
const sortedInputsSlice = sortedInputs.slice(b * numInputs, (b + 1) * numInputs);
const outputOffset = b * numValues;
for (let i = 0; i < numValues; ++i) {
output[outputOffset + i] = side === 'left' ?
lowerBound(sortedInputsSlice, values[i + outputOffset]) :
upperBound(sortedInputsSlice, values[i + outputOffset]);
}
}
return output;
}
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function searchSorted(args) {
const { inputs, backend, attrs } = args;
const { sortedSequence, values } = inputs;
const { side } = attrs;
const $sortedSequence = backend.data.get(sortedSequence.dataId).values;
const $values = backend.data.get(values.dataId).values;
const output = searchSortedImpl($sortedSequence, $values, sortedSequence.shape[0], sortedSequence.shape[1], values.shape[1], side);
return backend.makeTensorInfo(values.shape, 'int32', output);
}
const searchSortedConfig = {
kernelName: SearchSorted,
backendName: 'cpu',
kernelFunc: searchSorted,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function select(args) {
const { inputs, backend } = args;
const { condition, t, e } = inputs;
assertNotComplex([condition, t, e], 'select');
const conditionRank = condition.shape.length;
const values = backend.data.get(condition.dataId).values;
const tValues = backend.data.get(t.dataId).values;
const eValues = backend.data.get(e.dataId).values;
const resultDtype = upcastType(t.dtype, e.dtype);
const newValues = makeZerosTypedArray(sizeFromShape(t.shape), resultDtype);
let index = 0;
const offset = conditionRank === 0 || conditionRank > 1 || t.shape.length === 1 ?
1 :
sizeFromShape(t.shape.slice(1));
for (let i = 0; i < values.length; i++) {
for (let j = 0; j < offset; j++) {
if (values[i] === 1) {
newValues[index++] = tValues[i];
}
else {
newValues[index++] = eValues[i];
}
}
}
return backend.makeTensorInfo(t.shape, resultDtype, newValues);
}
const selectConfig = {
kernelName: Select,
backendName: 'cpu',
kernelFunc: select
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const scaleAlpha = SELU_SCALEALPHA;
const scale = SELU_SCALE;
const selu = unaryKernelFunc$1(Selu$1, (xi) => {
if (xi >= 0) {
return scale * xi;
}
else {
return scaleAlpha * (Math.exp(xi) - 1);
}
});
const seluConfig = {
kernelName: Selu$1,
backendName: 'cpu',
kernelFunc: selu,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sign = unaryKernelFunc$1(Sign, (xi) => {
if (xi < 0) {
return -1;
}
else if (xi > 0) {
return 1;
}
else {
return 0;
}
});
const signConfig = {
kernelName: Sign,
backendName: 'cpu',
kernelFunc: sign,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sin = unaryKernelFunc$1(Sin, (xi) => Math.sin(xi));
const sinConfig = {
kernelName: Sin,
backendName: 'cpu',
kernelFunc: sin,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sinh = unaryKernelFunc$1(Sinh, (xi) => Math.sinh(xi));
const sinhConfig = {
kernelName: Sinh,
backendName: 'cpu',
kernelFunc: sinh,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// mirrors the implementation of tf.nn.softplus: https://goo.gl/vkcvwX
// epsilon is the difference between 1.0 and the next representable float.
// For a single precision 32 bit float this should be 2^-23, see:
// https://math.byu.edu/~schow/work/IEEEFloatingPoint.htm
const epsilon$1 = 1.1920928955078125e-7;
const threshold = Math.log(epsilon$1) + 2.0;
const softplus = unaryKernelFunc$1(Softplus$1, (xi) => {
// Value above which exp(x) may overflow, but softplus(x) == x
// is within machine epsilon.
const tooLarge = xi > -threshold;
// Value below which exp(x) may underflow, but softplus(x) == exp(x)
// is within machine epsilon.
const tooSmall = xi < threshold;
const expX = Math.exp(xi);
let result;
if (tooSmall) {
result = expX;
}
else if (tooLarge) {
result = xi;
}
else {
result = Math.log(1.0 + expX);
}
return result;
});
const softplusConfig = {
kernelName: Softplus$1,
backendName: 'cpu',
kernelFunc: softplus,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function spaceToBatchND(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { blockShape, paddings } = attrs;
assertNotComplex([x], 'spaceToBatchND');
const prod = sizeFromShape(blockShape);
const completePaddings = [[0, 0]];
completePaddings.push(...paddings);
for (let i = 1 + blockShape.length; i < x.shape.length; ++i) {
completePaddings.push([0, 0]);
}
const paddedX = padV2Config.kernelFunc({
inputs: { x },
backend,
attrs: { paddings: completePaddings, constantValue: 0 }
});
const reshapedPaddedShape = getReshaped(paddedX.shape, blockShape, prod, false);
const permutedReshapedPaddedPermutation = getPermuted(reshapedPaddedShape.length, blockShape.length, false);
const flattenShape = getReshapedPermuted(paddedX.shape, blockShape, prod, false);
const reshapeInputs = { x: paddedX };
const reshapeAttrs = { shape: reshapedPaddedShape };
const paddedXReshaped = reshape({ inputs: reshapeInputs, backend, attrs: reshapeAttrs });
const transposeInputs = { x: paddedXReshaped };
const transposeAttrs = { perm: permutedReshapedPaddedPermutation };
const paddedXT = transpose$1({ inputs: transposeInputs, backend, attrs: transposeAttrs });
const resultReshapeInputs = { x: paddedXT };
const resultReshapeAttrs = { shape: flattenShape };
const result = reshape({ inputs: resultReshapeInputs, backend, attrs: resultReshapeAttrs });
backend.disposeIntermediateTensorInfo(paddedX);
backend.disposeIntermediateTensorInfo(paddedXReshaped);
backend.disposeIntermediateTensorInfo(paddedXT);
return result;
}
const spaceToBatchNDConfig = {
kernelName: SpaceToBatchND,
backendName: 'cpu',
kernelFunc: spaceToBatchND
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseFillEmptyRows(args) {
const { inputs, backend } = args;
const { indices, values, denseShape, defaultValue } = inputs;
if (denseShape.shape.length !== 1) {
throw new Error(`Dense shape must be a vector, saw:
${denseShape.shape}`);
}
if (indices.shape.length !== 2) {
throw new Error(`Indices must be a matrix, saw:
${indices.shape}`);
}
if (values.shape.length !== 1) {
throw new Error(`Values must be a vector, saw:
${values.shape}`);
}
if (defaultValue.shape.length !== 0) {
throw new Error(`Default value must be a scalar, saw:
${defaultValue.shape}`);
}
const $indices = backend.data.get(indices.dataId).values;
const $values = backend.data.get(values.dataId).values;
const $denseShape = backend.data.get(denseShape.dataId).values;
const $defaultValue = backend.data.get(defaultValue.dataId).values[0];
const [outputIndices, outputIndicesShape, outputValues, emptyRowIndicator, reverseIndexMap] = sparseFillEmptyRowsImpl($indices, indices.shape, indices.dtype, $values, values.dtype, $denseShape, $defaultValue);
return [
backend.makeTensorInfo(outputIndicesShape, indices.dtype, outputIndices),
backend.makeTensorInfo([outputIndicesShape[0]], values.dtype, outputValues),
backend.makeTensorInfo([emptyRowIndicator.length], 'bool', new Uint8Array(emptyRowIndicator.map((value) => Number(value)))),
backend.makeTensorInfo([reverseIndexMap.length], indices.dtype, new Int32Array(reverseIndexMap)),
];
}
const sparseFillEmptyRowsConfig = {
kernelName: SparseFillEmptyRows,
backendName: 'cpu',
kernelFunc: sparseFillEmptyRows,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseReshape(args) {
const { inputs, backend } = args;
const { inputIndices, inputShape, newShape } = inputs;
if (inputIndices.shape.length !== 2) {
throw new Error(`Input indices should be a matrix but received shape
${inputIndices.shape}`);
}
if (inputShape.shape.length !== 1) {
throw new Error(`Input shape should be a vector but received shape
${inputShape.shape}`);
}
if (newShape.shape.length !== 1) {
throw new Error(`Target shape should be a vector but received shape ${newShape.shape}`);
}
const $inputShape = Array.from(backend.data.get(inputShape.dataId).values);
const $inputIndices = backend.data.get(inputIndices.dataId).values;
const targetShape = Array.from(backend.data.get(newShape.dataId).values);
const [newIndices, indicesShape, outputShape] = sparseReshapeImpl($inputIndices, inputIndices.shape, inputIndices.dtype, $inputShape, targetShape);
return [
backend.makeTensorInfo(indicesShape, inputIndices.dtype, newIndices),
backend.makeTensorInfo([outputShape.length], newShape.dtype, new Int32Array(outputShape)),
];
}
const sparseReshapeConfig = {
kernelName: SparseReshape,
backendName: 'cpu',
kernelFunc: sparseReshape,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseSegmentMean(args) {
const { inputs, backend } = args;
const { data, indices, segmentIds } = inputs;
if (data.shape.length < 1) {
throw new Error(`Data should be at least 1 dimensional but received scalar`);
}
if (indices.shape.length !== 1) {
throw new Error(`Indices should be a vector but received shape
${indices.shape}`);
}
if (segmentIds.shape.length !== 1) {
throw new Error(`Segment ids should be a vector but received shape
${segmentIds.shape}`);
}
if (indices.shape[0] !== segmentIds.shape[0]) {
throw new Error(`segmentIds and indices should have same size.`);
}
const $data = backend.data.get(data.dataId).values;
const $indices = backend.data.get(indices.dataId).values;
const $segmentIds = backend.data.get(segmentIds.dataId).values;
const [outputData, outputDataShape] = sparseSegmentReductionImpl($data, data.shape, data.dtype, $indices, $segmentIds, true);
return backend.makeTensorInfo(outputDataShape, data.dtype, outputData);
}
const sparseSegmentMeanConfig = {
kernelName: SparseSegmentMean,
backendName: 'cpu',
kernelFunc: sparseSegmentMean,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseSegmentSum(args) {
const { inputs, backend } = args;
const { data, indices, segmentIds } = inputs;
if (data.shape.length < 1) {
throw new Error(`Data should be at least 1 dimensional but received scalar`);
}
if (indices.shape.length !== 1) {
throw new Error(`Indices should be a vector but received shape
${indices.shape}`);
}
if (segmentIds.shape.length !== 1) {
throw new Error(`Segment ids should be a vector but received shape
${segmentIds.shape}`);
}
if (indices.shape[0] !== segmentIds.shape[0]) {
throw new Error(`segmentIds and indices should have same size.`);
}
const $data = backend.data.get(data.dataId).values;
const $indices = backend.data.get(indices.dataId).values;
const $segmentIds = backend.data.get(segmentIds.dataId).values;
const [outputData, outputDataShape] = sparseSegmentReductionImpl($data, data.shape, data.dtype, $indices, $segmentIds);
return backend.makeTensorInfo(outputDataShape, data.dtype, outputData);
}
const sparseSegmentSumConfig = {
kernelName: SparseSegmentSum,
backendName: 'cpu',
kernelFunc: sparseSegmentSum,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function sparseToDense(args) {
const { inputs, backend, attrs } = args;
const { sparseIndices, sparseValues, defaultValue } = inputs;
const { outputShape } = attrs;
const { sliceRank, numUpdates, sliceSize, strides, outputSize } = calculateShapes(sparseValues, sparseIndices, outputShape);
const sumDupeIndices = false;
const indicesBuf = backend.bufferSync(sparseIndices);
let outBuf;
switch (sparseValues.dtype) {
case 'bool': {
const updatesBuf = backend.bufferSync(sparseValues);
const $defaultValue = Boolean(backend.data.get(defaultValue.dataId).values[0]);
outBuf = scatterImpl(indicesBuf, updatesBuf, outputShape, outputSize, sliceSize, numUpdates, sliceRank, strides, $defaultValue, sumDupeIndices);
break;
}
case 'float32': {
const updatesBuf = backend.bufferSync(sparseValues);
const $defaultValue = backend.data.get(defaultValue.dataId).values[0];
outBuf = scatterImpl(indicesBuf, updatesBuf, outputShape, outputSize, sliceSize, numUpdates, sliceRank, strides, $defaultValue, sumDupeIndices);
break;
}
case 'int32': {
const updatesBuf = backend.bufferSync(sparseValues);
const $defaultValue = backend.data.get(defaultValue.dataId).values[0];
outBuf = scatterImpl(indicesBuf, updatesBuf, outputShape, outputSize, sliceSize, numUpdates, sliceRank, strides, $defaultValue, sumDupeIndices);
break;
}
case 'string': {
const updatesBuf = backend.bufferSync(sparseValues);
const $defaultValue = decodeString(backend.data.get(defaultValue.dataId).values[0]);
outBuf = scatterImpl(indicesBuf, updatesBuf, outputShape, outputSize, sliceSize, numUpdates, sliceRank, strides, $defaultValue, sumDupeIndices);
break;
}
default:
throw new Error(`Unsupported type ${sparseValues.dtype}`);
}
return backend.makeTensorInfo(outputShape, outBuf.dtype, outBuf.values);
}
const sparseToDenseConfig = {
kernelName: SparseToDense,
backendName: 'cpu',
kernelFunc: sparseToDense
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function splitV(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { numOrSizeSplits, axis } = attrs;
const $axis = parseAxisParam(axis, x.shape)[0];
const splitSizes = prepareSplitSize(x, numOrSizeSplits, $axis);
const begin = new Array(x.shape.length).fill(0);
const size = x.shape.slice();
return splitSizes.map(s => {
const sliceSize = [...size];
sliceSize[$axis] = s;
const sliceT = slice$1({ inputs: { x }, backend, attrs: { begin, size: sliceSize } });
begin[$axis] += s;
return sliceT;
});
}
const splitVConfig = {
kernelName: SplitV,
backendName: 'cpu',
kernelFunc: splitV
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const squareConfig = {
kernelName: Square,
backendName: 'cpu',
kernelFunc: ({ inputs, backend }) => {
const { x } = inputs;
const cpuBackend = backend;
assertNotComplex(x, 'square');
const values = cpuBackend.data.get(x.dataId).values;
const newValues = new Float32Array(values.length);
for (let i = 0; i < values.length; ++i) {
const value = values[i];
newValues[i] = value * value;
}
const dataId = cpuBackend.write(newValues, x.shape, x.dtype);
return { dataId, shape: x.shape, dtype: x.dtype };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const step = unaryKernelFunc$1(Step, (xi, attrs) => {
const stepAttrs = attrs;
if (isNaN(xi)) {
return NaN;
}
else {
return xi > 0 ? 1 : stepAttrs.alpha;
}
});
const stepConfig = {
kernelName: Step,
backendName: 'cpu',
kernelFunc: step,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stridedSlice(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { begin, end, strides, beginMask, endMask, ellipsisMask, newAxisMask, shrinkAxisMask } = attrs;
assertNotComplex(x, 'stridedSlice');
const { finalShapeSparse, finalShape, isIdentity, sliceDim0, isSimpleSlice, begin: $begin, end: $end, strides: $strides } = sliceInfo(x.shape, begin, end, strides, beginMask, endMask, ellipsisMask, newAxisMask, shrinkAxisMask);
let result;
// ref:
// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/strided_slice_op.cc
if (isIdentity) {
// Optimization #1, slice is a no-op plus reshape
result = reshape({ inputs: { x }, backend, attrs: { shape: finalShape } });
}
else if (sliceDim0 || isSimpleSlice) {
// Optimization #2, slice is memory contiguous (only occurs in dim 0)
assert$1(x.shape.length >= 1, () => `Input must have rank at least 1, got: ${x.shape.length}`);
const size = computeOutShape$2($begin, $end, $strides);
// To tolerate begin[0] > end[0] (a 0-output slice), we min(begin, end).
const sliced = slice$1({ inputs: { x }, backend, attrs: { begin: $begin, size } });
result =
reshape({ inputs: { x: sliced }, backend, attrs: { shape: finalShape } });
backend.disposeIntermediateTensorInfo(sliced);
}
else {
const xBuf = backend.bufferSync(x);
const outBuf = stridedSliceImpl(finalShapeSparse, xBuf, $strides, $begin);
result = backend.makeTensorInfo(finalShape, outBuf.dtype, outBuf.values);
}
return result;
}
const stridedSliceConfig = {
kernelName: StridedSlice,
backendName: 'cpu',
kernelFunc: stridedSlice
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stringNGrams(args) {
const { inputs, backend, attrs } = args;
const { separator, nGramWidths, leftPad, rightPad, padWidth, preserveShortSequences } = attrs;
const { data, dataSplits } = inputs;
const $data = backend.data.get(data.dataId).values;
const $dataSplits = backend.data.get(dataSplits.dataId).values;
const [nGrams, nGramsSplits] = stringNGramsImpl($data, $dataSplits, separator, nGramWidths, leftPad, rightPad, padWidth, preserveShortSequences);
return [
backend.makeTensorInfo([nGrams.length], 'string', nGrams),
backend.makeTensorInfo(dataSplits.shape, 'int32', nGramsSplits),
];
}
const stringNGramsConfig = {
kernelName: StringNGrams,
backendName: 'cpu',
kernelFunc: stringNGrams,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stringSplit(args) {
const { inputs, backend, attrs } = args;
const { skipEmpty } = attrs;
const { input, delimiter } = inputs;
if (input.dtype !== 'string') {
throw new Error('Input must be of datatype string');
}
if (input.shape.length !== 1) {
throw new Error(`Input must be a vector, got shape: ${input.shape}`);
}
if (delimiter.shape.length !== 0) {
throw new Error(`Delimiter must be a scalar, got shape: ${delimiter.shape}`);
}
const $input = backend.data.get(input.dataId).values;
const $delimiter = backend.data.get(delimiter.dataId).values[0];
const [indices, values, shape] = stringSplitImpl($input, $delimiter, skipEmpty);
const outputSize = values.length;
return [
backend.makeTensorInfo([outputSize, 2], 'int32', indices),
backend.makeTensorInfo([outputSize], 'string', values),
backend.makeTensorInfo([2], 'int32', new Int32Array(shape))
];
}
const stringSplitConfig = {
kernelName: StringSplit,
backendName: 'cpu',
kernelFunc: stringSplit,
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function stringToHashBucketFast(args) {
const { inputs, backend, attrs } = args;
const { numBuckets } = attrs;
const { input } = inputs;
if (input.dtype !== 'string') {
throw new Error('Input must be of datatype string');
}
if (numBuckets <= 0) {
throw new Error(`Number of buckets must be at least 1`);
}
const $input = backend.data.get(input.dataId).values;
const output = stringToHashBucketFastImpl($input, numBuckets);
return backend.makeTensorInfo(input.shape, 'int32', output);
}
const stringToHashBucketFastConfig = {
kernelName: StringToHashBucketFast,
backendName: 'cpu',
kernelFunc: stringToHashBucketFast,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const tan = unaryKernelFunc$1(Tan, (xi) => Math.tan(xi));
const tanConfig = {
kernelName: Tan,
backendName: 'cpu',
kernelFunc: tan,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const tanh = unaryKernelFunc$1(Tanh$1, (xi) => Math.tanh(xi));
const tanhConfig = {
kernelName: Tanh$1,
backendName: 'cpu',
kernelFunc: tanh,
};
/**
* @license
* Copyright 2022 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function tensorScatterUpdate(args) {
const { inputs, backend } = args;
const { tensor, indices, updates } = inputs;
const { sliceRank, numUpdates, sliceSize, strides, outputSize } = calculateShapes(updates, indices, tensor.shape);
const sumDupeIndices = false;
const indicesBuf = backend.bufferSync(indices);
const updatesBuf = backend.bufferSync(updates);
const tensorBuf = backend.bufferSync(tensor);
const outBuf = scatterImpl(indicesBuf, updatesBuf, tensor.shape, outputSize, sliceSize, numUpdates, sliceRank, strides, tensorBuf, sumDupeIndices);
return backend.makeTensorInfo(tensor.shape, outBuf.dtype, outBuf.values);
}
const tensorScatterUpdateConfig = {
kernelName: TensorScatterUpdate,
backendName: 'cpu',
kernelFunc: tensorScatterUpdate
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function tile$1(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { reps } = attrs;
assertNotComplex(x, 'tile');
const outBuf = tileImpl(backend.bufferSync(x), reps);
return backend.makeTensorInfo(outBuf.shape, outBuf.dtype, outBuf.values);
}
const tileConfig = {
kernelName: Tile,
backendName: 'cpu',
kernelFunc: tile$1
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function topK(args) {
const { inputs, backend, attrs } = args;
const { x } = inputs;
const { k, sorted } = attrs;
assertNotComplex(x, 'topk');
const xVals = backend.data.get(x.dataId).values;
const [allTopKVals, allTopKIndices] = topKImpl(xVals, x.shape, x.dtype, k, sorted);
return [
backend.makeTensorInfo(allTopKVals.shape, allTopKVals.dtype, allTopKVals.values),
backend.makeTensorInfo(allTopKIndices.shape, allTopKIndices.dtype, allTopKIndices.values)
];
}
const topKConfig = {
kernelName: TopK,
backendName: 'cpu',
kernelFunc: topK
};
/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function transform(args) {
const { inputs, attrs, backend } = args;
const { image, transforms } = inputs;
const { interpolation, fillMode, fillValue, outputShape } = attrs;
const [batch, imageHeight, imageWidth, numChannels] = image.shape;
const [outHeight, outWidth] = outputShape != null ? outputShape : [imageHeight, imageWidth];
const outShape = [batch, outHeight, outWidth, numChannels];
const inStrides = computeStrides(image.shape);
const batchInStride = inStrides[0];
const rowInStride = inStrides[1];
const colInStride = inStrides[2];
const outStrides = computeStrides(outShape);
const batchOutStride = outStrides[0];
const rowOutStride = outStrides[1];
const colOutStride = outStrides[2];
const outVals = getTypedArrayFromDType(image.dtype, sizeFromShape(outShape));
outVals.fill(fillValue);
const imageVals = backend.data.get(image.dataId).values;
const transformVals = backend.data.get(transforms.dataId).values;
// Ref TF implementation:
// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/image/image_ops.h
for (let b = 0; b < batch; ++b) {
const transform = transforms.shape[0] === 1 ?
transformVals :
transformVals.subarray(b * 8, b * 8 + 8);
for (let outY = 0; outY < outHeight; ++outY) {
for (let outX = 0; outX < outWidth; ++outX) {
for (let channel = 0; channel < numChannels; ++channel) {
let val;
const projection = transform[6] * outX + transform[7] * outY + 1;
if (projection === 0) {
// Return the fill value for infinite coordinates,
// which are outside the input image
continue;
}
const inX = (transform[0] * outX + transform[1] * outY + transform[2]) /
projection;
const inY = (transform[3] * outX + transform[4] * outY + transform[5]) /
projection;
const x = mapCoord(inX, imageWidth, fillMode);
const y = mapCoord(inY, imageHeight, fillMode);
switch (interpolation) {
case 'nearest':
val = nearestInterpolation(imageVals, imageHeight, imageWidth, batchInStride, rowInStride, colInStride, b, y, x, channel, fillValue);
break;
case 'bilinear':
val = bilinearInterpolation(imageVals, imageHeight, imageWidth, batchInStride, rowInStride, colInStride, b, y, x, channel, fillValue);
break;
default:
throw new Error(`Error in Transform: Expect 'nearest' or ` +
`'bilinear', but got ${interpolation}`);
}
const ind = b * batchOutStride + outY * rowOutStride +
outX * colOutStride + channel;
outVals[ind] = val;
}
}
}
return backend.makeTensorInfo(outShape, image.dtype, outVals);
}
const dataId = backend.write(outVals, outShape, image.dtype);
return { dataId, shape: image.shape, dtype: image.dtype };
}
const transformConfig = {
kernelName: Transform,
backendName: 'cpu',
kernelFunc: transform
};
function mapCoord(outCoord, len, mode) {
switch (mode) {
case 'reflect':
return mapCoordReflect(outCoord, len);
case 'wrap':
return mapCoordWrap(outCoord, len);
case 'nearest':
return mapCoordNearest(outCoord, len);
case 'constant':
default:
return mapCoordConstant(outCoord);
}
}
function mapCoordReflect(outCoord, len) {
// Reflect [abcd] to [dcba|abcd|dcba].
let inCoord = outCoord;
if (inCoord < 0) {
if (len <= 1) {
inCoord = 0;
}
else {
const sz2 = 2 * len;
if (inCoord < sz2) {
inCoord = sz2 * Math.trunc(-inCoord / sz2) + inCoord;
}
inCoord = inCoord < -len ? inCoord + sz2 : -inCoord - 1;
}
}
else if (inCoord > len - 1) {
if (len <= 1) {
inCoord = 0;
}
else {
const sz2 = 2 * len;
inCoord -= sz2 * Math.trunc(inCoord / sz2);
if (inCoord >= len) {
inCoord = sz2 - inCoord - 1;
}
}
}
// clamp is necessary because when outCoord = 3.5 and len = 4,
// inCoord = 3.5 and will be rounded to 4 in nearest interpolation.
return clamp(0, inCoord, len - 1);
}
function mapCoordWrap(outCoord, len) {
// Wrap [abcd] to [abcd|abcd|abcd].
let inCoord = outCoord;
if (inCoord < 0) {
if (len <= 1) {
inCoord = 0;
}
else {
const sz = len - 1;
inCoord += len * (Math.trunc(-inCoord / sz) + 1);
}
}
else if (inCoord > len - 1) {
if (len <= 1) {
inCoord = 0;
}
else {
const sz = len - 1;
inCoord -= len * Math.trunc(inCoord / sz);
}
}
// clamp is necessary because when outCoord = -0.5 and len = 4,
// inCoord = 3.5 and will be rounded to 4 in nearest interpolation.
return clamp(0, inCoord, len - 1);
}
function mapCoordConstant(outCoord, len) {
return outCoord;
}
function mapCoordNearest(outCoord, len) {
return clamp(0, outCoord, len - 1);
}
function readWithFillValue(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, y, x, channel, fillValue) {
const ind = batch * batchStride + y * rowStride + x * colStride + channel;
if (0 <= y && y < imageHeight && 0 <= x && x < imageWidth) {
return imageVals[ind];
}
else {
return fillValue;
}
}
function nearestInterpolation(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, y, x, channel, fillValue) {
const $y = Math.round(y);
const $x = Math.round(x);
return readWithFillValue(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, $y, $x, channel, fillValue);
}
function bilinearInterpolation(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, y, x, channel, fillValue) {
const yFloor = Math.floor(y);
const xFloor = Math.floor(x);
const yCeil = yFloor + 1;
const xCeil = xFloor + 1;
// f(x, yFloor) = (xCeil - x) / (xCeil - xFloor) * f(xFloor, yFloor)
// + (x - xFloor) / (xCeil - xFloor) * f(xCeil, yFloor)
const valueYFloor = (xCeil - x) *
readWithFillValue(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, yFloor, xFloor, channel, fillValue) +
(x - xFloor) *
readWithFillValue(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, yFloor, xCeil, channel, fillValue);
// f(x, yCeil) = (xCeil - x) / (xCeil - xFloor) * f(xFloor, yCeil)
// + (x - xFloor) / (xCeil - xFloor) * f(xCeil, yCeil)
const valueYCeil = (xCeil - x) *
readWithFillValue(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, yCeil, xFloor, channel, fillValue) +
(x - xFloor) *
readWithFillValue(imageVals, imageHeight, imageWidth, batchStride, rowStride, colStride, batch, yCeil, xCeil, channel, fillValue);
// f(x, y) = (yCeil - y) / (yCeil - yFloor) * f(x, yFloor)
// + (y - yFloor) / (yCeil - yFloor) * f(x, yCeil)
return (yCeil - y) * valueYFloor + (y - yFloor) * valueYCeil;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function unique$1(args) {
const { inputs, attrs, backend } = args;
const { axis } = attrs;
const { x } = inputs;
assertNotComplex(x, 'unique');
const values = backend.data.get(x.dataId).values;
const { outputValues, outputShape, indices } = uniqueImpl(values, axis, x.shape, x.dtype);
return [
backend.makeTensorInfo(outputShape, x.dtype, outputValues),
backend.makeTensorInfo([indices.length], 'int32', indices),
];
}
const uniqueConfig = {
kernelName: Unique,
backendName: 'cpu',
kernelFunc: unique$1,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function unpack(args) {
const { inputs, backend, attrs } = args;
const { value } = inputs;
let { axis } = attrs;
if (axis < 0) {
axis += value.shape.length;
}
const valueRank = value.shape.length;
const num = value.shape[axis];
const outShape = new Array(valueRank - 1);
let outIndex = 0;
for (let i = 0; i < valueRank; i++) {
if (i !== axis) {
outShape[outIndex++] = value.shape[i];
}
}
const begin = new Array(valueRank).fill(0);
const size = value.shape.slice();
size[axis] = 1;
const res = new Array(num);
for (let i = 0; i < res.length; i++) {
begin[axis] = i;
const tempRes = slice$1({ inputs: { x: value }, backend, attrs: { begin, size } });
res[i] = reshape({ inputs: { x: tempRes }, backend, attrs: { shape: outShape } });
backend.disposeIntermediateTensorInfo(tempRes);
}
return res;
}
const unpackConfig = {
kernelName: Unpack,
backendName: 'cpu',
kernelFunc: unpack
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function unsortedSegmentSum(args) {
const { inputs, backend, attrs } = args;
const { x, segmentIds } = inputs;
const { numSegments } = attrs;
assertNotComplex(x, 'unsortedSegmentSum');
const xRank = x.shape.length;
const segmentIdsRank = segmentIds.shape.length;
const res = [];
const intermediates = [];
// Reshape the segment id's so that they can be broadcast with
// x. The new shape should be [segmentIds.shape, 1, ..., 1]
const numIters = xRank - segmentIdsRank;
let $segmentIds = segmentIds;
for (let i = 0; i < numIters; ++i) {
const expanded = expandDims$1({ inputs: { input: $segmentIds }, backend, attrs: { dim: i + 1 } });
$segmentIds = expanded;
intermediates.push(expanded);
}
for (let i = 0; i < numSegments; ++i) {
const scalarValue = createScalarValue(i, 'int32');
const segmentId = backend.makeTensorInfo([], 'int32', scalarValue);
const mask = equal$1({ inputs: { a: segmentId, b: $segmentIds }, backend });
const maskCasted = cast$2({ inputs: { x: mask }, backend, attrs: { dtype: 'float32' } });
const mul = multiply$1({ inputs: { a: maskCasted, b: x }, backend });
const sumTensorInfo = sum({ inputs: { x: mul }, backend, attrs: { axis: 0, keepDims: false } });
res.push(sumTensorInfo);
intermediates.push(segmentId);
intermediates.push(mask);
intermediates.push(maskCasted);
intermediates.push(mul);
intermediates.push(sumTensorInfo);
}
const result = pack({ inputs: res, backend, attrs: { axis: 0 } });
intermediates.forEach(t => backend.disposeIntermediateTensorInfo(t));
return result;
}
const unsortedSegmentSumConfig = {
kernelName: UnsortedSegmentSum,
backendName: 'cpu',
kernelFunc: unsortedSegmentSum
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// We explicitly import the modular kernels so they get registered in the
// global registry when we compile the library. A modular build would replace
// the contents of this file and import only the kernels that are needed.
// List all kernel configs here
const kernelConfigs = [
_fusedMatMulConfig,
absConfig$1,
acosConfig,
acoshConfig,
addConfig$1,
addNConfig,
allConfig,
anyConfig,
argMaxConfig,
argMinConfig,
asinConfig,
asinhConfig,
atanConfig,
atan2Config,
atanhConfig,
avgPoolConfig,
avgPool3DConfig,
avgPool3DGradConfig$1,
avgPoolGradConfig$1,
batchMatMulConfig,
batchNormConfig,
batchToSpaceNDConfig,
bincountConfig,
bitwiseAndConfig$1,
broadcastArgsConfig,
castConfig$1,
ceilConfig$1,
clipByValueConfig,
complexConfig$1,
complexAbsConfig,
concatConfig,
conv2DConfig,
conv2DBackpropFilterConfig,
conv2DBackpropInputConfig,
conv3DConfig,
conv3DBackpropFilterV2Config,
conv3DBackpropInputV2Config,
cosConfig,
coshConfig,
cropAndResizeConfig,
cumprodConfig,
cumsumConfig,
denseBincountConfig,
depthToSpaceConfig,
depthwiseConv2dNativeConfig,
depthwiseConv2dNativeBackpropFilterConfig,
depthwiseConv2dNativeBackpropInputConfig,
diagConfig,
dilation2DConfig,
dilation2DBackpropFilterConfig,
dilation2DBackpropInputConfig,
drawConfig,
einsumConfig,
eluConfig,
eluGradConfig$1,
equalConfig$1,
erfConfig,
expConfig$1,
expandDimsConfig,
expm1Config$1,
fftConfig,
fillConfig,
flipLeftRightConfig,
floorConfig$1,
floorDivConfig$1,
fusedConv2DConfig,
fusedDepthwiseConv2DConfig,
gatherNdConfig,
gatherV2Config,
greaterConfig$1,
greaterEqualConfig$1,
identityConfig$1,
ifftConfig,
imagConfig,
isFiniteConfig,
isInfConfig,
isNaNConfig,
leakyReluConfig,
lessConfig$1,
lessEqualConfig$1,
linSpaceConfig,
logConfig$1,
log1pConfig,
logicalAndConfig,
logicalNotConfig,
logicalOrConfig,
LRNConfig,
LRNGradConfig,
maxConfig,
maximumConfig$1,
maxPoolConfig,
maxPool3DConfig,
maxPool3DGradConfig$1,
maxPoolGradConfig$1,
maxPoolWithArgmaxConfig,
meanConfig,
minConfig,
minimumConfig$1,
mirrorPadConfig,
modConfig,
multinomialConfig,
multiplyConfig$1,
negConfig$1,
nonMaxSuppressionV3Config,
nonMaxSuppressionV4Config,
nonMaxSuppressionV5Config,
notEqualConfig$1,
oneHotConfig,
onesLikeConfig,
packConfig,
padV2Config,
powConfig,
preluConfig,
prodConfig$1,
raggedGatherConfig,
raggedRangeConfig,
raggedTensorToTensorConfig,
rangeConfig,
realConfig$1,
realDivConfig,
reciprocalConfig,
reluConfig,
relu6Config,
reshapeConfig,
resizeBilinearConfig,
resizeBilinearGradConfig$1,
resizeNearestNeighborConfig,
resizeNearestNeighborGradConfig$1,
reverseConfig,
rotateWithOffsetConfig,
roundConfig,
rsqrtConfig$1,
scatterNdConfig,
searchSortedConfig,
selectConfig,
seluConfig,
sigmoidConfig$1,
signConfig,
sinConfig,
sinhConfig,
sliceConfig$1,
softmaxConfig,
softplusConfig,
spaceToBatchNDConfig,
sparseFillEmptyRowsConfig,
sparseReshapeConfig,
sparseSegmentMeanConfig,
sparseSegmentSumConfig,
sparseToDenseConfig,
splitVConfig,
sqrtConfig$1,
squareConfig,
squaredDifferenceConfig$1,
staticRegexReplaceConfig$1,
stepConfig,
stridedSliceConfig,
stringNGramsConfig,
stringSplitConfig,
stringToHashBucketFastConfig,
subConfig$1,
sumConfig,
tanConfig,
tanhConfig,
tensorScatterUpdateConfig,
tileConfig,
topKConfig,
transformConfig,
transposeConfig$1,
uniqueConfig,
unpackConfig,
unsortedSegmentSumConfig,
zerosLikeConfig
];
for (const kernelConfig of kernelConfigs) {
registerKernel(kernelConfig);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const absGradConfig = {
kernelName: Abs,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(dy, step$2(cast$3(x, 'float32'), -1)) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const acosGradConfig = {
kernelName: Acos,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return {
x: () => {
const a = square$2(cast$3(x, 'float32'));
const b = sqrt$2(sub$2(scalar(1), a));
return neg$2(div$1(dy, b));
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const acoshGradConfig = {
kernelName: Acosh,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return {
x: () => {
const a = sqrt$2(sub$2(square$2(cast$3(x, 'float32')), 1));
return div$1(dy, a);
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const addGradConfig = {
kernelName: Add,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const outShape = assertAndGetBroadcastShape(a.shape, b.shape);
const derA = () => {
let res = dy;
const reduceAxes = getReductionAxes(a.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, a.shape);
};
const derB = () => {
let res = dy;
const reduceAxes = getReductionAxes(b.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, b.shape);
};
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const addNGradConfig = {
kernelName: AddN,
saveAllInputs: true,
gradFunc: (dy, saved) => {
const ders = {};
saved.forEach((_, i) => {
ders[i] = () => dy.clone();
});
return ders;
}
};
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const argMaxGradConfig = {
kernelName: ArgMax,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => zerosLike$2(x) };
}
};
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const argMinGradConfig = {
kernelName: ArgMin,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => zerosLike$2(x) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const asinGradConfig = {
kernelName: Asin,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, sqrt$2(sub$2(scalar(1), square$2(cast$3(x, 'float32'))))) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const asinhGradConfig = {
kernelName: Asinh,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return {
x: () => {
const a = sqrt$2(add$1(scalar(1), square$2(cast$3(x, 'float32'))));
return div$1(dy, a);
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const atan2GradConfig = {
kernelName: Atan2,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const outShape = assertAndGetBroadcastShape(a.shape, b.shape);
const derA = () => {
const d = add$1(square$2(a), square$2(b));
let res = mul(dy, div$1(b, d));
const reduceAxes = getReductionAxes(a.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, a.shape);
};
const derB = () => {
const d = add$1(square$2(a), square$2(b));
let res = neg$2(mul(dy, div$1(a, d)));
const reduceAxes = getReductionAxes(b.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, b.shape);
};
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const atanGradConfig = {
kernelName: Atan,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, add$1(square$2(cast$3(x, 'float32')), 1)) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const atanhGradConfig = {
kernelName: Atanh,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, sub$2(scalar(1), square$2(cast$3(x, 'float32')))) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the backprop of a 3d avg pool.
*
* @param dy The dy error, of rank 5 of shape
* [batchSize, depth, height, width, channels].
* assumed.
* @param input The original input image, of rank 5 or rank4 of shape
* [batchSize, depth, height, width, channels].
* @param filterSize The filter size:
* `[filterDepth, filterHeight, filterWidth]`.
* `filterSize` is a single number,
* then `filterDepth == filterHeight == filterWidth`.
* @param strides The strides of the pooling:
* `[strideDepth, strideHeight, strideWidth]`. If
* `strides` is a single number, then `strideHeight == strideWidth`.
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
* used in the forward prop of the op.
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*/
function avgPool3dGrad_(dy, input, filterSize, strides, pad, dimRoundingMode) {
const $dy = convertToTensor(dy, 'dy', 'avgPool3dGrad');
const $input = convertToTensor(input, 'input', 'avgPool3dGrad');
let dy5D = $dy;
let input5D = $input;
let reshapedTo5D = false;
if ($input.rank === 4) {
reshapedTo5D = true;
dy5D = reshape$2($dy, [1, $dy.shape[0], $dy.shape[1], $dy.shape[2], $dy.shape[3]]);
input5D = reshape$2($input, [
1, $input.shape[0], $input.shape[1], $input.shape[2], $input.shape[3]
]);
}
assert$1(dy5D.rank === 5, () => `Error in avgPool3dGrad: dy must be rank 5 but got rank ` +
`${dy5D.rank}.`);
assert$1(input5D.rank === 5, () => `Error in avgPool3dGrad: input must be rank 5 but got rank ` +
`${input5D.rank}.`);
checkPadOnDimRoundingMode('avgPool3dGrad', pad, dimRoundingMode);
const inputs = { dy: dy5D, input: input5D };
const attrs = { filterSize, strides, pad, dimRoundingMode };
// tslint:disable-next-line: no-unnecessary-type-assertion
const res = ENGINE.runKernel(AvgPool3DGrad, inputs, attrs);
if (reshapedTo5D) {
return reshape$2(res, [res.shape[1], res.shape[2], res.shape[3], res.shape[4]]);
}
return res;
}
const avgPool3dGrad = /* @__PURE__ */ op({ avgPool3dGrad_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const avgPool3DGradConfig = {
kernelName: AvgPool3D,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { filterSize, strides, pad, dimRoundingMode } = attrs;
return {
x: () => avgPool3dGrad(dy, x, filterSize, strides, pad, dimRoundingMode)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the backprop of an 2D avg pool.
*
* @param dy The dy error, of rank 4 or rank 3 of shape
* [batchSize, height, width, channels]. If rank 3, batch of 1 is
* assumed.
* @param input The input image, of rank 4 or rank 3 of shape
* [batchSize, height, width, channels]. If rank 3, batch of 1 is
* assumed.
* @param filterSize The filter size: `[filterHeight, filterWidth]`. If
* `filterSize` is a single number, then `filterHeight == filterWidth`.
* @param strides The strides of the pooling: `[strideHeight, strideWidth]`. If
* `strides` is a single number, then `strideHeight == strideWidth`.
* @param pad The type of padding algorithm used in the forward prop of the op.
* 'same', 'valid', for more info, see this guide:
* [https://www.tensorflow.org/api_docs/python/tf/nn/convolution](
* https://www.tensorflow.org/api_docs/python/tf/nn/convolution)
*/
function avgPoolGrad_(dy, input, filterSize, strides, pad) {
const $dy = convertToTensor(dy, 'dy', 'avgPoolGrad');
const $input = convertToTensor(input, 'input', 'avgPoolGrad');
assert$1($input.rank === $dy.rank, () => `Rank of input (${$input.rank}) does not match rank of dy (${$dy.rank})`);
let input4D = $input;
let dy4D = $dy;
let reshapedTo4D = false;
if ($input.rank === 3) {
reshapedTo4D = true;
input4D =
reshape$2($input, [1, $input.shape[0], $input.shape[1], $input.shape[2]]);
dy4D = reshape$2($dy, [1, $dy.shape[0], $dy.shape[1], $dy.shape[2]]);
}
assert$1(dy4D.rank === 4, () => `Error in avgPoolGrad: dy must be rank 4 but got rank ` +
`${dy4D.rank}.`);
assert$1(input4D.rank === 4, () => `Error in avgPoolGrad: input must be rank 4 but got rank ` +
`${input4D.rank}.`);
const inputs = { dy: dy4D, input: input4D };
const attrs = { filterSize, strides, pad };
// tslint:disable-next-line: no-unnecessary-type-assertion
const res = ENGINE.runKernel(AvgPoolGrad, inputs, attrs);
if (reshapedTo4D) {
return reshape$2(res, [res.shape[1], res.shape[2], res.shape[3]]);
}
return res;
}
const avgPoolGrad = /* @__PURE__ */ op({ avgPoolGrad_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const avgPoolGradConfig = {
kernelName: AvgPool,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { filterSize, strides, pad } = attrs;
return { x: () => avgPoolGrad(dy, x, filterSize, strides, pad) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const batchMatMulGradConfig = {
kernelName: BatchMatMul,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved, attrs) => {
const [a, b] = saved;
const { transposeA, transposeB } = attrs;
if (!transposeA && !transposeB) {
return {
a: () => matMul$1(dy, b, false, true),
b: () => matMul$1(a, dy, true, false)
};
}
else if (!transposeA && transposeB) {
return {
a: () => matMul$1(dy, b, false, false),
b: () => matMul$1(dy, a, true, false)
};
}
else if (transposeA && !transposeB) {
return {
a: () => matMul$1(b, dy, false, true),
b: () => matMul$1(a, dy, false, false)
};
}
else {
return {
a: () => matMul$1(b, dy, true, true),
b: () => matMul$1(dy, a, true, true)
};
}
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const batchToSpaceNDGradConfig = {
kernelName: BatchToSpaceND,
gradFunc: (dy, saved, attrs) => {
const { blockShape, crops } = attrs;
return { x: () => spaceToBatchND$2(dy, blockShape, crops) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const broadcastToGradConfig = {
kernelName: BroadcastTo,
gradFunc: (dy, saved, attrs) => {
const broadCastToAttrs = attrs;
const inputShape = broadCastToAttrs.inputShape;
const outputShape = broadCastToAttrs.shape;
const reps = Array.from(outputShape);
for (let i = inputShape.length - 1; i >= 0; i--) {
if (inputShape[i] === outputShape[i]) {
reps[i] = 1;
}
else if (inputShape[i] !== 1) {
throw new Error(`broadcastTo(): [${inputShape}] cannot be broadcast to [${outputShape}].`);
}
}
const axes = [];
for (let i = 0; i < reps.length; i++) {
if (reps[i] > 1) {
axes.push(i);
}
}
return { x: () => sum$2(dy, axes, true /* keepDims */) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const castGradConfig = {
kernelName: Cast,
gradFunc: (dy) => {
return { x: () => dy.clone() };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const ceilGradConfig = {
kernelName: Ceil,
gradFunc: (dy) => {
// TODO(manrajgrover): Return null for gradients when backprop supports it.
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const clipByValueGradConfig = {
kernelName: ClipByValue,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { clipValueMin, clipValueMax } = attrs;
return {
x: () => where(logicalAnd$2(greaterEqual$2(x, clipValueMin), lessEqual$2(x, clipValueMax)), dy, zerosLike$2(dy)),
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const complexAbsGradConfig = {
kernelName: ComplexAbs,
inputsToSave: ['x'],
gradFunc: absGradConfig.gradFunc,
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const concatGradConfig = {
kernelName: Concat,
saveAllInputs: true,
gradFunc: (dy, saved, attrs) => {
const shapes = saved.map(t => t.shape);
const { axis } = attrs;
const $axis = parseAxisParam(axis, saved[0].shape)[0];
const sizeSplits = shapes.map(s => s[$axis]);
const derTensors = split$1(dy, sizeSplits, $axis);
return derTensors.map(t => () => t);
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const conv2DGradConfig = {
kernelName: Conv2D,
inputsToSave: ['x', 'filter'],
gradFunc: (dy, saved, attrs) => {
const [x4D, $filter] = saved;
const { dilations, strides, pad, dataFormat } = attrs;
assert$1(tupleValuesAreOne(dilations), () => 'Error in gradient of conv2D: dilation rates greater than 1 ' +
`are not yet supported in gradients. Got dilations '${dilations}'`);
return {
x: () => conv2DBackpropInput$2(x4D.shape, dy, $filter, strides, pad, dataFormat),
filter: () => conv2DBackpropFilter$2(x4D, dy, $filter.shape, strides, pad, dataFormat)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const conv2DBackpropInputGradConfig = {
kernelName: Conv2DBackpropInput,
inputsToSave: ['dy', 'filter'],
gradFunc: (ddx, saved, attrs) => {
const [dy, filter] = saved;
const { strides, pad, dataFormat, dimRoundingMode } = attrs;
return {
dy: () => conv2d$1(ddx, filter, strides, pad, dataFormat, 1 /* dilations */, dimRoundingMode),
filter: () => conv2DBackpropFilter$2(ddx, dy, filter.shape, strides, pad, dataFormat, dimRoundingMode)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the derivative of the filter of a 3D convolution.
*
* @param x The input tensor, of rank 5 or rank 4 of shape
* [batch, depth, height, width, inChannels]. If rank 4, batch of 1 is
* assumed.
* @param dy The dy image, of rank 5 or rank 4, of shape
* [batch, depth, height, width, outDepth]. If rank 4, batch of 1 is
* assumed.
* @param filterShape The shape of the filter, length 5,
* [filterDepth, filterHeight, filterWidth, inDepth, outDepth].
* @param strides The strides of the convolution: [strideDepth, strideHeight,
* strideWidth].
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
* used in the forward prop of the op.
*/
function conv3DBackpropFilter_(x, dy, filterShape, strides, pad) {
let x5D = x;
if (x.rank === 4) {
x5D = reshape$2(x, [1, x.shape[0], x.shape[1], x.shape[2], x.shape[3]]);
}
let dy5D = dy;
if (dy5D.rank === 4) {
dy5D = reshape$2(dy, [1, dy.shape[0], dy.shape[1], dy.shape[2], dy.shape[3]]);
}
assert$1(x5D.rank === 5, () => `Error in conv3dDerFilter: input must be rank 5, but got shape ` +
`${x5D.shape}.`);
assert$1(dy5D.rank === 5, () => `Error in conv3dDerFilter: dy must be rank 5, but got shape ` +
`${dy5D.shape}.`);
assert$1(filterShape.length === 5, () => `Error in conv3dDerFilter: filterShape must be length 5, but got ` +
`${filterShape}.`);
assert$1(x5D.shape[4] === filterShape[3], () => `Error in conv3dDerFilter: depth of input ${x5D.shape[4]}) must ` +
`match input depth in filter (${filterShape[3]}.`);
assert$1(dy5D.shape[4] === filterShape[4], () => `Error in conv3dDerFilter: depth of dy (${dy5D.shape[4]}) must ` +
`match output depth for filter (${filterShape[4]}).`);
const inputs = { x: x5D, dy: dy5D };
const attrs = { strides, pad, filterShape };
// tslint:disable-next-line: no-unnecessary-type-assertion
return ENGINE.runKernel(Conv3DBackpropFilterV2, inputs, attrs);
}
const conv3DBackpropFilter = /* @__PURE__ */ op({ conv3DBackpropFilter_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const conv3DGradConfig = {
kernelName: Conv3D,
inputsToSave: ['x', 'filter'],
gradFunc: (dy, saved, attrs) => {
const { dilations, strides, pad } = attrs;
assert$1(tupleValuesAreOne(dilations), () => 'Error in gradient of conv3D: dilation rates greater than 1 are ' +
`not yet supported in gradients. Got dilations '${dilations}'`);
const [x5D, $filter] = saved;
return {
x: () => conv3DBackpropInput$1(x5D.shape, dy, $filter, strides, pad),
filter: () => conv3DBackpropFilter(x5D, dy, $filter.shape, strides, pad)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const cosGradConfig = {
kernelName: Cos,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(neg$2(sin$2(cast$3(x, 'float32'))), dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const coshGradConfig = {
kernelName: Cosh,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(sinh$2(cast$3(x, 'float32')), dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const cumsumGradConfig = {
kernelName: Cumsum,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { axis, exclusive, reverse } = attrs;
return {
x: () => {
const permutation = getAxesPermutation([axis], x.rank);
let out = cumsum$2(dy, axis, exclusive, !reverse);
if (permutation != null) {
out = transpose$2(out, permutation);
}
return out;
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const depthwiseConv2dNativeGradConfig = {
kernelName: DepthwiseConv2dNative,
inputsToSave: ['x', 'filter'],
gradFunc: (dy, saved, attrs) => {
const { dilations, strides, pad, dimRoundingMode } = attrs;
const $dilations = dilations == null ? [1, 1] : dilations;
assert$1(tupleValuesAreOne($dilations), () => 'Error in gradient of depthwiseConv2dNative: dilation rates ' +
`greater than 1 are not yet supported. Got dilations ` +
`'${$dilations}'`);
const [x, filter] = saved;
assert$1(x.rank === 4, () => `Error in gradient of depthwiseConv2dNative: input must be ` +
`rank 4, but got rank ${x.rank}.`);
assert$1(filter.rank === 4, () => `Error in gradient of depthwiseConv2dNative: filter must be ` +
`rank 4, but got rank ${filter.rank}.`);
assert$1(x.shape[3] === filter.shape[2], () => `Error in gradient of depthwiseConv2d: number of input ` +
`channels (${x.shape[3]}) must match the inChannels dimension ` +
`in filter ${filter.shape[2]}.`);
assert$1(eitherStridesOrDilationsAreOne(strides, $dilations), () => 'Error in gradient of depthwiseConv2d: Either strides or ' +
`dilations must be 1. Got strides ${strides} and dilations ` +
`'${$dilations}'.`);
checkPadOnDimRoundingMode('depthwiseConv2d', pad, dimRoundingMode);
return {
x: () => depthwiseConv2dNativeBackpropInput$2(x.shape, dy, filter, strides, pad, $dilations, dimRoundingMode),
filter: () => depthwiseConv2dNativeBackpropFilter$2(x, dy, filter.shape, strides, pad, $dilations, dimRoundingMode),
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const dilation2dGradConfig = {
kernelName: Dilation2D,
inputsToSave: ['x', 'filter'],
gradFunc: (dy, saved, attrs) => {
const [x, filter] = saved;
const inputInputs = { x, filter, dy };
const filterInputs = { x, filter, dy };
return {
x: () => ENGINE.runKernel(Dilation2DBackpropInput, inputInputs, attrs),
filter: () => ENGINE.runKernel(Dilation2DBackpropFilter, filterInputs, attrs)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const eluGradConfig = {
kernelName: Elu$1,
outputsToSave: [true],
gradFunc: (dy, saved) => {
const [y] = saved;
const inputs = { dy, y };
return { x: () => ENGINE.runKernel(EluGrad, inputs) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const erfGradConfig = {
kernelName: Erf,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
const a = mul(exp$2(neg$2(square$2(x))), 2 / Math.sqrt(Math.PI));
return { x: () => mul(dy, a) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const expGradConfig = {
kernelName: Exp,
outputsToSave: [true],
gradFunc: (dy, saved) => {
const [y] = saved;
return { x: () => mul(dy, y) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const expandDimsGradConfig = {
kernelName: ExpandDims,
inputsToSave: ['input'],
gradFunc: (dy, saved) => {
const [input] = saved;
return { input: () => reshape$2(dy, input.shape) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const expm1GradConfig = {
kernelName: Expm1,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(dy, exp$2(x)) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const floorGradConfig = {
kernelName: Floor,
gradFunc: (dy) => {
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const floorDivGradConfig = {
kernelName: FloorDiv,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const outShape = assertAndGetBroadcastShape(a.shape, b.shape);
const derA = () => {
const res = div$1(dy, cast$3(b, 'float32'));
const reduceAxes = getReductionAxes(a.shape, outShape);
if (reduceAxes.length > 0) {
return reshape$2(sum$2(res, reduceAxes), a.shape);
}
return res;
};
const derB = () => {
let res = mul(dy, cast$3(a, 'float32'));
const reduceAxes = getReductionAxes(b.shape, outShape);
if (reduceAxes.length > 0) {
res = reshape$2(sum$2(res, reduceAxes), b.shape);
}
const tmp = square$2(b);
return neg$2(div$1(res, cast$3(tmp, 'float32')));
};
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const fusedBatchNormGradConfig = {
kernelName: FusedBatchNorm,
inputsToSave: ['x', 'mean', 'variance', 'scale'],
gradFunc: (dy, saved, attrs) => {
const { varianceEpsilon } = attrs;
const [x, mean, variance, scale] = saved;
const scaleValue = scale == null ? scalar(1) : scale;
const reductionAxes = getReductionAxes(mean.shape, x.shape);
const tileShape = [];
if (mean.rank === 1) {
for (let i = 0; i < x.shape.length - 1; ++i) {
tileShape.push(x.shape[i]);
}
tileShape.push(1);
}
const xMinusMean = sub$2(x, mean);
const dyTimesScaleValue = mul(dy, scaleValue);
const oneOverSqrtVariance = rsqrt$2(add$1(variance, scalar(varianceEpsilon)));
const minusHalfRCube = mul(mul(mul(oneOverSqrtVariance, oneOverSqrtVariance), oneOverSqrtVariance), scalar(-0.5));
const derX = () => {
if (mean.rank === 1) {
return reshape$2(mul(mul(dy, tile$3(reshape$2(oneOverSqrtVariance, [1, 1, 1, mean.shape[0]]), tileShape)), scaleValue), x.shape);
}
else {
return reshape$2(mul(mul(dy, oneOverSqrtVariance), scaleValue), x.shape);
}
};
const derMean = () => {
let meanDer = mul(mul(oneOverSqrtVariance, scalar(-1)), dyTimesScaleValue);
if (mean.rank === 1) {
meanDer = sum$2(meanDer, reductionAxes);
}
return reshape$2(meanDer, mean.shape);
};
const derVariance = () => {
let varianceDer = mul(mul(minusHalfRCube, xMinusMean), dyTimesScaleValue);
if (mean.rank === 1) {
varianceDer = sum$2(varianceDer, reductionAxes);
}
return reshape$2(varianceDer, mean.shape);
};
const derScale = () => {
const xMinusMean2TimesRsqrt = mul(xMinusMean, oneOverSqrtVariance);
let scaleDer = mul(dy, xMinusMean2TimesRsqrt);
if (mean.rank === 1) {
scaleDer = sum$2(scaleDer, reductionAxes);
}
return reshape$2(scaleDer, mean.shape);
};
const derOffset = () => {
let offsetDer = dy;
if (mean.rank === 1) {
offsetDer = sum$2(offsetDer, reductionAxes);
}
return reshape$2(offsetDer, mean.shape);
};
return {
x: derX,
mean: derMean,
variance: derVariance,
scale: derScale,
offset: derOffset
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const gatherGradConfig = {
kernelName: GatherV2,
inputsToSave: ['x', 'indices'],
gradFunc: (dy, saved, attrs) => {
const [x, indices] = saved;
const { axis, batchDims } = attrs;
const parsedAxis = parseAxisParam(axis, x.shape)[0];
const derXBatch = (x, indices, dy) => {
return () => {
const paramsShape = x.shape;
const indicesSize = indices.size;
const outerShape = paramsShape.slice(0, parsedAxis);
const outerDims = outerShape.length;
const innerShape = paramsShape.slice(axis, paramsShape.length).slice(1);
const innerDims = innerShape.length;
const outerAxesIndices = arrayRange(0, outerDims);
const innerAxesIndices = arrayRange(outerDims + 1, outerDims + 1 + innerDims);
const valuesShape = arrayConcat([outerShape, [indicesSize],
innerShape]);
const values = reshape$2(dy, valuesShape);
const reshapedIndices = reshape$2(indices, [indicesSize]);
const transposeDims = arrayConcat([[outerDims], outerAxesIndices, innerAxesIndices]);
const valuesTranspose = transpose$2(values, transposeDims);
let paramsGrad = unsortedSegmentSum$2(valuesTranspose, reshapedIndices, x.shape[parsedAxis]);
const invertTransposeDims = getUndoAxesPermutation(transposeDims);
paramsGrad = transpose$2(paramsGrad, invertTransposeDims);
return paramsGrad;
};
};
if (batchDims === 1) {
const batchSize = x.shape[0];
const xBatch = x.split(batchSize, 0);
const derXBatched = () => {
const stacked = stack(xBatch.map((x, i) => {
return derXBatch(x, indices.slice(i, 1), dy.slice(i, 1))();
}));
return stacked.reshape(x.shape);
};
return { x: derXBatched, indices: () => indices };
}
else {
return { x: derXBatch(x, indices, dy), indices: () => indices };
}
}
};
function arrayRange(start, stop) {
const result = [];
for (let i = start; i < stop; ++i) {
result.push(i);
}
return result;
}
function arrayConcat(arrays) {
const result = [];
for (let i = 0; i < arrays.length; ++i) {
for (let j = 0; j < arrays[i].length; ++j) {
result.push(arrays[i][j]);
}
}
return result;
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const greaterEqualGradConfig = {
kernelName: GreaterEqual,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
return { a: () => zerosLike$2(a), b: () => zerosLike$2(b) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const identityGradConfig = {
kernelName: Identity$1,
gradFunc: (dy) => {
return { x: () => cast$3(dy, 'float32') };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const isFiniteGradConfig = {
kernelName: IsFinite,
gradFunc: (dy) => {
// TODO(nsthorat): Let gradients be null for cases where we want to stop
// backpropgation.
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const isInfGradConfig = {
kernelName: IsInf,
gradFunc: (dy) => {
// TODO(nsthorat): Let gradients be null for cases where we want to stop
// backpropgation.
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const isNanGradConfig = {
kernelName: IsNan,
gradFunc: (dy) => {
// TODO(nsthorat): Let gradients be null for cases where we want to stop
// backpropgation.
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const leakyReluGradConfig = {
kernelName: LeakyRelu,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { alpha } = attrs;
const mask = greater$2(x, 0);
// Returns `gradients * (features > 0) + alpha * gradients * (features <=
// 0)`.
return { x: () => where(mask, dy, mul(dy, alpha)) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const log1pGradConfig = {
kernelName: Log1p,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, add$1(x, 1)) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const logGradConfig = {
kernelName: Log,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, cast$3(x, 'float32')) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const logSoftmaxGradConfig = {
kernelName: LogSoftmax$1,
inputsToSave: [],
outputsToSave: [true],
gradFunc: (dy, saved, attrs) => {
const [value] = saved;
const { axis } = attrs;
return {
logits: () => {
const keepDims = true;
const softmax = exp$2(value);
return sub$2(dy, mul(sum$2(dy, axis, keepDims), softmax));
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
function localResponseNormalizationBackprop_(x, y, dy, depthRadius = 5, bias = 1, alpha = 1, beta = 0.5) {
const inputs = { x, y, dy };
const attrs = { depthRadius, bias, alpha, beta };
return ENGINE.runKernel(LRNGrad, inputs, attrs);
}
const localResponseNormalizationBackprop = op({ localResponseNormalizationBackprop_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const lrnGradConfig = {
kernelName: LRN,
inputsToSave: ['x'],
outputsToSave: [true],
gradFunc: (dy, saved, attrs) => {
const [x, y] = saved;
const { depthRadius, bias, alpha, beta } = attrs;
return {
x: () => localResponseNormalizationBackprop(x, y, dy, depthRadius, bias, alpha, beta)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Gradient helper function for the min and max operations.
*/
function gradForMinAndMax(dy, y, xOrig, origAxes) {
if (y.rank < xOrig.rank) {
y = reshape$2(y, expandShapeToKeepDim(y.shape, origAxes));
}
if (dy.rank < xOrig.rank) {
dy = reshape$2(dy, expandShapeToKeepDim(dy.shape, origAxes));
}
return {
x: () => {
const dx = mul(dy, cast$3(equal$2(xOrig, y), dy.dtype));
return dx;
}
};
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const maxGradConfig = {
kernelName: Max,
inputsToSave: ['x'],
outputsToSave: [true],
gradFunc: (dy, saved, attrs) => {
const maxAttrs = attrs;
const { reductionIndices } = maxAttrs;
const x = saved[0];
const y = saved[1];
const origAxes = parseAxisParam(reductionIndices, x.shape);
const maxGrad = gradForMinAndMax(dy, y, x, origAxes);
return {
x: () => {
return maxGrad['x']();
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const maximumGradConfig = {
kernelName: Maximum,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const derA = () => mul(dy, cast$3(greaterEqual$2(a, b), 'float32'));
const derB = () => mul(dy, cast$3(less$2(a, b), 'float32'));
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the backprop of a 3d max pool.
*
* @param dy The dy error, of rank 5 of shape
* [batchSize, depth, height, width, channels].
* assumed.
* @param input The original input image, of rank 5 or rank 4 of shape
* [batchSize, depth, height, width, channels].
* @param output The original output image, of rank 5 of shape
* [batchSize, outDepth, outHeight, outWidth, channels].
* @param filterSize The filter size:
* `[filterDepth, filterHeight, filterWidth]`.
* `filterSize` is a single number,
* then `filterDepth == filterHeight == filterWidth`.
* @param strides The strides of the pooling:
* `[strideDepth, strideHeight, strideWidth]`. If
* `strides` is a single number, then `strideHeight == strideWidth`.
* @param pad A string from: 'same', 'valid'. The type of padding algorithm
* used in the forward prop of the op.
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*/
function maxPool3dGrad_(dy, input, output, filterSize, strides, pad, dimRoundingMode) {
const $dy = convertToTensor(dy, 'dy', 'maxPool3dGrad');
const $input = convertToTensor(input, 'input', 'maxPool3dGrad');
const $output = convertToTensor(output, 'output', 'maxPool3dGrad');
let dy5D = $dy;
let input5D = $input;
let output5D = $output;
let reshapedTo5D = false;
if ($input.rank === 4) {
reshapedTo5D = true;
dy5D = reshape$2($dy, [1, $dy.shape[0], $dy.shape[1], $dy.shape[2], $dy.shape[3]]);
input5D = reshape$2($input, [
1, $input.shape[0], $input.shape[1], $input.shape[2], $input.shape[3]
]);
output5D = reshape$2($output, [
1, $output.shape[0], $output.shape[1], $output.shape[2], $output.shape[3]
]);
}
assert$1(dy5D.rank === 5, () => `Error in maxPool3dGrad: dy must be rank 5 but got rank ` +
`${dy5D.rank}.`);
assert$1(input5D.rank === 5, () => `Error in maxPool3dGrad: input must be rank 5 but got rank ` +
`${input5D.rank}.`);
assert$1(output5D.rank === 5, () => `Error in maxPool3dGrad: output must be rank 5 but got rank ` +
`${output5D.rank}.`);
checkPadOnDimRoundingMode('maxPool3dGrad', pad, dimRoundingMode);
const inputs = { dy: dy5D, input: input5D, output: output5D };
const attrs = { filterSize, strides, pad, dimRoundingMode };
// tslint:disable-next-line: no-unnecessary-type-assertion
const res = ENGINE.runKernel(MaxPool3DGrad, inputs, attrs);
if (reshapedTo5D) {
return reshape$2(res, [res.shape[1], res.shape[2], res.shape[3], res.shape[4]]);
}
return res;
}
const maxPool3dGrad = /* @__PURE__ */ op({ maxPool3dGrad_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const maxPool3DGradConfig = {
kernelName: MaxPool3D,
inputsToSave: ['x'],
outputsToSave: [true],
gradFunc: (dy, saved, attrs) => {
const [x, y] = saved;
const { filterSize, strides, pad, dimRoundingMode } = attrs;
return {
x: () => maxPool3dGrad(dy, x, y, filterSize, strides, pad, dimRoundingMode)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/**
* Computes the backprop of a 2D max pool.
*
* @param dy The dy error, of rank 4 or rank 3 of shape
* [batchSize, height, width, channels]. If rank 3, batch of 1 is
* assumed.
* @param input The original input image, of rank 4, of shape
* [batchSize, height, width, channels].
* @param output The original output image, of rank 4, of shape
* [batchSize, outHeight, outWidth, channels].
* @param filterSize The filter size: `[filterHeight, filterWidth]`. If
* `filterSize` is a single number, then `filterHeight == filterWidth`.
* @param strides The strides of the pooling: `[strideHeight, strideWidth]`. If
* `strides` is a single number, then `strideHeight == strideWidth`.
* @param pad The type of padding algorithm used in the forward prop of the op.
* 'same', 'valid', for more info, see this guide:
* [https://www.tensorflow.org/api_docs/python/tf/nn/convolution](
* https://www.tensorflow.org/api_docs/python/tf/nn/convolution)
* @param dimRoundingMode A string from: 'ceil', 'round', 'floor'. If none is
* provided, it will default to truncate.
*/
function maxPoolGrad_(dy, input, output, filterSize, strides, pad, dimRoundingMode) {
const $dy = convertToTensor(dy, 'dy', 'maxPoolGrad');
const $input = convertToTensor(input, 'input', 'maxPoolGrad');
const $output = convertToTensor(output, 'output', 'maxPoolGrad');
assert$1($input.rank === $dy.rank, () => `Rank of input (${$input.rank}) does not match rank of dy ` +
`(${$dy.rank})`);
assert$1($dy.rank === 4, () => `Error in maxPoolGrad: dy must be rank 4 but got rank ` +
`${$dy.rank}.`);
assert$1($input.rank === 4, () => `Error in maxPoolGrad: input must be rank 4 but got rank ` +
`${$input.rank}.`);
checkPadOnDimRoundingMode('maxPoolGrad', pad, dimRoundingMode);
const inputs = { dy: $dy, input: $input, output: $output };
const attrs = { filterSize, strides, pad, dimRoundingMode };
// tslint:disable-next-line: no-unnecessary-type-assertion
return ENGINE.runKernel(MaxPoolGrad, inputs, attrs);
}
const maxPoolGrad = /* @__PURE__ */ op({ maxPoolGrad_ });
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const maxPoolGradConfig = {
kernelName: MaxPool,
inputsToSave: ['x'],
outputsToSave: [true],
gradFunc: (dy, saved, attrs) => {
const [x, y] = saved;
const { filterSize, strides, pad } = attrs;
return {
x: () => maxPoolGrad(dy, x, y, filterSize, strides, pad)
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const meanGradConfig = {
kernelName: Mean,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { axis } = attrs;
const axes = parseAxisParam(axis, x.shape);
const shapes = computeOutAndReduceShapes(x.shape, axes);
const reduceShape = shapes[1];
const reduceSize = sizeFromShape(reduceShape);
const derX = () => {
const expandedDyShape = x.shape.slice();
axes.forEach(axis => {
expandedDyShape[axis] = 1;
});
const expandedDy = reshape$2(dy, expandedDyShape);
const res = div$1(mul(expandedDy, ones(x.shape, 'float32')), reduceSize);
return res;
};
return { x: derX };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const minGradConfig = {
kernelName: Min,
inputsToSave: ['x'],
outputsToSave: [true],
gradFunc: (dy, saved, attrs) => {
const minAttrs = attrs;
const { axis } = minAttrs;
const [x, y] = saved;
const origAxes = parseAxisParam(axis, x.shape);
const minGrad = gradForMinAndMax(dy, y, x, origAxes);
return {
x: () => {
return minGrad['x']();
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const minimumGradConfig = {
kernelName: Minimum,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const derA = () => mul(dy, cast$3(lessEqual$2(a, b), 'float32'));
const derB = () => mul(dy, cast$3(greater$2(a, b), 'float32'));
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const mirrorPadGradConfig = {
kernelName: MirrorPad,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
// Pad introduces values around the original tensor, so the gradient
// slices the original shape out of the gradient.
const x = saved[0];
const { paddings } = attrs;
const begin = paddings.map(p => p[0]);
return { x: () => slice$2(dy, begin, x.shape) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const modGradConfig = {
kernelName: Mod,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const outShape = assertAndGetBroadcastShape(a.shape, b.shape);
const derA = () => {
const reduceAxes = getReductionAxes(a.shape, outShape);
if (reduceAxes.length > 0) {
return reshape$2(sum$2(dy, reduceAxes), a.shape);
}
return dy;
};
const derB = () => {
const res = mul(dy, neg$2(floor$2(div$1(a, b))));
const reduceAxes = getReductionAxes(b.shape, outShape);
if (reduceAxes.length > 0) {
return reshape$2(sum$2(res, reduceAxes), b.shape);
}
return res;
};
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const multiplyGradConfig = {
kernelName: Multiply,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const outShape = assertAndGetBroadcastShape(a.shape, b.shape);
const derA = () => {
const res = mul(dy, cast$3(b, 'float32'));
const reduceAxes = getReductionAxes(a.shape, outShape);
if (reduceAxes.length > 0) {
return reshape$2(sum$2(res, reduceAxes), a.shape);
}
return res;
};
const derB = () => {
const res = mul(dy, cast$3(a, 'float32'));
const reduceAxes = getReductionAxes(b.shape, outShape);
if (reduceAxes.length > 0) {
return reshape$2(sum$2(res, reduceAxes), b.shape);
}
return res;
};
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const negGradConfig = {
kernelName: Neg,
gradFunc: (dy) => {
return { x: () => neg$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const oneHotGradConfig = {
kernelName: OneHot,
inputsToSave: ['indices'],
gradFunc: (dy, saved) => {
const indices = saved[0];
return { indices: () => zeros$1(indices.shape, 'float32') };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const onesLikeGradConfig = {
kernelName: OnesLike,
gradFunc: (dy) => {
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const packGradConfig = {
kernelName: Pack,
saveAllInputs: true,
gradFunc: (dy, saved, attrs) => {
const { axis } = attrs;
const derTensors = unstack(dy, axis);
return derTensors.map(t => () => t);
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const padV2GradConfig = {
kernelName: PadV2,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
// Pad introduces values around the original tensor, so the gradient
// slices the original shape out of the gradient.
const x = saved[0];
const { paddings } = attrs;
const begin = paddings.map(p => p[0]);
return { x: () => slice$2(dy, begin, x.shape) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const powGradConfig = {
kernelName: Pow,
inputsToSave: ['a', 'b'],
outputsToSave: [true],
gradFunc: (dy, saved) => {
const [a, b, y] = saved;
const base = a;
const exp = b;
const outShape = assertAndGetBroadcastShape(base.shape, exp.shape);
const derBase = () => {
const expFloat = cast$3(exp, 'float32');
let res = mul(dy, mul(expFloat, pow$2(base, sub$2(expFloat, scalar(1)))));
const reduceAxes = getReductionAxes(base.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, base.shape);
};
const derExp = () => {
const condition = greater$2(base, 0);
const logBase = where(condition, log$2(base), zerosLike$2(base));
let res = mul(dy, mul(y, logBase));
const reduceAxes = getReductionAxes(exp.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, exp.shape);
};
return { a: derBase, b: derExp };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const preluGradConfig = {
kernelName: Prelu,
inputsToSave: ['x', 'alpha'],
gradFunc: (dy, saved) => {
const [x, alpha] = saved;
const mask = greater$2(x, 0);
return {
x: () => where(mask, dy, mul(dy, alpha)),
alpha: () => {
let res = where(mask, zerosLike$2(dy), mul(dy, x));
const reduceAxes = getReductionAxes(alpha.shape, dy.shape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, alpha.shape);
}
};
}
};
/**
* @license
* Copyright 2022 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Gradient for product operation on a single axis.
function prodGradFn_(x, dy, axis) {
// The gradient tensor (dy) has a set of axes removed, so we create re-shaped
// versions (of size 1) for the removed axis; this supports broadcasting over
// those dimensions.
const expandedYShape = x.shape.slice();
expandedYShape[axis] = 1;
// The actual gradient computation.
const expandedDy = reshape$2(dy, expandedYShape);
const xCumProd = cumprod$2(x, axis, true, false);
const xCumRevProd = cumprod$2(x, axis, true, true);
const dx = mul(xCumProd, xCumRevProd);
return mul(expandedDy, dx);
}
// Support gradients when the product is done on many axes at once.
// This done py pushing all the axes on which the product is applied into a
// single axis.
function prodsGradFn_(x, dy, axis) {
// Move all axes for doing prod over to the end of the tensor.
const xRank = x.shape.length;
const finalProdAxis = xRank - axis.length;
const xPermutation = getAxesPermutation(axis, xRank);
let permutedX = x;
if (xPermutation != null) {
permutedX = transpose$2(x, xPermutation);
}
// Reshape all the prod dimensions into a single one, and do compute prod
// gradients on that.
const newShape = permutedX.shape.slice();
const removedShape = newShape.splice(xRank - axis.length, axis.length);
const endPartShape = removedShape.reduce((p, c) => p * c, 1);
newShape.push(endPartShape);
const reshapedPermutedX = permutedX.reshape(newShape);
let prodGrad = prodGradFn_(reshapedPermutedX, dy, finalProdAxis);
// Undo the re-shaping now we have the dx vector, and permute back to
// original axes order.
prodGrad = prodGrad.reshape(permutedX.shape);
if (xPermutation != null) {
const undoPermutation = getUndoAxesPermutation(xPermutation);
prodGrad = transpose$2(prodGrad, undoPermutation);
}
return prodGrad;
}
// Running example:
// [
// [
// [3.0, 4.0],
// [5.0, 6.0],
// [7.0, 8.0]
// ],
// [
// [3.0, 5.0],
// [0.0, 6.0],
// [5.0, 6.0]
// ]
// ]
//
const prodGradConfig = {
kernelName: Prod,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { axis } = attrs;
let axisArr = [];
if (axis === undefined || axis === null) {
axisArr = x.shape.map((_, i) => i);
}
else if (typeof axis === 'number') {
axisArr = [axis];
}
else {
axisArr = axis;
}
return { x: () => prodsGradFn_(x, dy, axisArr) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const divGradConfig = {
kernelName: RealDiv,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const outShape = assertAndGetBroadcastShape(a.shape, b.shape);
const derA = () => {
const res = div$1(dy, cast$3(b, 'float32'));
const reduceAxes = getReductionAxes(a.shape, outShape);
if (reduceAxes.length > 0) {
return reshape$2(sum$2(res, reduceAxes), a.shape);
}
return res;
};
const derB = () => {
let res = mul(dy, cast$3(a, 'float32'));
const reduceAxes = getReductionAxes(b.shape, outShape);
if (reduceAxes.length > 0) {
res = reshape$2(sum$2(res, reduceAxes), b.shape);
}
const tmp = square$2(b);
return neg$2(div$1(res, cast$3(tmp, 'float32')));
};
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const reciprocalGradConfig = {
kernelName: Reciprocal,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, neg$2(square$2(x))) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const relu6GradConfig = {
kernelName: Relu6$1,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
const mask = mul(lessEqual$2(x, 6), step$2(x));
return { x: () => mul(dy, cast$3(mask, 'float32')) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const reluGradConfig = {
kernelName: Relu$1,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(dy, cast$3(step$2(x), 'float32')) };
}
};
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const reshapeGradConfig = {
kernelName: Reshape$1,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => reshape$2(dy, x.shape) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const resizeBilinearGradConfig = {
kernelName: ResizeBilinear,
inputsToSave: ['images'],
gradFunc: (dy, saved, attrs) => {
const [images] = saved;
const inputs = { dy, images };
const imagesDer = () =>
// tslint:disable-next-line: no-unnecessary-type-assertion
ENGINE.runKernel(ResizeBilinearGrad, inputs, attrs);
return { images: imagesDer };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const resizeNearestNeighborGradConfig = {
kernelName: ResizeNearestNeighbor,
inputsToSave: ['images'],
gradFunc: (dy, saved, attrs) => {
const [images] = saved;
const inputs = { dy, images };
const imagesDer = () =>
// tslint:disable-next-line: no-unnecessary-type-assertion
ENGINE.runKernel(ResizeNearestNeighborGrad, inputs, attrs);
return { images: imagesDer };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const reverseGradConfig = {
kernelName: Reverse,
gradFunc: (dy, saved, attrs) => {
const { dims } = attrs;
const axes = parseAxisParam(dims, dy.shape);
return { x: () => reverse$2(dy, axes) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const roundGradConfig = {
kernelName: Round,
gradFunc: (dy) => {
// TODO(nsthorat): Let gradients be null for cases where we want to stop
// backpropgation.
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const rsqrtGradConfig = {
kernelName: Rsqrt,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => neg$2(div$1(dy, mul(pow$2(x, 1.5), 2))) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const selectGradConfig = {
kernelName: Select,
inputsToSave: ['condition'],
gradFunc: (dy, saved) => {
const [condition] = saved;
return {
// TODO(julianoks): Return null for condition gradient
// when backprop supports it.
condition: () => cast$3(zerosLike$2(condition), 'float32'),
t: () => mul(dy, cast$3(condition, dy.dtype)),
e: () => mul(dy, cast$3(logicalNot$2(condition), dy.dtype))
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const seluGradConfig = {
kernelName: Selu$1,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return {
x: () => {
const mask = greater$2(x, scalar(0));
const scaleAlpha = scalar(SELU_SCALEALPHA);
const scale = scalar(SELU_SCALE);
const greaterThanZeroDer = mul(dy, scale);
const lessEqualZeroDer = mul(mul(dy, scaleAlpha), exp$2(cast$3(x, 'float32')));
return where(mask, greaterThanZeroDer, lessEqualZeroDer);
}
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sigmoidGradConfig = {
kernelName: Sigmoid$1,
outputsToSave: [true],
gradFunc: (dy, saved) => {
const [y] = saved;
return { x: () => mul(dy, mul(y, sub$2(scalar(1), y))) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const signGradConfig = {
kernelName: Sign,
gradFunc: (dy) => {
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sinGradConfig = {
kernelName: Sin,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(cos$2(cast$3(x, 'float32')), dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sinhGradConfig = {
kernelName: Sinh,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(cosh$2(cast$3(x, 'float32')), dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sliceGradConfig = {
kernelName: Slice,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { begin, size } = attrs;
const inputShape = x.shape;
const [begin_, size_] = parseSliceParams(x, begin, size);
// Create an Nx2 padding where the first column represents how many
// zeros are prepended (at start) for each dimension, and the second
// column indicates how many zeros are appended (at end).
// The number of zeros to append is the shape of the input
// elementwise-subtracted by both the begin vector and sizes vector.
const paddings = [];
for (let i = 0; i < dy.rank; i++) {
paddings.push([begin_[i], inputShape[i] - begin_[i] - size_[i]]);
}
return { x: () => pad(dy, paddings) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const softmaxGradConfig = {
kernelName: Softmax$1,
outputsToSave: [true],
gradFunc: (dy, saved, attrs) => {
const [y] = saved;
const { dim } = attrs;
const keepDims = true;
const dyTimesY = mul(dy, y);
return {
logits: () => sub$2(dyTimesY, mul(sum$2(dyTimesY, [dim], keepDims), y))
};
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const softplusGradConfig = {
kernelName: Softplus$1,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(dy, sigmoid$2(x)) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const spaceToBatchNDGradConfig = {
kernelName: SpaceToBatchND,
gradFunc: (dy, saved, attrs) => {
const { blockShape, paddings } = attrs;
return { x: () => batchToSpaceND$2(dy, blockShape, paddings) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const splitVGradConfig = {
kernelName: SplitV,
gradFunc: (dy, saved, attrs) => {
const { axis } = attrs;
return { x: () => concat$2(dy, axis) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sqrtGradConfig = {
kernelName: Sqrt,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, mul(sqrt$2(cast$3(x, 'float32')), 2)) };
}
};
/**
* @license
* Copyright 2019 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const squareGradConfig = {
kernelName: Square,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => mul(dy, mul(cast$3(x, 'float32'), 2)) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const squaredDifferenceGradConfig = {
kernelName: SquaredDifference,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const two = scalar(2);
const derA = () => mul(dy, mul(two, sub$2(a, b)));
const derB = () => mul(dy, mul(two, sub$2(b, a)));
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const stepGradConfig = {
kernelName: Step,
gradFunc: (dy) => {
// TODO(manrajgrover): Return null for gradients when backprop supports
// it.
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const subGradConfig = {
kernelName: Sub,
inputsToSave: ['a', 'b'],
gradFunc: (dy, saved) => {
const [a, b] = saved;
const outShape = assertAndGetBroadcastShape(a.shape, b.shape);
const derA = () => {
let res = dy;
const reduceAxes = getReductionAxes(a.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(res, a.shape);
};
const derB = () => {
let res = dy;
const reduceAxes = getReductionAxes(b.shape, outShape);
if (reduceAxes.length > 0) {
res = sum$2(res, reduceAxes);
}
return reshape$2(neg$2(res), b.shape);
};
return { a: derA, b: derB };
}
};
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const sumGradConfig = {
kernelName: Sum,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const expandedDyShape = x.shape.slice();
const { axis } = attrs;
const axes = parseAxisParam(axis, x.shape);
axes.forEach(axis => {
expandedDyShape[axis] = 1;
});
const expandedDy = reshape$2(dy, expandedDyShape);
const derX = mul(expandedDy, ones(x.shape, 'float32'));
return { x: () => derX };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const tanGradConfig = {
kernelName: Tan,
inputsToSave: ['x'],
gradFunc: (dy, saved) => {
const [x] = saved;
return { x: () => div$1(dy, square$2(cos$2(x))) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const tanhGradConfig = {
kernelName: Tanh$1,
outputsToSave: [true],
gradFunc: (dy, saved) => {
const [y] = saved;
return { x: () => mul(sub$2(scalar(1), square$2(y)), dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const tileGradConfig = {
kernelName: Tile,
inputsToSave: ['x'],
gradFunc: (dy, saved, attrs) => {
const [x] = saved;
const { reps } = attrs;
const derX = () => {
let xGrad = zerosLike$2(x);
// TODO(cais): Maybe reduce memory footprint by avoiding repeated
// slicing.
if (x.rank === 1) {
for (let i = 0; i < reps[0]; ++i) {
xGrad = add$1(xGrad, slice$2(dy, [i * x.shape[0]], [x.shape[0]]));
}
}
else if (x.rank === 2) {
for (let i = 0; i < reps[0]; ++i) {
for (let j = 0; j < reps[1]; ++j) {
xGrad = add$1(xGrad, slice$2(dy, [i * x.shape[0], j * x.shape[1]], [
x.shape[0], x.shape[1]
]));
}
}
}
else if (x.rank === 3) {
for (let i = 0; i < reps[0]; ++i) {
for (let j = 0; j < reps[1]; ++j) {
for (let k = 0; k < reps[2]; ++k) {
xGrad =
add$1(xGrad, slice$2(dy, [i * x.shape[0], j * x.shape[1], k * x.shape[2]], [x.shape[0], x.shape[1], x.shape[2]]));
}
}
}
}
else if (x.rank === 4) {
for (let i = 0; i < reps[0]; ++i) {
for (let j = 0; j < reps[1]; ++j) {
for (let k = 0; k < reps[2]; ++k) {
for (let l = 0; l < reps[3]; ++l) {
xGrad =
add$1(xGrad, slice$2(dy, [
i * x.shape[0], j * x.shape[1], k * x.shape[2],
l * x.shape[3]
], [x.shape[0], x.shape[1], x.shape[2], x.shape[3]]));
}
}
}
}
}
else {
throw new Error(`Gradient for tile operation is not implemented for rank-` +
`${x.rank} tensors yet.`);
}
return xGrad;
};
return { x: derX };
},
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const transposeGradConfig = {
kernelName: Transpose,
gradFunc: (dy, saved, attrs) => {
const transposeAttrs = attrs;
const { perm } = transposeAttrs;
const undoPerm = getUndoAxesPermutation(perm);
return { x: () => transpose$2(dy, undoPerm) };
}
};
/**
* @license
* Copyright 2020 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const unpackGradConfig = {
kernelName: Unpack,
gradFunc: (dy, saved, attrs) => {
const unpackAttrs = attrs;
const { axis } = unpackAttrs;
return { value: () => stack(dy, axis) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const unsortedSegmentSumGradConfig = {
kernelName: UnsortedSegmentSum,
inputsToSave: ['segmentIds'],
gradFunc: (dy, saved) => {
const [segmentIds] = saved;
const derX = () => {
return gatherDropNegatives(dy, segmentIds);
};
return { x: derX };
}
};
function gatherDropNegatives(x, indices) {
// Helper function for unsorted segment ops. Gathers params for
// positive segment ids and gathers 0 for inputs with negative segment id.
// Mirrors _GatherDropNegatives from tensorflow/python/ops/math_grad.py
const zeroClippedIndices = maximum$2(indices, zerosLike$2(indices));
const gathered = gather$1(x, zeroClippedIndices);
let isPositive = greaterEqual$2(indices, scalar(0, 'int32'));
const numIters = gathered.rank - isPositive.rank;
for (let i = 0; i < numIters; ++i) {
isPositive = expandDims$3(isPositive, i + 1);
}
isPositive = logicalAnd$2(isPositive, ones(gathered.shape, 'bool'));
const zeroSlice = zerosLike$2(gathered);
return where(isPositive, gathered, zeroSlice);
}
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
const zerosLikeGradConfig = {
kernelName: ZerosLike,
gradFunc: (dy) => {
return { x: () => zerosLike$2(dy) };
}
};
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
// Export all kernel configs here so that the package can auto register them
const gradConfigs = [
absGradConfig,
acosGradConfig,
acoshGradConfig,
addGradConfig,
addNGradConfig,
argMaxGradConfig,
argMinGradConfig,
asinGradConfig,
asinhGradConfig,
atan2GradConfig,
atanGradConfig,
atanhGradConfig,
avgPool3DGradConfig,
avgPoolGradConfig,
batchMatMulGradConfig,
batchToSpaceNDGradConfig,
broadcastToGradConfig,
castGradConfig,
ceilGradConfig,
clipByValueGradConfig,
complexAbsGradConfig,
concatGradConfig,
conv2DBackpropInputGradConfig,
conv2DGradConfig,
conv3DGradConfig,
cosGradConfig,
coshGradConfig,
cumsumGradConfig,
depthwiseConv2dNativeGradConfig,
dilation2dGradConfig,
divGradConfig,
eluGradConfig,
erfGradConfig,
expGradConfig,
expandDimsGradConfig,
expm1GradConfig,
floorDivGradConfig,
floorGradConfig,
fusedBatchNormGradConfig,
gatherGradConfig,
greaterEqualGradConfig,
identityGradConfig,
isFiniteGradConfig,
isInfGradConfig,
isNanGradConfig,
leakyReluGradConfig,
log1pGradConfig,
logGradConfig,
logSoftmaxGradConfig,
lrnGradConfig,
maxGradConfig,
maxGradConfig,
maximumGradConfig,
maxPool3DGradConfig,
maxPoolGradConfig,
meanGradConfig,
minGradConfig,
minimumGradConfig,
mirrorPadGradConfig,
modGradConfig,
multiplyGradConfig,
negGradConfig,
oneHotGradConfig,
onesLikeGradConfig,
packGradConfig,
padV2GradConfig,
padV2GradConfig,
powGradConfig,
preluGradConfig,
prodGradConfig,
reciprocalGradConfig,
relu6GradConfig,
reluGradConfig,
reshapeGradConfig,
resizeBilinearGradConfig,
resizeNearestNeighborGradConfig,
reverseGradConfig,
roundGradConfig,
rsqrtGradConfig,
selectGradConfig,
seluGradConfig,
sigmoidGradConfig,
signGradConfig,
sinGradConfig,
sinhGradConfig,
sliceGradConfig,
softmaxGradConfig,
softplusGradConfig,
spaceToBatchNDGradConfig,
spaceToBatchNDGradConfig,
splitVGradConfig,
splitVGradConfig,
sqrtGradConfig,
squaredDifferenceGradConfig,
squareGradConfig,
stepGradConfig,
subGradConfig,
sumGradConfig,
tanGradConfig,
tanhGradConfig,
tileGradConfig,
transposeGradConfig,
unpackGradConfig,
unsortedSegmentSumGradConfig,
zerosLikeGradConfig
];
for (const gradientConfig of gradConfigs) {
registerGradient(gradientConfig);
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Explicit error types.
*
* See the following link for more information about why the code includes
* calls to setPrototypeOf:
*
* https://github.com/Microsoft/TypeScript-wiki/blob/master/Breaking-Changes.md#extending-built-ins-like-error-array-and-map-may-no-longer-work
*/
// tslint:enable
/**
* Equivalent of Python's AttributeError.
*/
class AttributeError extends Error {
constructor(message) {
super(message);
// Set the prototype explicitly.
Object.setPrototypeOf(this, AttributeError.prototype);
}
}
/**
* Equivalent of Python's RuntimeError.
*/
class RuntimeError extends Error {
constructor(message) {
super(message);
// Set the prototype explicitly.
Object.setPrototypeOf(this, RuntimeError.prototype);
}
}
/**
* Equivalent of Python's ValueError.
*/
class ValueError extends Error {
constructor(message) {
super(message);
// Set the prototype explicitly.
Object.setPrototypeOf(this, ValueError.prototype);
}
}
/**
* Equivalent of Python's NotImplementedError.
*/
class NotImplementedError extends Error {
constructor(message) {
super(message);
// Set the prototype explicitly.
Object.setPrototypeOf(this, NotImplementedError.prototype);
}
}
/**
* Equivalent of Python's AssertionError.
*/
class AssertionError extends Error {
constructor(message) {
super(message);
// Set the prototype explicitly.
Object.setPrototypeOf(this, AssertionError.prototype);
}
}
/**
* @license
* Copyright 2022 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* LruCache: A mapping from the String to T. If the number of the entries is
* exceeding the `maxEntries`, the LruCache will delete the least recently
* used entry.
*/
class LruCache {
constructor(maxEntries) {
this.maxEntries = maxEntries || 100;
this.cache = new Map();
}
/**
* Get the entry for the key and mark it as used recently.
*/
get(key) {
let entry;
if (this.cache.has(key)) {
entry = this.cache.get(key);
this.cache.delete(key);
this.cache.set(key, entry);
}
return entry;
}
/**
* Put the entry into the cache. If the key already existed, mark the key as
* used recently.
*/
put(key, value) {
if (this.cache.has(key)) {
this.cache.delete(key);
}
else if (this.cache.size >= this.maxEntries) {
const keyToDelete = this.cache.keys().next().value;
this.cache.delete(keyToDelete);
}
this.cache.set(key, value);
}
/**
* Get the MaxEntries of the cache.
*/
getMaxEntries() {
return this.maxEntries;
}
/**
* Set the MaxEntries of the cache. If the maxEntries is decreased, reduce
* entries in the cache.
*/
setMaxEntries(maxEntries) {
if (maxEntries < 0) {
throw new Error(`The maxEntries of LRU caches must be at least 0, but got ${maxEntries}.`);
}
if (this.maxEntries > maxEntries) {
for (let i = 0; i < this.maxEntries - maxEntries; i++) {
const keyToDelete = this.cache.keys().next().value;
this.cache.delete(keyToDelete);
}
}
this.maxEntries = maxEntries;
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original source: utils/generic_utils.py */
// tslint:enable
/**
* If `value` is an Array, equivalent to Python's `value * numValues`.
* If `value` is not an Array, equivalent to Python's `[value] * numValues`
*/
// tslint:disable-next-line:no-any
function pyListRepeat(value, numValues) {
if (Array.isArray(value)) {
// tslint:disable-next-line:no-any
let newArray = [];
for (let i = 0; i < numValues; i++) {
newArray = newArray.concat(value);
}
return newArray;
}
else {
const newArray = new Array(numValues);
newArray.fill(value);
return newArray;
}
}
function assert(val, message) {
if (!val) {
throw new AssertionError(message);
}
}
/**
* Count the number of elements of the `array` that are equal to `reference`.
*/
function count(array, refernce) {
let counter = 0;
for (const item of array) {
if (item === refernce) {
counter++;
}
}
return counter;
}
/**
* If an array is of length 1, just return the first element. Otherwise, return
* the full array.
* @param tensors
*/
function singletonOrArray(xs) {
if (xs.length === 1) {
return xs[0];
}
return xs;
}
/**
* Normalizes a list/tensor into a list.
*
* If a tensor is passed, we return
* a list of size 1 containing the tensor.
*
* @param x target object to be normalized.
*/
// tslint:disable-next-line:no-any
function toList(x) {
if (Array.isArray(x)) {
return x;
}
return [x];
}
/**
* Converts string to snake-case.
* @param name
*/
function toSnakeCase(name) {
const intermediate = name.replace(/(.)([A-Z][a-z0-9]+)/g, '$1_$2');
const insecure = intermediate.replace(/([a-z])([A-Z])/g, '$1_$2').toLowerCase();
/*
If the class is private the name starts with "_" which is not secure
for creating scopes. We prefix the name with "private" in this case.
*/
if (insecure[0] !== '_') {
return insecure;
}
return 'private' + insecure;
}
function toCamelCase(identifier) {
// quick return for empty string or single character strings
if (identifier.length <= 1) {
return identifier;
}
// Check for the underscore indicating snake_case
if (identifier.indexOf('_') === -1) {
return identifier;
}
return identifier.replace(/[_]+(\w|$)/g, (m, p1) => p1.toUpperCase());
}
// tslint:disable-next-line:no-any
let _GLOBAL_CUSTOM_OBJECTS = {};
function serializeKerasObject(instance) {
if (instance === null || instance === undefined) {
return null;
}
const dict = {};
dict['className'] = instance.getClassName();
dict['config'] = instance.getConfig();
return dict;
}
/**
* Replace ndarray-style scalar objects in serialization objects with numbers.
*
* Background: In some versions of tf.keras, certain scalar values in the HDF5
* model save file can be serialized as: `{'type': 'ndarray', 'value': num}`,
* where in `num` is a plain number. This method converts such serialization
* to a `number`.
*
* @param config The keras-format serialization object to be processed
* (in place).
*/
function convertNDArrayScalarsInConfig(config) {
if (config == null || typeof config !== 'object') {
return;
}
else if (Array.isArray(config)) {
config.forEach(configItem => convertNDArrayScalarsInConfig(configItem));
}
else {
const fields = Object.keys(config);
for (const field of fields) {
const value = config[field];
if (value != null && typeof value === 'object') {
if (!Array.isArray(value) && value['type'] === 'ndarray' &&
typeof value['value'] === 'number') {
config[field] = value['value'];
}
else {
convertNDArrayScalarsInConfig(value);
}
}
}
}
}
/**
* Deserialize a saved Keras Object
* @param identifier either a string ID or a saved Keras dictionary
* @param moduleObjects a list of Python class names to object constructors
* @param customObjects a list of Python class names to object constructors
* @param printableModuleName debug text for the object being reconstituted
* @param fastWeightInit Optional flag to use fast weight initialization
* during deserialization. This is applicable to cases in which
* the initialization will be immediately overwritten by loaded weight
* values. Default: `false`.
* @returns a TensorFlow.js Layers object
*/
// tslint:disable:no-any
function deserializeKerasObject(identifier, moduleObjects = {}, customObjects = {}, printableModuleName = 'object', fastWeightInit = false) {
// tslint:enable
if (typeof identifier === 'string') {
const functionName = identifier;
let fn;
if (functionName in customObjects) {
fn = customObjects[functionName];
}
else if (functionName in _GLOBAL_CUSTOM_OBJECTS) {
fn = _GLOBAL_CUSTOM_OBJECTS[functionName];
}
else {
fn = moduleObjects[functionName];
if (fn == null) {
throw new ValueError(`Unknown ${printableModuleName}: ${identifier}. ` +
`This may be due to one of the following reasons:\n` +
`1. The ${printableModuleName} is defined in Python, in which ` +
`case it needs to be ported to TensorFlow.js or your JavaScript ` +
`code.\n` +
`2. The custom ${printableModuleName} is defined in JavaScript, ` +
`but is not registered properly with ` +
`tf.serialization.registerClass().`);
// TODO(cais): Add link to tutorial page on custom layers.
}
}
return fn;
}
else {
// In this case we are dealing with a Keras config dictionary.
const config = identifier;
if (config['className'] == null || config['config'] == null) {
throw new ValueError(`${printableModuleName}: Improper config format: ` +
`${JSON.stringify(config)}.\n` +
`'className' and 'config' must set.`);
}
const className = config['className'];
let cls, fromConfig;
if (className in customObjects) {
[cls, fromConfig] = customObjects[className];
}
else if (className in _GLOBAL_CUSTOM_OBJECTS) {
[cls, fromConfig] = _GLOBAL_CUSTOM_OBJECTS['className'];
}
else if (className in moduleObjects) {
[cls, fromConfig] = moduleObjects[className];
}
if (cls == null) {
throw new ValueError(`Unknown ${printableModuleName}: ${className}. ` +
`This may be due to one of the following reasons:\n` +
`1. The ${printableModuleName} is defined in Python, in which ` +
`case it needs to be ported to TensorFlow.js or your JavaScript ` +
`code.\n` +
`2. The custom ${printableModuleName} is defined in JavaScript, ` +
`but is not registered properly with ` +
`tf.serialization.registerClass().`);
// TODO(cais): Add link to tutorial page on custom layers.
}
if (fromConfig != null) {
// Porting notes: Instead of checking to see whether fromConfig accepts
// customObjects, we create a customObjects dictionary and tack it on to
// config['config'] as config['config'].customObjects. Objects can use it,
// if they want.
// tslint:disable-next-line:no-any
const customObjectsCombined = {};
for (const key of Object.keys(_GLOBAL_CUSTOM_OBJECTS)) {
customObjectsCombined[key] = _GLOBAL_CUSTOM_OBJECTS[key];
}
for (const key of Object.keys(customObjects)) {
customObjectsCombined[key] = customObjects[key];
}
// Add the customObjects to config
const nestedConfig = config['config'];
nestedConfig['customObjects'] = customObjectsCombined;
const backupCustomObjects = Object.assign({}, _GLOBAL_CUSTOM_OBJECTS);
for (const key of Object.keys(customObjects)) {
_GLOBAL_CUSTOM_OBJECTS[key] = customObjects[key];
}
convertNDArrayScalarsInConfig(config['config']);
const returnObj = fromConfig(cls, config['config'], customObjects, fastWeightInit);
_GLOBAL_CUSTOM_OBJECTS = Object.assign({}, backupCustomObjects);
return returnObj;
}
else {
// Then `cls` may be a function returning a class.
// In this case by convention `config` holds
// the kwargs of the function.
const backupCustomObjects = Object.assign({}, _GLOBAL_CUSTOM_OBJECTS);
for (const key of Object.keys(customObjects)) {
_GLOBAL_CUSTOM_OBJECTS[key] = customObjects[key];
}
// In python this is **config['config'], for tfjs-layers we require
// classes that use this fall-through construction method to take
// a config interface that mimics the expansion of named parameters.
const returnObj = new cls(config['config']);
_GLOBAL_CUSTOM_OBJECTS = Object.assign({}, backupCustomObjects);
return returnObj;
}
}
}
/**
* Compares two numbers for sorting.
* @param a
* @param b
*/
function numberCompare(a, b) {
return (a < b) ? -1 : ((a > b) ? 1 : 0);
}
/**
* Comparison of two numbers for reverse sorting.
* @param a
* @param b
*/
function reverseNumberCompare(a, b) {
return -1 * numberCompare(a, b);
}
/**
* Get the unique elements of an array.
* @param xs Array.
* @returns An Array consisting of the unique elements in `xs`.
*/
function unique(xs) {
if (xs == null) {
return xs;
}
const out = [];
// TODO(cais): Maybe improve performance by sorting.
for (const x of xs) {
if (out.indexOf(x) === -1) {
out.push(x);
}
}
return out;
}
/**
* Determine if an Object is empty (i.e., does not have own properties).
* @param obj Object
* @returns Whether the Object is empty.
* @throws ValueError: If object is `null` or `undefined`.
*/
function isObjectEmpty(obj) {
if (obj == null) {
throw new ValueError(`Invalid value in obj: ${JSON.stringify(obj)}`);
}
for (const key in obj) {
if (obj.hasOwnProperty(key)) {
return false;
}
}
return true;
}
/**
* Helper function used to build type union/enum run-time checkers.
* @param values The list of allowed values.
* @param label A string name for the type
* @param value The value to test.
* @throws ValueError: If the value is not in values nor `undefined`/`null`.
*/
function checkStringTypeUnionValue(values, label, value) {
if (value == null) {
return;
}
if (values.indexOf(value) < 0) {
throw new ValueError(`${value} is not a valid ${label}. Valid values are ${values} or null/undefined.`);
}
}
// tslint:enable:no-any
/**
* Assert that a value or an array of value are positive integer.
*
* @param value The value being asserted on. May be a single number or an array
* of numbers.
* @param name Name of the value, used to make the error message.
*/
function assertPositiveInteger(value, name) {
if (Array.isArray(value)) {
assert$1(value.length > 0, () => `${name} is unexpectedly an empty array.`);
value.forEach((v, i) => assertPositiveInteger(v, `element ${i + 1} of ${name}`));
}
else {
assert$1(Number.isInteger(value) && value > 0, () => `Expected ${name} to be a positive integer, but got ` +
`${formatAsFriendlyString(value)}.`);
}
}
/**
* Format a value into a display-friendly, human-readable fashion.
*
* - `null` is formatted as `'null'`
* - Strings are formated with flanking pair of quotes.
* - Arrays are formatted with flanking pair of square brackets.
*
* @param value The value to display.
* @return Formatted string.
*/
// tslint:disable-next-line:no-any
function formatAsFriendlyString(value) {
if (value === null) {
return 'null';
}
else if (Array.isArray(value)) {
return '[' + value.map(v => formatAsFriendlyString(v)).join(',') + ']';
}
else if (typeof value === 'string') {
return `"${value}"`;
}
else {
return `${value}`;
}
}
/**
* Returns a function `f2` (decorator) which wraps the original function
* `f`. `f2` guarantees that `f` can be called at most once
* every `waitMs` ms. If `f2` is called more often, it will return
* the last returned result of `f`.
*
* @param f The original function `f` to wrap.
* @param waitMs The time between two consecutive calls to `f` in ms.
*/
function debounce(f, waitMs, nowFunc) {
let lastTime = nowFunc != null ? nowFunc() : now();
let lastResult;
const f2 = (...args) => {
const now$1 = nowFunc != null ? nowFunc() : now();
if (now$1 - lastTime < waitMs) {
return lastResult;
}
lastTime = now$1;
lastResult = f(...args);
return lastResult;
};
return f2;
}
/**
* Returns the fusable activation given a layers identifier.
*
* @param activationName The layers identifier string.
* @return The name of the fusable activation.
*/
function mapActivationToFusedKernel(activationName) {
if (activationName === 'relu') {
return 'relu';
}
if (activationName === 'linear') {
return 'linear';
}
if (activationName === 'elu') {
return 'elu';
}
return null;
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Utilities related to persistent state in the backend.
*/
/**
* An ID to track `tf.SymbolicTensor`s and derived classes.
* Required in different places in engine/topology.ts to identify unique
* tensors.
*/
let _nextUniqueTensorId = 0;
function getNextUniqueTensorId() {
return _nextUniqueTensorId++;
}
const _uidPrefixes = {};
/**
* Provides a unique UID given a string prefix.
*
* @param prefix
*/
function getUid(prefix = '') {
if (!(prefix in _uidPrefixes)) {
_uidPrefixes[prefix] = 0;
}
_uidPrefixes[prefix] += 1;
return prefix + _uidPrefixes[prefix].toString();
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
const VALID_DATA_FORMAT_VALUES = ['channelsFirst', 'channelsLast'];
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Common functions for TensorFlow.js Layers.
*/
// A map from the requested scoped name of a Tensor to the number of Tensors
// wanting that name so far. This allows enforcing name uniqueness by appending
// an incrementing index, e.g. scope/name, scope/name_1, scope/name_2, etc.
const nameMap = new Map();
function checkDataFormat(value) {
checkStringTypeUnionValue(VALID_DATA_FORMAT_VALUES, 'DataFormat', value);
}
const _nameScopeStack = [];
const _nameScopeDivider = '/';
/**
* Enter namescope, which can be nested.
*/
function nameScope(name, fn) {
_nameScopeStack.push(name);
try {
const val = fn();
_nameScopeStack.pop();
return val;
}
catch (e) {
_nameScopeStack.pop();
throw e;
}
}
/**
* Get the current namescope as a flat, concatenated string.
*/
function currentNameScopePrefix() {
if (_nameScopeStack.length === 0) {
return '';
}
else {
return _nameScopeStack.join(_nameScopeDivider) + _nameScopeDivider;
}
}
/**
* Get the name a Tensor (or Variable) would have if not uniqueified.
* @param tensorName
* @return Scoped name string.
*/
function getScopedTensorName(tensorName) {
if (!isValidTensorName(tensorName)) {
throw new Error('Not a valid tensor name: \'' + tensorName + '\'');
}
return currentNameScopePrefix() + tensorName;
}
/**
* Get unique names for Tensors and Variables.
* @param scopedName The fully-qualified name of the Tensor, i.e. as produced by
* `getScopedTensorName()`.
* @return A unique version of the given fully scoped name.
* If this is the first time that the scoped name is seen in this session,
* then the given `scopedName` is returned unaltered. If the same name is
* seen again (producing a collision), an incrementing suffix is added to the
* end of the name, so it takes the form 'scope/name_1', 'scope/name_2', etc.
*/
function getUniqueTensorName(scopedName) {
if (!isValidTensorName(scopedName)) {
throw new Error('Not a valid tensor name: \'' + scopedName + '\'');
}
if (!nameMap.has(scopedName)) {
nameMap.set(scopedName, 0);
}
const index = nameMap.get(scopedName);
nameMap.set(scopedName, nameMap.get(scopedName) + 1);
if (index > 0) {
const result = `${scopedName}_${index}`;
// Mark the composed name as used in case someone wants
// to call getUniqueTensorName("name_1").
nameMap.set(result, 1);
return result;
}
else {
return scopedName;
}
}
const tensorNameRegex = new RegExp(/^[A-Za-z0-9][-A-Za-z0-9\._\/]*$/);
/**
* Determine whether a string is a valid tensor name.
* @param name
* @returns A Boolean indicating whether `name` is a valid tensor name.
*/
function isValidTensorName(name) {
return !!name.match(tensorNameRegex);
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Math utility functions.
*
* This file contains some frequently used math function that operates on
* number[] or Float32Array and return a number. Many of these functions are
* not-so-thick wrappers around TF.js Core functions. But they offer the
* convenience of
* 1) not having to convert the inputs into Tensors,
* 2) not having to convert the returned Tensors to numbers.
*/
/**
* Calculate the product of an array of numbers.
* @param array The array to calculate the product over.
* @param begin Beginning index, inclusive.
* @param end Ending index, exclusive.
* @return The product.
*/
function arrayProd(array, begin, end) {
if (begin == null) {
begin = 0;
}
if (end == null) {
end = array.length;
}
let prod = 1;
for (let i = begin; i < end; ++i) {
prod *= array[i];
}
return prod;
}
/**
* Generate an array of integers in [begin, end).
* @param begin Beginning integer, inclusive.
* @param end Ending integer, exclusive.
* @returns Range array.
* @throws ValueError, iff `end` < `begin`.
*/
function range(begin, end) {
if (end < begin) {
throw new ValueError(`end (${end}) < begin (${begin}) is forbidden.`);
}
const out = [];
for (let i = begin; i < end; ++i) {
out.push(i);
}
return out;
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
let _epsilon;
/**
* Returns the value of the fuzz factor used in numeric expressions.
*/
function epsilon() {
if (_epsilon == null) {
_epsilon = backend().epsilon();
}
return _epsilon;
}
/**
* Returns the default image data format convention.
*/
function imageDataFormat() {
return 'channelsLast';
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* deeplearn.js backend.
*/
/**
* Casts a tensor to a different dtype and returns it.
* @param x Input tensor.
* @param dtype String: 'float32'|'int32'|'bool'.
* @returns Tensor of the specified `dtype`.
*/
function cast(x, dtype) {
return cast$3(x, dtype);
}
/**
* Adds a 1-sized dimension at index "axis".
* @param x Input tensor.
* @param axis Position where to add the new axis.
* @returns Result of the dimension expansion.
*/
function expandDims(x, axis = -1) {
const outShape = x.shape.slice();
if (axis < 0) {
axis = outShape.length + axis + 1;
}
outShape.splice(axis, 0, 1);
return reshape$2(x, outShape);
}
/**
* Repeats a 2D tensor.
*
* If `x` has shape `[samples, dim]` and `n` is 2, for example, the output
* will have shape `[samples, 2, dim]`.
*
* @param x Input tensor.
* @param n Integer, number of times to repeat.
* @returns The result of the repeat operation.
* @throws ValueError: If input tensor is not 2D.
*/
function repeat(x, n) {
return tidy(() => {
if (x.shape.length !== 2) {
throw new ValueError(`repeat() expects a rank-2 tensor, but received a ` +
`rank-${x.shape.length} tensor.`);
}
const y = expandDims(x, 1);
return tile(y, [1, n, 1]);
});
}
/**
* Flatten a Tensor into 1D.
* @param x Input tensor.
* @return The result of the flattening `x`.
*/
function flatten(x) {
const newShape = [arrayProd(x.shape)];
return reshape$2(x, newShape);
}
/**
* Turn a nD tensor into a 2D tensor with same 0th dimension.
* In other words, it flattens each data samples of a batch.
*
* @param x The tensor to flatten. The rank of this tensor is required to be 2
* or higher.
* @return The result of the flattening.
*/
function batchFlatten(x) {
if (x.rank <= 1) {
throw new ValueError(`batchFlatten requires a minimum rank of 2. Got rank: ${x.rank}.`);
}
const newShape = [x.shape[0], arrayProd(x.shape, 1)];
return reshape$2(x, newShape);
}
/**
* Do slicing along the first axis.
* @param array input `tf.Tensor`.
* @param start starting index, inclusive.
* @param size size of the slice along the first axis.
* @returns result of the slicing.
* @throws ValueError: If `array` is of an unsupported subtype of `tf.Tensor`.
*/
function sliceAlongFirstAxis(array, start, size) {
return tidy(() => {
switch (array.rank) {
case 1:
return slice1d(array, start, size);
case 2:
return slice2d(array, [start, 0], [size, array.shape[1]]);
case 3:
return slice3d(array, [start, 0, 0], [size, array.shape[1], array.shape[2]]);
case 4:
return slice4d(array, [start, 0, 0, 0], [size, array.shape[1], array.shape[2], array.shape[3]]);
case 5:
return slice$2(array, [start, 0, 0, 0, 0], [
size, array.shape[1], array.shape[2], array.shape[3], array.shape[4]
]);
case 6:
return slice$2(array, [start, 0, 0, 0, 0, 0], [
size, array.shape[1], array.shape[2], array.shape[3], array.shape[4],
array.shape[5]
]);
default:
throw new ValueError(`sliceAlongFirstAxis() received an unsupported tensor rank: ` +
`${array.rank}`);
}
});
}
/**
* Creates a tensor by tiling `x` by `n`.
* @param x A tensor.
* @param n An Array of integers or a single integer. If an Array, the length
* must be the same as the number of dimensions in `x`. If a single integer,
* it will be treated as an Array of length 1.
*/
function tile(x, n) {
if (!Array.isArray(n)) {
n = [n];
}
if (x.rank !== n.length) {
throw new ValueError(`The length of input n (${n.length}) does not match ` +
`the number of dimensions in input x (${x.rank})`);
}
return tile$3(x, n);
}
/* Creation of random tensors. */
/**
* Get a tensor with normal distribution of values.
*
* @param shape Shape of the tensor.
* @param mean mean value of the normal distribution.
* @param stddev standard deviation of the normal distribution.
* @param dtype
* @param seed
* @return The normal tensor.
*/
function randomNormal(shape, mean = 0.0, stddev = 1.0, dtype, seed) {
return randomNormal$1(shape, mean, stddev, dtype, seed);
}
/* Linear Algebra */
/**
* Multiply two tensors and returns the result as a tensor.
*
* For 2D tensors, this is equivalent to matrix multiplication (matMul).
* For tensors of higher ranks, it follows the Theano behavior,
* (e.g. `(2, 3) * (4, 3, 5) -> (2, 4, 5)`). From the Theano documentation:
*
* For N dimensions it is a sum product over the last axis of x and the
* second-to-last of y:
*
* @param a A tensor of at least rank 2.
* @param b A tensor of at least rank 2.
* @param activation (optional) A string identifying the activation
* function.
* @return Result of the dot operation.
*/
function dot(a, b, activation, bias) {
if ((a.rank < 2) || (b.rank < 2)) {
throw new NotImplementedError(`dot requires both inputs to be rank >= 2` +
` but got x shape = ${a.shape} and y shape = ${b.shape}`);
}
if (b.rank >= 3) {
const xLastDim = a.shape.slice(-1)[0];
const ySecondLastDim = b.shape.slice(-2)[0];
if (xLastDim !== ySecondLastDim) {
throw new NotImplementedError(`If rank y >= 3, then the second last dim` +
` of y must equal the last dim of x but got x shape = ${a.shape} and ` +
` y shape = ${b.shape}`);
}
}
// Handle basic 2D x 2D case.
if ((a.rank === 2) && (b.rank === 2)) {
const transposeA = false;
const transposeB = false;
// tfc.fused.matMul only fuses certain activation functions. Unsupported
// activation functions are treated as 'linear' activations, which is
// equivalent to a no-op.
return matMul({
a,
b: b,
transposeA,
transposeB,
bias: bias ? reshapeBias(a.rank, bias, imageDataFormat()) : null,
activation
});
}
else {
// Reshape x into the analogous 2D Tensor.
const aFirstDims = a.shape.slice(); // Holds all but the last dim of x.
const aLastDim = aFirstDims.pop();
a = reshape$2(a, [-1, aLastDim]);
// Reshape y into the analogous 2D Tensor, and keep track of the
// required dimensions to reproduce the output shape.
const bShape = b.shape.slice();
const bLastDim = bShape.pop();
const ySecondLastDim = bShape.pop();
const yOtherDims = [...bShape, bLastDim];
// permutation should be like [r-2, 0, 1, 2, ... r-4, r-3, r-1]
// where r is the rank of y.
const perm = Array.from({ length: b.rank }, (_, i) => {
if (i === 0) {
return b.rank - 2;
}
else if (i <= b.rank - 2) {
return i - 1;
}
return i;
});
b = reshape$2(transpose$2(b, perm), [ySecondLastDim, -1]);
// Multiply x and y as 2D Tensors, and then reshape back to original.
const outputShape = [...aFirstDims, ...yOtherDims];
const transposeA = false;
const transposeB = false;
return reshape$2(matMul({
a,
b,
transposeA,
transposeB,
bias: bias ? reshapeBias(a.rank, bias, imageDataFormat()) : null,
activation
}), outputShape);
}
}
/* Elementary math functions. */
/**
* Retrieves the elements of indices `indices` in the tensor `reference`.
* @param reference A tensor.
* @param indices An integer tensor of indices or an `Array` of integers.
* @param axis Axis along which to perform the gather operation.
* @returns The result of the gathering as a tensor.
*/
function gather(reference, indices, axis) {
return tidy(() => {
if (Array.isArray(indices)) {
indices = tensor1d(indices, 'int32');
}
else {
indices = cast$3(indices, 'int32');
}
return gather$1(reference, indices, axis);
});
}
/**
* Element-wise square.
* @param x Input tensor.
* @return element-wise x^2
*/
function square(x) {
return mul(x, x);
}
/**
* Reshapes bias tensor according to rank of x.
*/
function reshapeBias(xRank, bias, dataFormat) {
const biasShape = bias.shape;
if (bias.rank !== 1 && bias.rank !== xRank) {
throw new ValueError(`Unexpected bias dimensions: ${bias.rank}` +
`; expected it to be 1 or ${xRank}`);
}
if (xRank === 5) {
if (dataFormat === 'channelsFirst') {
if (biasShape.length === 1) {
return reshape$2(bias, [1, biasShape[0], 1, 1, 1]);
}
else {
return reshape$2(bias, [1, biasShape[3], biasShape[0], biasShape[1], biasShape[2]]);
}
}
else if (dataFormat === 'channelsLast') {
if (biasShape.length === 1) {
return reshape$2(bias, [1, 1, 1, 1, biasShape[0]]);
}
else {
return reshape$2(bias, [1].concat(biasShape));
}
}
}
else if (xRank === 4) {
if (dataFormat === 'channelsFirst') {
if (biasShape.length === 1) {
return reshape$2(bias, [1, biasShape[0], 1, 1]);
}
else {
return reshape$2(bias, [1, biasShape[2], biasShape[0], biasShape[1]]);
}
}
else if (dataFormat === 'channelsLast') {
if (biasShape.length === 1) {
return reshape$2(bias, [1, 1, 1, biasShape[0]]);
}
else {
return reshape$2(bias, [1].concat(biasShape));
}
}
}
else if (xRank === 3) {
if (dataFormat === 'channelsFirst') {
if (biasShape.length === 1) {
return reshape$2(bias, [1, biasShape[0], 1]);
}
else {
return reshape$2(bias, [1, biasShape[1], biasShape[0]]);
}
}
else if (dataFormat === 'channelsLast') {
if (biasShape.length === 1) {
return reshape$2(bias, [1, 1, biasShape[0]]);
}
else {
return reshape$2(bias, [1].concat(biasShape));
}
}
}
else if (xRank < 3) {
return bias;
}
throw new ValueError(`Unsupported input rank by biasAdd: ${bias.rank}`);
}
/* Neural-network operations. */
/**
* Add a bias to a tensor.
*
* @param x The tensor to add the bias to.
* @param bias The bias to add to `x`. Must be 1D or the same rank as `x`.
* @return Result of the bias adding.
* @throws ValueError: If the rank of `bias` is incorrect.
*/
function biasAdd(x, bias, dataFormat) {
return tidy(() => {
if (dataFormat == null) {
dataFormat = imageDataFormat();
}
checkDataFormat(dataFormat);
return add$1(x, reshapeBias(x.rank, bias, dataFormat));
});
}
/**
* Exponential linear unit (ELU).
* @param x A tensor or variable to compute the activation function for.
* @param alpha: A scalar, a scaling factor for the negative section.
* @return Output of the ELU operation.
*/
function elu(x, alpha = 1) {
// TODO(cais): Add support for alpha values other than 1.
if (alpha !== 1) {
throw new NotImplementedError(`Support for alpha values other than 1 (${alpha}) is not implemented ` +
`yet.`);
}
return elu$3(x);
}
/**
* Softsign of a tensor.
*
* Defined as x / (abs(x) + 1), element-wise.
*
* @param x: Input.
* @returns Output.
*/
function softsign(x) {
return tidy(() => div$1(x, add$1(abs$2(x), 1)));
}
/**
* Sets entries in `x` to zero at random, while scaling the entire tensor.
*
* @param x input tensor.
* @param level fraction of the entries in the tensor that will be set to 0.
* @param noiseShape shape of randomly generated keep/drop flags, must be
* broadcastable to the shape of `x`. Optional.
* @param seed random seed to ensure determinism. Optional.
* @returns Result of the dropout operation.
*/
function dropout$1(x, level, noiseShape, seed) {
return tidy(() => dropout$2(x, level, noiseShape, seed));
}
/**
* Element-wise, segment-wise linear approximation of sigmoid.
*
* Returns `0.` if `x < -2.5`, `1.` if `x > 2.5`.
* In `-2.5 <= x <= 2.5`, returns `0.2 * x + 0.5`.
*
* @param x Input tensor.
* @returns Output tensor.
*/
function hardSigmoid(x) {
return tidy(() => {
const y = add$1(.5, mul(.2, x));
return clipByValue$2(y, 0, 1);
});
}
/**
* Invoke `x` in the training phase, and `alt` otherwise.
*
* Porting Note: We do not create placeholder tensors for the `training`
* boolean flag here, because there is no such thing in the TF.js imperative
* backend.
*
* @param x The function to invoke iff `training` is `true`.
* @param alt The function to invoke iff `training` is `false`.
* @param training Boolean flag for whether training phase is active.
* @returns The return value of `x()` if `training` is `true`, or the return
* value of `alt()` if `training` is `false`.
*/
function inTrainPhase(x, alt, training = false) {
return training ? x() : alt();
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
const VALID_FAN_MODE_VALUES = ['fanIn', 'fanOut', 'fanAvg'];
const VALID_DISTRIBUTION_VALUES = ['normal', 'uniform', 'truncatedNormal'];
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
function checkFanMode(value) {
checkStringTypeUnionValue(VALID_FAN_MODE_VALUES, 'FanMode', value);
}
function checkDistribution(value) {
checkStringTypeUnionValue(VALID_DISTRIBUTION_VALUES, 'Distribution', value);
}
/**
* Initializer base class.
*
* @doc {
* heading: 'Initializers', subheading: 'Classes', namespace: 'initializers'}
*/
class Initializer extends Serializable {
fromConfigUsesCustomObjects() {
return false;
}
getConfig() {
return {};
}
}
class Zeros extends Initializer {
apply(shape, dtype) {
return zeros$1(shape, dtype);
}
}
/** @nocollapse */
Zeros.className = 'Zeros';
registerClass(Zeros);
class Ones extends Initializer {
apply(shape, dtype) {
return ones(shape, dtype);
}
}
/** @nocollapse */
Ones.className = 'Ones';
registerClass(Ones);
class Constant extends Initializer {
constructor(args) {
super();
if (typeof args !== 'object') {
throw new ValueError(`Expected argument of type ConstantConfig but got ${args}`);
}
if (args.value === undefined) {
throw new ValueError(`config must have value set but got ${args}`);
}
this.value = args.value;
}
apply(shape, dtype) {
return tidy(() => mul(scalar(this.value), ones(shape, dtype)));
}
getConfig() {
return {
value: this.value,
};
}
}
/** @nocollapse */
Constant.className = 'Constant';
registerClass(Constant);
class RandomUniform extends Initializer {
constructor(args) {
super();
this.DEFAULT_MINVAL = -0.05;
this.DEFAULT_MAXVAL = 0.05;
this.minval = args.minval || this.DEFAULT_MINVAL;
this.maxval = args.maxval || this.DEFAULT_MAXVAL;
this.seed = args.seed;
}
apply(shape, dtype) {
return randomUniform(shape, this.minval, this.maxval, dtype, this.seed);
}
getConfig() {
return { minval: this.minval, maxval: this.maxval, seed: this.seed };
}
}
/** @nocollapse */
RandomUniform.className = 'RandomUniform';
registerClass(RandomUniform);
class RandomNormal extends Initializer {
constructor(args) {
super();
this.DEFAULT_MEAN = 0.;
this.DEFAULT_STDDEV = 0.05;
this.mean = args.mean || this.DEFAULT_MEAN;
this.stddev = args.stddev || this.DEFAULT_STDDEV;
this.seed = args.seed;
}
apply(shape, dtype) {
dtype = dtype || 'float32';
if (dtype !== 'float32' && dtype !== 'int32') {
throw new NotImplementedError(`randomNormal does not support dType ${dtype}.`);
}
return randomNormal(shape, this.mean, this.stddev, dtype, this.seed);
}
getConfig() {
return { mean: this.mean, stddev: this.stddev, seed: this.seed };
}
}
/** @nocollapse */
RandomNormal.className = 'RandomNormal';
registerClass(RandomNormal);
class TruncatedNormal extends Initializer {
constructor(args) {
super();
this.DEFAULT_MEAN = 0.;
this.DEFAULT_STDDEV = 0.05;
this.mean = args.mean || this.DEFAULT_MEAN;
this.stddev = args.stddev || this.DEFAULT_STDDEV;
this.seed = args.seed;
}
apply(shape, dtype) {
dtype = dtype || 'float32';
if (dtype !== 'float32' && dtype !== 'int32') {
throw new NotImplementedError(`truncatedNormal does not support dType ${dtype}.`);
}
return truncatedNormal(shape, this.mean, this.stddev, dtype, this.seed);
}
getConfig() {
return { mean: this.mean, stddev: this.stddev, seed: this.seed };
}
}
/** @nocollapse */
TruncatedNormal.className = 'TruncatedNormal';
registerClass(TruncatedNormal);
class Identity extends Initializer {
constructor(args) {
super();
this.gain = args.gain != null ? args.gain : 1.0;
}
apply(shape, dtype) {
return tidy(() => {
if (shape.length !== 2 || shape[0] !== shape[1]) {
throw new ValueError('Identity matrix initializer can only be used for' +
' 2D square matrices.');
}
else {
return mul(this.gain, eye(shape[0]));
}
});
}
getConfig() {
return { gain: this.gain };
}
}
/** @nocollapse */
Identity.className = 'Identity';
registerClass(Identity);
/**
* Computes the number of input and output units for a weight shape.
* @param shape Shape of weight.
* @param dataFormat data format to use for convolution kernels.
* Note that all kernels in Keras are standardized on the
* CHANNEL_LAST ordering (even when inputs are set to CHANNEL_FIRST).
* @return An length-2 array: fanIn, fanOut.
*/
function computeFans(shape, dataFormat = 'channelsLast') {
let fanIn;
let fanOut;
checkDataFormat(dataFormat);
if (shape.length === 2) {
fanIn = shape[0];
fanOut = shape[1];
}
else if ([3, 4, 5].indexOf(shape.length) !== -1) {
if (dataFormat === 'channelsFirst') {
const receptiveFieldSize = arrayProd(shape, 2);
fanIn = shape[1] * receptiveFieldSize;
fanOut = shape[0] * receptiveFieldSize;
}
else if (dataFormat === 'channelsLast') {
const receptiveFieldSize = arrayProd(shape, 0, shape.length - 2);
fanIn = shape[shape.length - 2] * receptiveFieldSize;
fanOut = shape[shape.length - 1] * receptiveFieldSize;
}
}
else {
const shapeProd = arrayProd(shape);
fanIn = Math.sqrt(shapeProd);
fanOut = Math.sqrt(shapeProd);
}
return [fanIn, fanOut];
}
class VarianceScaling extends Initializer {
/**
* Constructor of VarianceScaling.
* @throws ValueError for invalid value in scale.
*/
constructor(args) {
super();
if (args.scale < 0.0) {
throw new ValueError(`scale must be a positive float. Got: ${args.scale}`);
}
this.scale = args.scale == null ? 1.0 : args.scale;
this.mode = args.mode == null ? 'fanIn' : args.mode;
checkFanMode(this.mode);
this.distribution =
args.distribution == null ? 'normal' : args.distribution;
checkDistribution(this.distribution);
this.seed = args.seed;
}
apply(shape, dtype) {
const fans = computeFans(shape);
const fanIn = fans[0];
const fanOut = fans[1];
let scale = this.scale;
if (this.mode === 'fanIn') {
scale /= Math.max(1, fanIn);
}
else if (this.mode === 'fanOut') {
scale /= Math.max(1, fanOut);
}
else {
scale /= Math.max(1, (fanIn + fanOut) / 2);
}
if (this.distribution === 'normal') {
const stddev = Math.sqrt(scale);
dtype = dtype || 'float32';
if (dtype !== 'float32' && dtype !== 'int32') {
throw new NotImplementedError(`${this.getClassName()} does not support dType ${dtype}.`);
}
return truncatedNormal(shape, 0, stddev, dtype, this.seed);
}
else {
const limit = Math.sqrt(3 * scale);
return randomUniform(shape, -limit, limit, dtype, this.seed);
}
}
getConfig() {
return {
scale: this.scale,
mode: this.mode,
distribution: this.distribution,
seed: this.seed
};
}
}
/** @nocollapse */
VarianceScaling.className = 'VarianceScaling';
registerClass(VarianceScaling);
class GlorotUniform extends VarianceScaling {
/**
* Constructor of GlorotUniform
* @param scale
* @param mode
* @param distribution
* @param seed
*/
constructor(args) {
super({
scale: 1.0,
mode: 'fanAvg',
distribution: 'uniform',
seed: args == null ? null : args.seed
});
}
getClassName() {
// In Python Keras, GlorotUniform is not a class, but a helper method
// that creates a VarianceScaling object. Use 'VarianceScaling' as
// class name to be compatible with that.
return VarianceScaling.className;
}
}
/** @nocollapse */
GlorotUniform.className = 'GlorotUniform';
registerClass(GlorotUniform);
class GlorotNormal extends VarianceScaling {
/**
* Constructor of GlorotNormal.
* @param scale
* @param mode
* @param distribution
* @param seed
*/
constructor(args) {
super({
scale: 1.0,
mode: 'fanAvg',
distribution: 'normal',
seed: args == null ? null : args.seed
});
}
getClassName() {
// In Python Keras, GlorotNormal is not a class, but a helper method
// that creates a VarianceScaling object. Use 'VarianceScaling' as
// class name to be compatible with that.
return VarianceScaling.className;
}
}
/** @nocollapse */
GlorotNormal.className = 'GlorotNormal';
registerClass(GlorotNormal);
class HeNormal extends VarianceScaling {
constructor(args) {
super({
scale: 2.0,
mode: 'fanIn',
distribution: 'normal',
seed: args == null ? null : args.seed
});
}
getClassName() {
// In Python Keras, HeNormal is not a class, but a helper method
// that creates a VarianceScaling object. Use 'VarianceScaling' as
// class name to be compatible with that.
return VarianceScaling.className;
}
}
/** @nocollapse */
HeNormal.className = 'HeNormal';
registerClass(HeNormal);
class HeUniform extends VarianceScaling {
constructor(args) {
super({
scale: 2.0,
mode: 'fanIn',
distribution: 'uniform',
seed: args == null ? null : args.seed
});
}
getClassName() {
// In Python Keras, HeUniform is not a class, but a helper method
// that creates a VarianceScaling object. Use 'VarianceScaling' as
// class name to be compatible with that.
return VarianceScaling.className;
}
}
/** @nocollapse */
HeUniform.className = 'HeUniform';
registerClass(HeUniform);
class LeCunNormal extends VarianceScaling {
constructor(args) {
super({
scale: 1.0,
mode: 'fanIn',
distribution: 'normal',
seed: args == null ? null : args.seed
});
}
getClassName() {
// In Python Keras, LeCunNormal is not a class, but a helper method
// that creates a VarianceScaling object. Use 'VarianceScaling' as
// class name to be compatible with that.
return VarianceScaling.className;
}
}
/** @nocollapse */
LeCunNormal.className = 'LeCunNormal';
registerClass(LeCunNormal);
class LeCunUniform extends VarianceScaling {
constructor(args) {
super({
scale: 1.0,
mode: 'fanIn',
distribution: 'uniform',
seed: args == null ? null : args.seed
});
}
getClassName() {
// In Python Keras, LeCunUniform is not a class, but a helper method
// that creates a VarianceScaling object. Use 'VarianceScaling' as
// class name to be compatible with that.
return VarianceScaling.className;
}
}
/** @nocollapse */
LeCunUniform.className = 'LeCunUniform';
registerClass(LeCunUniform);
class Orthogonal extends Initializer {
constructor(args) {
super();
this.DEFAULT_GAIN = 1;
this.ELEMENTS_WARN_SLOW = 2000;
this.gain = args.gain == null ? this.DEFAULT_GAIN : args.gain;
this.seed = args.seed;
}
apply(shape, dtype) {
return tidy(() => {
if (shape.length < 2) {
throw new NotImplementedError('Shape must be at least 2D.');
}
if (dtype !== 'int32' && dtype !== 'float32' && dtype !== undefined) {
throw new TypeError(`Unsupported data type ${dtype}.`);
}
dtype = dtype;
// flatten the input shape with the last dimension remaining its
// original shape so it works for conv2d
const numRows = sizeFromShape(shape.slice(0, -1));
const numCols = shape[shape.length - 1];
const numElements = numRows * numCols;
if (numElements > this.ELEMENTS_WARN_SLOW) {
console.warn(`Orthogonal initializer is being called on a matrix with more ` +
`than ${this.ELEMENTS_WARN_SLOW} (${numElements}) elements: ` +
`Slowness may result.`);
}
const flatShape = [Math.max(numCols, numRows), Math.min(numCols, numRows)];
// Generate a random matrix
const randNormalMat = randomNormal(flatShape, 0, 1, dtype, this.seed);
// Compute QR factorization
const qr = linalg.qr(randNormalMat, false);
let qMat = qr[0];
const rMat = qr[1];
// Make Q uniform
const diag = rMat.flatten().stridedSlice([0], [Math.min(numCols, numRows) * Math.min(numCols, numRows)], [Math.min(numCols, numRows) + 1]);
qMat = mul(qMat, diag.sign());
if (numRows < numCols) {
qMat = qMat.transpose();
}
return mul(scalar(this.gain), qMat.reshape(shape));
});
}
getConfig() {
return {
gain: this.gain,
seed: this.seed,
};
}
}
/** @nocollapse */
Orthogonal.className = 'Orthogonal';
registerClass(Orthogonal);
// Maps the JavaScript-like identifier keys to the corresponding registry
// symbols.
const INITIALIZER_IDENTIFIER_REGISTRY_SYMBOL_MAP = {
'constant': 'Constant',
'glorotNormal': 'GlorotNormal',
'glorotUniform': 'GlorotUniform',
'heNormal': 'HeNormal',
'heUniform': 'HeUniform',
'identity': 'Identity',
'leCunNormal': 'LeCunNormal',
'leCunUniform': 'LeCunUniform',
'ones': 'Ones',
'orthogonal': 'Orthogonal',
'randomNormal': 'RandomNormal',
'randomUniform': 'RandomUniform',
'truncatedNormal': 'TruncatedNormal',
'varianceScaling': 'VarianceScaling',
'zeros': 'Zeros'
};
function deserializeInitializer(config, customObjects = {}) {
return deserializeKerasObject(config, SerializationMap.getMap().classNameMap, customObjects, 'initializer');
}
function serializeInitializer(initializer) {
return serializeKerasObject(initializer);
}
function getInitializer(identifier) {
if (typeof identifier === 'string') {
const className = identifier in INITIALIZER_IDENTIFIER_REGISTRY_SYMBOL_MAP ?
INITIALIZER_IDENTIFIER_REGISTRY_SYMBOL_MAP[identifier] :
identifier;
/* We have four 'helper' classes for common initializers that
all get serialized as 'VarianceScaling' and shouldn't go through
the deserializeInitializer pathway. */
if (className === 'GlorotNormal') {
return new GlorotNormal();
}
else if (className === 'GlorotUniform') {
return new GlorotUniform();
}
else if (className === 'HeNormal') {
return new HeNormal();
}
else if (className === 'HeUniform') {
return new HeUniform();
}
else if (className === 'LeCunNormal') {
return new LeCunNormal();
}
else if (className === 'LeCunUniform') {
return new LeCunUniform();
}
else {
const config = {};
config['className'] = className;
config['config'] = {};
return deserializeInitializer(config);
}
}
else if (identifier instanceof Initializer) {
return identifier;
}
else {
return deserializeInitializer(identifier);
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Special case of normalizing shapes to lists.
*
* @param x A shape or list of shapes to normalize into a list of Shapes.
* @return A list of Shapes.
*/
function normalizeShapeList(x) {
if (x.length === 0) {
return [];
}
if (!Array.isArray(x[0])) {
return [x];
}
return x;
}
/**
* Helper function to obtain exactly one Tensor.
* @param xs: A single `tf.Tensor` or an `Array` of `tf.Tensor`s.
* @return A single `tf.Tensor`. If `xs` is an `Array`, return the first one.
* @throws ValueError: If `xs` is an `Array` and its length is not 1.
*/
function getExactlyOneTensor(xs) {
let x;
if (Array.isArray(xs)) {
if (xs.length !== 1) {
throw new ValueError(`Expected Tensor length to be 1; got ${xs.length}`);
}
x = xs[0];
}
else {
x = xs;
}
return x;
}
/**
* Helper function to obtain exactly on instance of Shape.
*
* @param shapes Input single `Shape` or Array of `Shape`s.
* @returns If input is a single `Shape`, return it unchanged. If the input is
* an `Array` containing exactly one instance of `Shape`, return the instance.
* Otherwise, throw a `ValueError`.
* @throws ValueError: If input is an `Array` of `Shape`s, and its length is not
* 1.
*/
function getExactlyOneShape(shapes) {
if (Array.isArray(shapes) && Array.isArray(shapes[0])) {
if (shapes.length === 1) {
shapes = shapes;
return shapes[0];
}
else {
throw new ValueError(`Expected exactly 1 Shape; got ${shapes.length}`);
}
}
else {
return shapes;
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Count the elements in an Array of LayerVariables.
*
* @param weights: The LayerVariables of which the constituent numbers are to
* be counted.
* @returns A count of the elements in all the LayerVariables
*/
function countParamsInWeights(weights) {
let count = 0;
for (const weight of weights) {
if (weight.shape.length === 0) {
count += 1;
}
else {
count += weight.shape.reduce((a, b) => a * b);
}
}
return count;
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
const DEFAULT_VARIABLE_NAME_PREFIX = 'Variable';
/**
* A `tf.layers.LayerVariable` is similar to a `tf.Tensor` in that it has a
* dtype and shape, but its value is mutable. The value is itself represented
* as a`tf.Tensor`, and can be read with the `read()` method and updated with
* the `write()` method.
*/
class LayerVariable {
/**
* Construct Variable from a `tf.Tensor`.
*
* If not explicitly named, the Variable will be given a name with the
* prefix 'Variable'. Variable names are unique. In the case of name
* collision, suffixies '_<num>' will be added to the name.
*
* @param val Initial value of the Variable.
* @param name Name of the variable. If `null` or `undefined` is provided, it
* will default a name with the prefix 'Variable'.
* @param constraint Optional, projection function to be applied to the
* variable after optimize updates
* @throws ValueError if `name` is `null` or `undefined`.
*/
constructor(val, dtype = 'float32', name = DEFAULT_VARIABLE_NAME_PREFIX, trainable = true, constraint = null) {
this.dtype = dtype == null ? 'float32' : dtype;
this.shape = val.shape;
this.id = getNextUniqueTensorId();
name = name == null ? DEFAULT_VARIABLE_NAME_PREFIX : name;
this.originalName = getScopedTensorName(name);
this.name = getUniqueTensorName(this.originalName);
this.trainable_ = trainable;
this.constraint = constraint;
this.val = variable(val, this.trainable_, this.name, this.dtype);
}
/**
* Get a snapshot of the Variable's value.
*
* The returned value is a snapshot of the Variable's value at the time of
* the invocation. Future mutations in the value of the tensor will only
* be reflected by future calls to this method.
*/
read() {
this.assertNotDisposed();
return this.val;
}
/**
* Update the value of the Variable.
*
* @param newVal: The new value to update to. Must be consistent with the
* dtype and shape of the Variable.
* @return This Variable.
*/
write(newVal) {
// TODO(cais): Once TF.js Core supports Tensor.dtype, check dtype match.
this.assertNotDisposed();
checkShapesMatch(this.val, newVal);
// Skip updating if this is the exact same tensor.
if (this.val.id !== newVal.id) {
this.val.assign(newVal);
if (this.constraint != null) {
this.val.assign(this.constraint.apply(this.val));
}
}
return this;
}
/**
* Dispose this LayersVariable instance from memory.
*/
dispose() {
this.assertNotDisposed();
this.val.dispose();
}
assertNotDisposed() {
if (this.val.isDisposed) {
throw new Error(`LayersVariable ${this.name} is already disposed.`);
}
}
get trainable() {
return this.trainable_;
}
set trainable(trainable) {
this.trainable_ = trainable;
this.val.trainable = trainable;
}
}
function checkShapesMatch(x, y) {
if (x.shape.toString() !== y.shape.toString()) {
throw new Error('Shape mismatch: ' + JSON.stringify(x.shape) + ' vs. ' +
JSON.stringify(y.shape));
}
}
/**
* Get the values of an array of Variables.
*
* @param tensors An `Array` of `Variable`s to get the values of.
* @return The values of the inputs, as an `Array` of`tf.Tensor`s.
*/
function batchGetValue(xs) {
return xs.map(x => x.read());
}
/**
* Update the value of multiple Variables at once.
*
* @param variablesAndValues An `Array`, each element is of type
* [Variable, Tensor]. The first item is the
* `Variable` of which the value is to be updated. The second item
* carries the new value.
*/
function batchSetValue(variablesAndValues) {
variablesAndValues.forEach(variableAndValue => {
const variable = variableAndValue[0];
variable.write(variableAndValue[1]);
});
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original source: keras/engine/topology.py */
/**
* Specifies the ndim, dtype and shape of every input to a layer.
*
* Every layer should expose (if appropriate) an `inputSpec` attribute:
* a list of instances of InputSpec (one per input tensor).
*
* A null entry in a shape is compatible with any dimension,
* a null shape is compatible with any shape.
*/
class InputSpec {
constructor(args) {
this.dtype = args.dtype;
this.shape = args.shape;
/*
TODO(michaelterry): Could throw error if ndim and shape are both defined
(then backport).
*/
if (args.shape != null) {
this.ndim = args.shape.length;
}
else {
this.ndim = args.ndim;
}
this.maxNDim = args.maxNDim;
this.minNDim = args.minNDim;
this.axes = args.axes || {};
}
}
/**
* `tf.SymbolicTensor` is a placeholder for a Tensor without any concrete value.
*
* They are most often encountered when building a graph of `Layer`s for a
* `tf.LayersModel` and the input data's shape, but not values are known.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
class SymbolicTensor {
/**
*
* @param dtype
* @param shape
* @param sourceLayer The Layer that produced this symbolic tensor.
* @param inputs The inputs passed to sourceLayer's __call__() method.
* @param nodeIndex
* @param tensorIndex
* @param callArgs The keyword arguments passed to the __call__() method.
* @param name
* @param outputTensorIndex The index of this tensor in the list of outputs
* returned by apply().
*/
constructor(dtype, shape, sourceLayer, inputs, callArgs, name, outputTensorIndex) {
this.dtype = dtype;
this.shape = shape;
this.sourceLayer = sourceLayer;
this.inputs = inputs;
this.callArgs = callArgs;
this.outputTensorIndex = outputTensorIndex;
this.id = getNextUniqueTensorId();
if (name != null) {
this.originalName = getScopedTensorName(name);
this.name = getUniqueTensorName(this.originalName);
}
this.rank = shape.length;
}
}
let _nextNodeID = 0;
/**
* A `Node` describes the connectivity between two layers.
*
* Each time a layer is connected to some new input,
* a node is added to `layer.inboundNodes`.
*
* Each time the output of a layer is used by another layer,
* a node is added to `layer.outboundNodes`.
*
* `nodeIndices` and `tensorIndices` are basically fine-grained coordinates
* describing the origin of the `inputTensors`, verifying the following:
*
* `inputTensors[i] ==
* inboundLayers[i].inboundNodes[nodeIndices[i]].outputTensors[
* tensorIndices[i]]`
*
* A node from layer A to layer B is added to:
* A.outboundNodes
* B.inboundNodes
*/
class Node {
constructor(args,
// TODO(michaelterry): Define actual type for this.
callArgs) {
this.callArgs = callArgs;
this.id = _nextNodeID++;
/*
Layer instance (NOT a list).
this is the layer that takes a list of input tensors
and turns them into a list of output tensors.
the current node will be added to
the inboundNodes of outboundLayer.
*/
this.outboundLayer = args.outboundLayer;
/*
The following 3 properties describe where
the input tensors come from: which layers,
and for each layer, which node and which
tensor output of each node.
*/
// List of layer instances.
this.inboundLayers = args.inboundLayers;
// List of integers, 1:1 mapping with inboundLayers.
this.nodeIndices = args.nodeIndices;
// List of integers, 1:1 mapping with inboundLayers.
this.tensorIndices = args.tensorIndices;
/*
Following 2 properties:
tensor inputs and outputs of outboundLayer.
*/
// List of tensors. 1:1 mapping with inboundLayers.
this.inputTensors = args.inputTensors;
// List of tensors, created by outboundLayer.call().
this.outputTensors = args.outputTensors;
/*
Following 2 properties: input and output masks.
List of tensors, 1:1 mapping with inputTensor.
*/
this.inputMasks = args.inputMasks;
// List of tensors, created by outboundLayer.computeMask().
this.outputMasks = args.outputMasks;
// Following 2 properties: input and output shapes.
// List of shape tuples, shapes of inputTensors.
this.inputShapes = args.inputShapes;
// List of shape tuples, shapes of outputTensors.
this.outputShapes = args.outputShapes;
// Add nodes to all layers involved.
for (const layer of args.inboundLayers) {
if (layer != null) {
layer.outboundNodes.push(this);
}
}
args.outboundLayer.inboundNodes.push(this);
}
getConfig() {
const inboundNames = [];
for (const layer of this.inboundLayers) {
if (layer != null) {
inboundNames.push(layer.name);
}
else {
inboundNames.push(null);
}
}
return {
outboundLayer: this.outboundLayer ? this.outboundLayer.name : null,
inboundLayers: inboundNames,
nodeIndices: this.nodeIndices,
tensorIndices: this.tensorIndices
};
}
}
let _nextLayerID = 0;
/**
* A layer is a grouping of operations and weights that can be composed to
* create a `tf.LayersModel`.
*
* Layers are constructed by using the functions under the
* [tf.layers](#Layers-Basic) namespace.
*
* @doc {heading: 'Layers', subheading: 'Classes', namespace: 'layers'}
*/
class Layer extends Serializable {
constructor(args = {}) {
super();
this._callHook = null;
this._addedWeightNames = [];
// Porting Notes: PyKeras does not have this property in this base Layer
// class. Instead lets Layer subclass set it dynamically and checks the
// value with `hasattr`. In tfjs-layers, we let this be a member of this
// base class.
this._stateful = false;
this.id = _nextLayerID++;
this.activityRegularizer = null;
this.inputSpec = null;
this.supportsMasking = false;
// These properties will be set upon call of this.build()
this._trainableWeights = [];
this._nonTrainableWeights = [];
this._losses = [];
this._updates = [];
this._built = false;
/*
These lists will be filled via successive calls
to this.addInboundNode().
*/
this.inboundNodes = [];
this.outboundNodes = [];
let name = args.name;
if (!name) {
const prefix = this.getClassName();
name = toSnakeCase(prefix) + '_' + getUid(prefix);
}
this.name = name;
this.trainable_ = args.trainable == null ? true : args.trainable;
if (args.inputShape != null || args.batchInputShape != null) {
/*
In this case we will later create an input layer
to insert before the current layer
*/
let batchInputShape;
if (args.batchInputShape != null) {
batchInputShape = args.batchInputShape;
}
else if (args.inputShape != null) {
let batchSize = null;
if (args.batchSize != null) {
batchSize = args.batchSize;
}
batchInputShape = [batchSize].concat(args.inputShape);
}
this.batchInputShape = batchInputShape;
// Set dtype.
let dtype = args.dtype;
if (dtype == null) {
dtype = args.inputDType;
}
if (dtype == null) {
dtype = 'float32';
}
this.dtype = dtype;
}
if (args.weights != null) {
this.initialWeights = args.weights;
}
else {
this.initialWeights = null;
}
// The value of `_refCount` is initialized to null. When the layer is used
// in a symbolic way for the first time, it will be set to 1.
this._refCount = null;
this.fastWeightInitDuringBuild = false;
}
/**
* Converts a layer and its index to a unique (immutable type) name.
* This function is used internally with `this.containerNodes`.
* @param layer The layer.
* @param nodeIndex The layer's position (e.g. via enumerate) in a list of
* nodes.
*
* @returns The unique name.
*/
static nodeKey(layer, nodeIndex) {
return layer.name + '_ib-' + nodeIndex.toString();
}
/**
* Returns this.inboundNode at index nodeIndex.
*
* Porting note: This is a replacement for _get_node_attribute_at_index()
* @param nodeIndex
* @param attrName The name of the attribute related to request for this node.
*/
getNodeAtIndex(nodeIndex, attrName) {
if (this.inboundNodes.length === 0) {
throw new RuntimeError('The layer has never been called ' +
`and thus has no defined ${attrName}.`);
}
if (this.inboundNodes.length <= nodeIndex) {
throw new ValueError(`Asked to get ${attrName} at node ${nodeIndex}, ` +
`but the layer has only ${this.inboundNodes.length} inbound nodes.`);
}
return this.inboundNodes[nodeIndex];
}
/**
* Retrieves the input tensor(s) of a layer at a given node.
*
* @param nodeIndex Integer, index of the node from which to retrieve the
* attribute. E.g. `nodeIndex=0` will correspond to the first time the layer
* was called.
*
* @return A tensor (or list of tensors if the layer has multiple inputs).
*/
getInputAt(nodeIndex) {
return singletonOrArray(this.getNodeAtIndex(nodeIndex, 'input').inputTensors);
}
/**
* Retrieves the output tensor(s) of a layer at a given node.
*
* @param nodeIndex Integer, index of the node from which to retrieve the
* attribute. E.g. `nodeIndex=0` will correspond to the first time the layer
* was called.
*
* @return A tensor (or list of tensors if the layer has multiple outputs).
*/
getOutputAt(nodeIndex) {
return singletonOrArray(this.getNodeAtIndex(nodeIndex, 'output').outputTensors);
}
// Properties
/**
* Retrieves the input tensor(s) of a layer.
*
* Only applicable if the layer has exactly one inbound node,
* i.e. if it is connected to one incoming layer.
*
* @return Input tensor or list of input tensors.
*
* @exception AttributeError if the layer is connected to more than one
* incoming layers.
*/
get input() {
if (this.inboundNodes.length > 1) {
throw new AttributeError(`Layer ${this.name}` +
' has multiple inbound nodes, ' +
'hence the notion of "layer input" ' +
'is ill-defined. ' +
'Use `getInputAt(nodeIndex)` instead.');
}
else if (this.inboundNodes.length === 0) {
throw new AttributeError(`Layer ${this.name}` +
' is not connected, no input to return.');
}
return singletonOrArray(this.getNodeAtIndex(0, 'input').inputTensors);
}
/**
* Retrieves the output tensor(s) of a layer.
*
* Only applicable if the layer has exactly one inbound node,
* i.e. if it is connected to one incoming layer.
*
* @return Output tensor or list of output tensors.
*
* @exception AttributeError if the layer is connected to more than one
* incoming layers.
*/
get output() {
if (this.inboundNodes.length === 0) {
throw new AttributeError(`Layer ${this.name}` +
' has no inbound nodes.');
}
if (this.inboundNodes.length > 1) {
throw new AttributeError(`Layer ${this.name}` +
' has multiple inbound nodes, ' +
'hence the notion of "layer output" ' +
'is ill-defined. ' +
'Use `getOutputAt(nodeIndex)` instead.');
}
return singletonOrArray(this.getNodeAtIndex(0, 'output').outputTensors);
}
get losses() {
return this._losses;
}
/**
* Retrieves the Layer's current loss values.
*
* Used for regularizers during training.
*/
calculateLosses() {
// Porting Node: This is an augmentation to Layer.loss in PyKeras.
// In PyKeras, Layer.loss returns symbolic tensors. Here a concrete
// Tensor (specifically Scalar) values are returned. This is due to the
// imperative backend.
return this.losses.map(lossFn => lossFn());
}
get updates() {
return this._updates;
}
get built() {
return this._built;
}
set built(built) {
this._built = built;
}
get trainable() {
return this.trainable_;
}
set trainable(trainable) {
this._trainableWeights.forEach(w => w.trainable = trainable);
this.trainable_ = trainable;
}
get trainableWeights() {
if (this.trainable_) {
return this._trainableWeights.filter(w => w.trainable);
}
else {
return [];
}
}
set trainableWeights(weights) {
this._trainableWeights = weights;
}
get nonTrainableWeights() {
if (this.trainable) {
return this._trainableWeights.filter(w => !w.trainable)
.concat(this._nonTrainableWeights);
}
else {
return this._trainableWeights.concat(this._nonTrainableWeights);
}
}
set nonTrainableWeights(weights) {
this._nonTrainableWeights = weights;
}
/**
* The concatenation of the lists trainableWeights and nonTrainableWeights
* (in this order).
*/
get weights() {
return this.trainableWeights.concat(this.nonTrainableWeights);
}
get stateful() {
return this._stateful;
}
/**
* Reset the states of the layer.
*
* This method of the base Layer class is essentially a no-op.
* Subclasses that are stateful (e.g., stateful RNNs) should override this
* method.
*/
resetStates() {
if (!this.stateful) {
throw new Error('Cannot call the resetStates() method of a non-stateful Layer ' +
'object.');
}
}
/**
* Checks compatibility between the layer and provided inputs.
*
* This checks that the tensor(s) `input`
* verify the input assumptions of the layer
* (if any). If not, exceptions are raised.
*
* @param inputs Input tensor or list of input tensors.
*
* @exception ValueError in case of mismatch between
* the provided inputs and the expectations of the layer.
*/
assertInputCompatibility(inputs) {
const inputsList = toList(inputs);
if (this.inputSpec == null || this.inputSpec.length === 0) {
return;
}
const inputSpec = toList(this.inputSpec);
if (inputsList.length !== inputSpec.length) {
throw new ValueError(`Layer ${this.name} expects ${inputSpec.length} inputs, ` +
`but it received ${inputsList.length} input tensors. ` +
`Input received: ${inputs}`);
}
for (let inputIndex = 0; inputIndex < inputsList.length; inputIndex++) {
const x = inputsList[inputIndex];
const spec = inputSpec[inputIndex];
if (spec == null) {
continue;
}
// Check ndim.
const ndim = x.rank;
if (spec.ndim != null) {
if (ndim !== spec.ndim) {
throw new ValueError(`Input ${inputIndex} is incompatible with layer ${this.name}: ` +
`expected ndim=${spec.ndim}, found ndim=${ndim}`);
}
}
if (spec.maxNDim != null) {
if (ndim > spec.maxNDim) {
throw new ValueError(`Input ${inputIndex} is incompatible with layer ${this.name}` +
`: expected max_ndim=${spec.maxNDim}, found ndim=${ndim}`);
}
}
if (spec.minNDim != null) {
if (ndim < spec.minNDim) {
throw new ValueError(`Input ${inputIndex} is incompatible with layer ${this.name}` +
`: expected min_ndim=${spec.minNDim}, found ndim=${ndim}.`);
}
}
// Check dtype.
if (spec.dtype != null) {
if (x.dtype !== spec.dtype) {
throw new ValueError(`Input ${inputIndex} is incompatible with layer ${this.name} ` +
`: expected dtype=${spec.dtype}, found dtype=${x.dtype}.`);
}
}
// Check specific shape axes.
if (spec.axes) {
const xShape = x.shape;
for (const key in spec.axes) {
const axis = Number(key);
const value = spec.axes[key];
// Perform Python-style slicing in case axis < 0;
// TODO(cais): Use https://github.com/alvivi/typescript-underscore to
// ensure type safety through Underscore calls.
const xShapeAtAxis = axis >= 0 ? xShape[axis] : xShape[xShape.length + axis];
if (value != null && [value, null].indexOf(xShapeAtAxis) === -1) {
throw new ValueError(`Input ${inputIndex} is incompatible with layer ` +
`${this.name}: expected axis ${axis} of input shape to ` +
`have value ${value} but got shape ${xShape}.`);
}
}
}
// Check shape.
if (spec.shape != null) {
for (let i = 0; i < spec.shape.length; ++i) {
const specDim = spec.shape[i];
const dim = x.shape[i];
if (specDim != null && dim != null) {
if (specDim !== dim) {
throw new ValueError(`Input ${inputIndex} is incompatible with layer ` +
`${this.name}: expected shape=${spec.shape}, ` +
`found shape=${x.shape}.`);
}
}
}
}
}
}
/**
* This is where the layer's logic lives.
*
* @param inputs Input tensor, or list/tuple of input tensors.
* @param kwargs Additional keyword arguments.
*
* @return A tensor or list/tuple of tensors.
*/
call(inputs, kwargs) {
return inputs;
}
invokeCallHook(inputs, kwargs) {
if (this._callHook != null) {
this._callHook(inputs, kwargs);
}
}
/**
* Set call hook.
* This is currently used for testing only.
* @param callHook
*/
setCallHook(callHook) {
this._callHook = callHook;
}
/**
* Clear call hook.
* This is currently used for testing only.
*/
clearCallHook() {
this._callHook = null;
}
/**
* Builds or executes a `Layer`'s logic.
*
* When called with `tf.Tensor`(s), execute the `Layer`'s computation and
* return Tensor(s). For example:
*
* ```js
* const denseLayer = tf.layers.dense({
* units: 1,
* kernelInitializer: 'zeros',
* useBias: false
* });
*
* // Invoke the layer's apply() method with a `tf.Tensor` (with concrete
* // numeric values).
* const input = tf.ones([2, 2]);
* const output = denseLayer.apply(input);
*
* // The output's value is expected to be [[0], [0]], due to the fact that
* // the dense layer has a kernel initialized to all-zeros and does not have
* // a bias.
* output.print();
* ```
*
* When called with `tf.SymbolicTensor`(s), this will prepare the layer for
* future execution. This entails internal book-keeping on shapes of
* expected Tensors, wiring layers together, and initializing weights.
*
* Calling `apply` with `tf.SymbolicTensor`s are typically used during the
* building of non-`tf.Sequential` models. For example:
*
* ```js
* const flattenLayer = tf.layers.flatten();
* const denseLayer = tf.layers.dense({units: 1});
*
* // Use tf.layers.input() to obtain a SymbolicTensor as input to apply().
* const input = tf.input({shape: [2, 2]});
* const output1 = flattenLayer.apply(input);
*
* // output1.shape is [null, 4]. The first dimension is the undetermined
* // batch size. The second dimension comes from flattening the [2, 2]
* // shape.
* console.log(JSON.stringify(output1.shape));
*
* // The output SymbolicTensor of the flatten layer can be used to call
* // the apply() of the dense layer:
* const output2 = denseLayer.apply(output1);
*
* // output2.shape is [null, 1]. The first dimension is the undetermined
* // batch size. The second dimension matches the number of units of the
* // dense layer.
* console.log(JSON.stringify(output2.shape));
*
* // The input and output can be used to construct a model that consists
* // of the flatten and dense layers.
* const model = tf.model({inputs: input, outputs: output2});
* ```
*
* @param inputs a `tf.Tensor` or `tf.SymbolicTensor` or an Array of them.
* @param kwargs Additional keyword arguments to be passed to `call()`.
*
* @return Output of the layer's `call` method.
*
* @exception ValueError error in case the layer is missing shape information
* for its `build` call.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
// Porting Note: This is a replacement for __call__() in Python.
apply(inputs, kwargs) {
kwargs = kwargs || {};
this.assertNotDisposed();
// Ensure inputs are all the same type.
const inputsList = toList(inputs);
const allAreSymbolic = checkAllSymbolic(inputs);
const noneAreSymbolic = checkNoneSymbolic(inputs);
if (allAreSymbolic === noneAreSymbolic) {
throw new ValueError('Arguments to apply() must be all ' +
'SymbolicTensors or all Tensors');
}
// TODO(michaelterry): nameScope() may not be necessary.
return nameScope(this.name, () => {
// Handle laying building (weight creating, input spec locking).
if (!this.built) {
/*
Throw exceptions in case the input is not compatible
with the inputSpec specified in the layer constructor.
*/
this.assertInputCompatibility(inputs);
// Collect input shapes to build layer.
const inputShapes = [];
for (const xElem of toList(inputs)) {
inputShapes.push(xElem.shape);
}
this.build(singletonOrArray(inputShapes));
this.built = true;
// Load weights that were specified at layer instantiation.
if (this.initialWeights) {
this.setWeights(this.initialWeights);
}
if (this._refCount === null && noneAreSymbolic) {
// The first use of this layer is a non-symbolic call, set ref count
// to 1 so the Layer can be properly disposed if its dispose() method
// is called.
this._refCount = 1;
}
}
/*
Throw exceptions in case the input is not compatible
with the inputSpec set at build time.
*/
this.assertInputCompatibility(inputs);
// Handle mask propagation.
// TODO(michaelterry): Mask propagation not currently implemented.
// Actually call the layer, collecting output(s), mask(s), and shape(s).
if (noneAreSymbolic) {
let output = this.call(inputs, kwargs);
// Apply masks to the output tensors if the layer supports it.
if (this.supportsMasking) {
// TODO(mattsoulanille): pass the input tensors' masks to computeMask
this.setMaskMetadata(inputs, output);
}
// If the layer returns tensors from its inputs, unmodified,
// we copy them to avoid loss of tensor metadata.
const outputList = toList(output);
const outputListCopy = [];
// TODO(michaelterry): This copying may not be necessary given our eager
// backend.
for (let x of outputList) {
if (inputsList.indexOf(x) !== -1) {
x = x.clone();
}
outputListCopy.push(x);
}
output = singletonOrArray(outputListCopy);
if (this.activityRegularizer != null) {
throw new NotImplementedError('Layer invocation in the presence of activity ' +
'regularizer(s) is not supported yet.');
}
// TODO(michaelterry): Call addInboundNode()?
return output;
}
else {
const inputShape = collectInputShape(inputs);
const outputShape = this.computeOutputShape(inputShape);
let output;
const outputDType = guessOutputDType();
this.warnOnIncompatibleInputShape(Array.isArray(inputs) ? inputShape[0] :
inputShape);
if (outputShape != null && outputShape.length > 0 &&
Array.isArray(outputShape[0])) {
// We have multiple output shapes. Create multiple output tensors.
output = outputShape
.map((shape, index) => new SymbolicTensor(outputDType, shape, this, toList(inputs), kwargs, this.name, index));
}
else {
output = new SymbolicTensor(outputDType, outputShape, this, toList(inputs), kwargs, this.name);
}
/*
Add an inbound node to the layer, so that it keeps track
of the call and of all new variables created during the call.
This also updates the layer history of the output tensor(s).
If the input tensor(s) had no previous history,
this does nothing.
*/
this.addInboundNode(inputs, output, null, null, inputShape, outputShape, kwargs);
this._refCount++;
if (this.activityRegularizer != null) {
throw new NotImplementedError('Layer invocation in the presence of activity ' +
'regularizer(s) is not supported yet.');
}
return output;
}
});
}
/**
* Check compatibility between input shape and this layer's batchInputShape.
*
* Print warning if any incompatibility is found.
*
* @param inputShape Input shape to be checked.
*/
warnOnIncompatibleInputShape(inputShape) {
if (this.batchInputShape == null) {
return;
}
else if (inputShape.length !== this.batchInputShape.length) {
console.warn(`The rank of the input tensor provided (shape: ` +
`${JSON.stringify(inputShape)}) does not match that of the ` +
`batchInputShape (${JSON.stringify(this.batchInputShape)}) ` +
`of the layer ${this.name}`);
}
else {
let dimMismatch = false;
this.batchInputShape.forEach((dimension, i) => {
if (dimension != null && inputShape[i] != null &&
inputShape[i] !== dimension) {
dimMismatch = true;
}
});
if (dimMismatch) {
console.warn(`The shape of the input tensor ` +
`(${JSON.stringify(inputShape)}) does not ` +
`match the expectation of layer ${this.name}: ` +
`${JSON.stringify(this.batchInputShape)}`);
}
}
}
/**
* Retrieves the output shape(s) of a layer.
*
* Only applicable if the layer has only one inbound node, or if all inbound
* nodes have the same output shape.
*
* @returns Output shape or shapes.
* @throws AttributeError: if the layer is connected to more than one incoming
* nodes.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
get outputShape() {
if (this.inboundNodes == null || this.inboundNodes.length === 0) {
throw new AttributeError(`The layer ${this.name} has never been called and thus has no ` +
`defined output shape.`);
}
const allOutputShapes = [];
for (const node of this.inboundNodes) {
const shapeString = JSON.stringify(node.outputShapes);
if (allOutputShapes.indexOf(shapeString) === -1) {
allOutputShapes.push(shapeString);
}
}
if (allOutputShapes.length === 1) {
const outputShapes = this.inboundNodes[0].outputShapes;
if (Array.isArray(outputShapes) && Array.isArray(outputShapes[0]) &&
outputShapes.length === 1) {
return outputShapes[0];
}
else {
return outputShapes;
}
}
else {
throw new AttributeError(`The layer ${this.name} has multiple inbound nodes with different ` +
`output shapes. Hence the notion of "output shape" is ill-defined ` +
`for the layer.`);
// TODO(cais): Implement getOutputShapeAt().
}
}
/**
* Counts the total number of numbers (e.g., float32, int32) in the
* weights.
*
* @returns An integer count.
* @throws RuntimeError: If the layer is not built yet (in which case its
* weights are not defined yet.)
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
countParams() {
if (!this.built) {
throw new RuntimeError(`You tried to call countParams() on ${this.name}, ` +
`but the layer is not built yet. Build it first by calling ` +
`build(batchInputShape).`);
}
return countParamsInWeights(this.weights);
}
/**
* Creates the layer weights.
*
* Must be implemented on all layers that have weights.
*
* Called when apply() is called to construct the weights.
*
* @param inputShape A `Shape` or array of `Shape` (unused).
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
build(inputShape) {
this.built = true;
}
/**
* Returns the current values of the weights of the layer.
*
* @param trainableOnly Whether to get the values of only trainable weights.
* @returns Weight values as an `Array` of `tf.Tensor`s.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
getWeights(trainableOnly = false) {
return batchGetValue(trainableOnly ? this.trainableWeights : this.weights);
}
/**
* Sets the weights of the layer, from Tensors.
*
* @param weights a list of Tensors. The number of arrays and their shape
* must match number of the dimensions of the weights of the layer (i.e.
* it should match the output of `getWeights`).
*
* @exception ValueError If the provided weights list does not match the
* layer's specifications.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
setWeights(weights) {
tidy(() => {
const params = this.weights;
if (params.length !== weights.length) {
// TODO(cais): Restore the following and use `providedWeights`, instead
// of `weights` in the error message, once the deeplearn.js bug is
// fixed: https://github.com/PAIR-code/deeplearnjs/issues/498 const
// providedWeights = JSON.stringify(weights).slice(0, 50);
throw new ValueError(`You called setWeights(weights) on layer "${this.name}" ` +
`with a weight list of length ${weights.length}, ` +
`but the layer was expecting ${params.length} weights. ` +
`Provided weights: ${weights}...`);
}
if (params.length === 0) {
return;
}
const weightValueTuples = [];
const paramValues = batchGetValue(params);
for (let i = 0; i < paramValues.length; ++i) {
const pv = paramValues[i];
const p = params[i];
const w = weights[i];
if (!arraysEqual(pv.shape, w.shape)) {
throw new ValueError(`Layer weight shape ${pv.shape} ` +
`not compatible with provided weight shape ${w.shape}`);
}
weightValueTuples.push([p, w]);
}
batchSetValue(weightValueTuples);
});
}
/**
* Adds a weight variable to the layer.
*
* @param name Name of the new weight variable.
* @param shape The shape of the weight.
* @param dtype The dtype of the weight.
* @param initializer An initializer instance.
* @param regularizer A regularizer instance.
* @param trainable Whether the weight should be trained via backprop or not
* (assuming that the layer itself is also trainable).
* @param constraint An optional trainable.
* @return The created weight variable.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
addWeight(name, shape, dtype, initializer, regularizer, trainable, constraint, getInitializerFunc) {
// Reject duplicate weight names.
if (this._addedWeightNames.indexOf(name) !== -1) {
throw new ValueError(`Duplicate weight name ${name} for layer ${this.name}`);
}
this._addedWeightNames.push(name);
if (dtype == null) {
dtype = 'float32';
}
if (this.fastWeightInitDuringBuild) {
initializer = getInitializerFunc != null ? getInitializerFunc() :
getInitializer('zeros');
}
const initValue = initializer.apply(shape, dtype);
const weight = new LayerVariable(initValue, dtype, name, trainable, constraint);
initValue.dispose();
// Request backend not to dispose the weights of the model on scope() exit.
if (regularizer != null) {
this.addLoss(() => regularizer.apply(weight.read()));
}
if (trainable == null) {
trainable = true;
}
if (trainable) {
this._trainableWeights.push(weight);
}
else {
this._nonTrainableWeights.push(weight);
}
return weight;
}
/**
* Set the fast-weight-initialization flag.
*
* In cases where the initialized weight values will be immediately
* overwritten by loaded weight values during model loading, setting
* the flag to `true` saves unnecessary calls to potentially expensive
* initializers and speeds up the loading process.
*
* @param value Target value of the flag.
*/
setFastWeightInitDuringBuild(value) {
this.fastWeightInitDuringBuild = value;
}
/**
* Add losses to the layer.
*
* The loss may potentially be conditional on some inputs tensors,
* for instance activity losses are conditional on the layer's inputs.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
addLoss(losses) {
if (losses == null || Array.isArray(losses) && losses.length === 0) {
return;
}
// Update this.losses
losses = toList(losses);
if (this._losses !== undefined && this._losses !== null) {
this.losses.push(...losses);
}
}
/**
* Computes the output shape of the layer.
*
* Assumes that the layer will be built to match that input shape provided.
*
* @param inputShape A shape (tuple of integers) or a list of shape tuples
* (one per output tensor of the layer). Shape tuples can include null for
* free dimensions, instead of an integer.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
computeOutputShape(inputShape) {
return inputShape;
}
/**
* Computes an output mask tensor.
*
* @param inputs Tensor or list of tensors.
* @param mask Tensor or list of tensors.
*
* @return null or a tensor (or list of tensors, one per output tensor of the
* layer).
*/
computeMask(inputs, mask) {
if (!this.supportsMasking) {
if (mask != null) {
if (Array.isArray(mask)) {
mask.forEach(maskElement => {
if (maskElement != null) {
throw new TypeError(`Layer ${this.name} does not support masking, ` +
'but was passed an inputMask.');
}
});
}
else {
throw new TypeError(`Layer ${this.name} does not support masking, ` +
'but was passed an inputMask.');
}
}
// masking not explicitly supported: return null as mask
return null;
}
// if masking is explictly supported, by default
// carry over the input mask
return mask;
}
setMaskMetadata(inputs, outputs, previousMask) {
if (!this.supportsMasking) {
return;
}
const outputMasks = this.computeMask(inputs, previousMask);
const outputsList = toList(outputs);
const outputMasksList = toList(outputMasks);
if (outputsList.length !== outputMasksList.length) {
throw new Error(`${this.name} outputs ${outputsList.length} tensors ` +
`but ${outputsList.length} masks for those tensors`);
}
for (let i = 0; i < outputsList.length; i++) {
outputsList[i].kerasMask = outputMasksList[i];
}
}
/**
* Internal method to create an inbound node for the layer.
*
* @param inputTensors List of input tensors.
* @param outputTensors List of output tensors.
* @param inputMasks List of input masks (a mask can be a tensor, or null).
* @param outputMasks List of output masks (a mask can be a tensor, or null).
* @param inputShapes List of input shape tuples.
* @param outputShapes List of output shape tuples.
* @param kwargs Dictionary of keyword arguments that were passed to the
* `call` method of the layer at the call that created the node.
*/
addInboundNode(inputTensors, outputTensors, inputMasks, outputMasks, inputShapes, outputShapes, kwargs = null) {
const inputTensorList = toList(inputTensors);
outputTensors = toList(outputTensors);
inputMasks = toList(inputMasks);
outputMasks = toList(outputMasks);
inputShapes = normalizeShapeList(inputShapes);
outputShapes = normalizeShapeList(outputShapes);
// Collect input tensor(s) coordinates.
const inboundLayers = [];
const nodeIndices = [];
const tensorIndices = [];
for (const x of inputTensorList) {
/*
* TODO(michaelterry): Keras adds this value to tensors; it's not
* clear whether we'll use this or not.
*/
inboundLayers.push(x.sourceLayer);
nodeIndices.push(x.nodeIndex);
tensorIndices.push(x.tensorIndex);
}
// Create node, add it to inbound nodes.
// (This call has side effects.)
// tslint:disable-next-line:no-unused-expression
new Node({
outboundLayer: this,
inboundLayers,
nodeIndices,
tensorIndices,
inputTensors: inputTensorList,
outputTensors,
inputMasks,
outputMasks,
inputShapes,
outputShapes
}, kwargs);
// Update tensor history
for (let i = 0; i < outputTensors.length; i++) {
// TODO(michaelterry: _uses_learning_phase not tracked.
outputTensors[i].sourceLayer = this;
outputTensors[i].nodeIndex = this.inboundNodes.length - 1;
outputTensors[i].tensorIndex = i;
}
}
/**
* Returns the config of the layer.
*
* A layer config is a TS dictionary (serializable)
* containing the configuration of a layer.
* The same layer can be reinstantiated later
* (without its trained weights) from this configuration.
*
* The config of a layer does not include connectivity
* information, nor the layer class name. These are handled
* by 'Container' (one layer of abstraction above).
*
* Porting Note: The TS dictionary follows TS naming standards for
* keys, and uses tfjs-layers type-safe Enums. Serialization methods
* should use a helper function to convert to the pythonic storage
* standard. (see serialization_utils.convertTsToPythonic)
*
* @returns TS dictionary of configuration.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
getConfig() {
const config = { name: this.name, trainable: this.trainable };
if (this.batchInputShape != null) {
config['batchInputShape'] = this.batchInputShape;
}
if (this.dtype != null) {
config['dtype'] = this.dtype;
}
return config;
}
/**
* Dispose the weight variables that this Layer instance holds.
*
* @returns {number} Number of disposed variables.
*/
disposeWeights() {
this.weights.forEach(weight => weight.dispose());
return this.weights.length;
}
assertNotDisposed() {
if (this._refCount === 0) {
throw new Error(`Layer '${this.name}' is already disposed.`);
}
}
/**
* Attempt to dispose layer's weights.
*
* This method decreases the reference count of the Layer object by 1.
*
* A Layer is reference-counted. Its reference count is incremented by 1
* the first item its `apply()` method is called and when it becomes a part
* of a new `Node` (through calling the `apply()` method on a
* `tf.SymbolicTensor`).
*
* If the reference count of a Layer becomes 0, all the weights will be
* disposed and the underlying memory (e.g., the textures allocated in WebGL)
* will be freed.
*
* Note: If the reference count is greater than 0 after the decrement, the
* weights of the Layer will *not* be disposed.
*
* After a Layer is disposed, it cannot be used in calls such as `apply()`,
* `getWeights()` or `setWeights()` anymore.
*
* @returns A DisposeResult Object with the following fields:
* - refCountAfterDispose: The reference count of the Container after this
* `dispose()` call.
* - numDisposedVariables: Number of `tf.Variable`s (i.e., weights) disposed
* during this `dispose()` call.
* @throws {Error} If the layer is not built yet, or if the layer has already
* been disposed.
*
* @doc {heading: 'Models', 'subheading': 'Classes'}
*/
dispose() {
if (!this.built) {
throw new Error(`Cannot dispose Layer ${this.name} because it has not been ` +
`built yet.`);
}
if (this._refCount === null) {
throw new Error(`Cannot dispose Layer ${this.name} because it has not been used ` +
`yet.`);
}
this.assertNotDisposed();
let numDisposedVariables = 0;
if (--this._refCount === 0) {
numDisposedVariables = this.disposeWeights();
}
return { refCountAfterDispose: this._refCount, numDisposedVariables };
}
}
/**
* Collects the input shape(s) of a list of `tf.Tensor`s or
* `tf.SymbolicTensor`s.
*
* TODO(michaelterry): Update PyKeras docs (backport).
*
* @param inputTensors List of input tensors (or single input tensor).
*
* @return List of shape tuples (or single tuple), one tuple per input.
*/
function collectInputShape(inputTensors) {
inputTensors =
toList(inputTensors);
const shapes = [];
for (const x of inputTensors) {
shapes.push(x.shape);
}
return singletonOrArray(shapes);
}
/**
* Guesses output dtype based on inputs.
*
* At present, just returns 'float32' for any input.
*
* @param inputTensors List of input tensors (or single input tensor).
*
* @return The guessed DType. At present, always returns 'float32'.
*/
function guessOutputDType(inputTensors) {
return 'float32';
}
/**
* Returns the list of input tensors necessary to compute `tensor`.
*
* Output will always be a list of tensors (potentially with 1 element).
*
* @param tensor The tensor to start from.
* @param layer Origin layer of the tensor.
* @param nodeIndex Origin node index of the tensor.
*
* @return Array of input tensors.
*/
function getSourceInputs(tensor, layer, nodeIndex) {
if (layer == null || (nodeIndex != null && nodeIndex > 0)) {
layer = tensor.sourceLayer;
nodeIndex = tensor.nodeIndex;
}
if (layer.inboundNodes.length === 0) {
return [tensor];
}
else {
const node = layer.inboundNodes[nodeIndex];
if (node.inboundLayers.length === 0) {
return node.inputTensors;
}
else {
const sourceTensors = [];
for (let i = 0; i < node.inboundLayers.length; i++) {
const x = node.inputTensors[i];
const layer = node.inboundLayers[i];
const nodeIndex = node.nodeIndices[i];
const previousSources = getSourceInputs(x, layer, nodeIndex);
// Avoid input redundancy.
for (const x of previousSources) {
if (sourceTensors.indexOf(x) === -1) {
sourceTensors.push(x);
}
}
}
return sourceTensors;
}
}
}
function checkAllSymbolic(tensors) {
let allAreSymbolic = true;
for (const tensor of toList(tensors)) {
if (!(tensor instanceof SymbolicTensor)) {
allAreSymbolic = false;
break;
}
}
return allAreSymbolic;
}
function checkNoneSymbolic(tensors) {
let noneAreSymbolic = true;
for (const tensor of toList(tensors)) {
if (tensor instanceof SymbolicTensor) {
noneAreSymbolic = false;
break;
}
}
return noneAreSymbolic;
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
class InputLayer extends Layer {
constructor(args) {
super({
dtype: args.dtype,
name: args.name != null ? args.name : getUid('input').toString()
});
// Normalize config.batchSize and config.sparse
if (args.batchSize == null) {
args.batchSize = null;
}
if (args.sparse == null) {
args.sparse = false;
}
this.trainable = false;
this.built = true;
this.sparse = args.sparse;
if (args.inputShape != null && args.batchInputShape != null) {
throw new ValueError('Only provide the inputShape OR ' +
'batchInputShape argument to inputLayer, not both at the same time.');
}
let batchInputShape = args.batchInputShape;
if (batchInputShape == null) {
if (args.inputShape == null) {
throw new ValueError('An InputLayer should be passed either a ' +
'`batchInputShape` or an `inputShape`.');
}
else {
batchInputShape = [args.batchSize].concat(args.inputShape);
}
}
else {
// TODO(michaelterry): Backport to PyKeras
if (args.batchSize != null) {
throw new ValueError('Cannot specify batchSize if batchInputShape is ' +
'specified when creating an InputLayer.');
}
}
const dtype = args.dtype || 'float32';
this.batchInputShape = batchInputShape;
this.dtype = dtype;
// TODO(michaelterry): Backport this to PyKeras?
this.inputSpec = [{ shape: batchInputShape }];
const inputTensor = new SymbolicTensor(this.dtype, this.batchInputShape, this, [], {}, this.name);
inputTensor.nodeIndex = 0;
inputTensor.tensorIndex = 0;
// Create an input node to add to this.outboundNode.
// (This call has side effects.)
// tslint:disable-next-line:no-unused-expression
new Node({
outboundLayer: this,
inboundLayers: [],
nodeIndices: [],
tensorIndices: [],
inputTensors: [inputTensor],
outputTensors: [inputTensor],
inputMasks: [null],
outputMasks: [null],
inputShapes: [batchInputShape],
outputShapes: [batchInputShape]
});
}
apply(inputs, kwargs) {
throw new ValueError('Cannot pass any input to an ' +
`InputLayer's apply() method. InputLayer name: ${this.name}`);
}
dispose() {
// dispose() for InputLayer is overridden as no-op.
return { refCountAfterDispose: this._refCount, numDisposedVariables: 0 };
}
getConfig() {
return {
batchInputShape: this.batchInputShape,
dtype: this.dtype,
sparse: this.sparse,
name: this.name
};
}
}
/** @nocollapse */
InputLayer.className = 'InputLayer';
registerClass(InputLayer);
function Input(config) {
if (config.batchShape == null && config.shape == null) {
throw new Error('Please provide to Input either a `shape`' +
' or a `batchShape` argument. Note that ' +
'`shape` does not include the batch ' +
'dimension.');
}
if (config.batchShape != null && config.shape != null) {
// TODO(michaelterry): Backport to PyKeras.
throw new ValueError('Please provide either a `shape` or `batchShape` ' +
'argument to Input, but not both.');
}
let batchShape = config.batchShape;
if (config.shape != null && batchShape == null) {
batchShape = [null].concat(config.shape);
}
let dtype = config.dtype;
if (dtype == null) {
dtype = 'float32';
}
const inputLayer = new InputLayer({
batchInputShape: batchShape,
name: config.name,
dtype,
sparse: config.sparse
});
const outputs = inputLayer.inboundNodes[0].outputTensors;
return outputs[0];
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Executor: Evaluates SymbolicTensor based on feeds.
*/
/**
* Helper function to check the dtype and shape compatibility of a feed value.
*/
function assertFeedCompatibility(key, val) {
// Check dtype compatibility.
if (key.dtype == null || key.dtype === val.dtype) {
// a. If types match, return val tensor as is.
return val;
}
try {
// b. Attempt to convert to expected type.
return cast$3(val, key.dtype);
}
catch (err) {
// c. If conversion fails, return helpful error.
throw new ValueError(`The dtype of the feed (${val.dtype}) can not be cast to the dtype ` +
`of the key '${key.name}' (${key.dtype}).`);
}
}
/**
* FeedDict: A mapping from unique SymbolicTensors to feed values for them.
* A feed value is a concrete value represented as an `Tensor`.
*/
class FeedDict {
/**
* Constructor, optionally does copy-construction.
* @param feeds An Array of `Feed`s, or another `FeedDict`, in which case
* copy-construction will be performed.
*/
constructor(feeds) {
this.id2Value = {};
this.id2Mask = {};
this.name2Id = {};
if (feeds instanceof FeedDict) {
for (const id in feeds.id2Value) {
this.id2Value[id] = feeds.id2Value[id];
if (id in feeds.id2Mask) {
this.id2Mask[id] = feeds.id2Mask[id];
}
}
}
else {
if (feeds == null) {
return;
}
for (const feed of feeds) {
this.add(feed.key, feed.value);
}
}
}
/**
* Add a key-value pair to the FeedDict.
*
* @param key The key of the feed.
* @param value The value of the tensor feed.
* @param mask The value of the mask feed (optional).
* @returns This `FeedDict`.
* @throws ValueError: If the key `SymbolicTensor` already exists in the
* `FeedDict`.
*/
add(key, value, mask) {
if (this.id2Value[key.id] == null) {
this.id2Value[key.id] = assertFeedCompatibility(key, value);
this.name2Id[key.name] = key.id;
if (mask != null) {
this.id2Mask[key.id] = mask;
}
}
else {
throw new ValueError(`Duplicate key: name=${key.name}, id=${key.id}`);
}
return this;
}
/**
* Add a Feed to the FeedDict.
* @param feed The new `Feed` to add.
* @returns This `FeedDict`.
*/
addFeed(feed) {
this.add(feed.key, feed.value);
}
/**
* Probe whether a key already exists in the FeedDict.
* @param key
*/
hasKey(key) {
return this.id2Value[key.id] != null;
}
/**
* Get all the SymbolicTensor available in this FeedDict.
*/
names() {
return Object.keys(this.name2Id);
}
/**
* Get the feed value for given key.
* @param key The SymbolicTensor, or its name (as a string), of which the
* value is sought.
* @returns If `key` exists, the corresponding feed value.
* @throws ValueError: If `key` does not exist in this `FeedDict`.
*/
getValue(key) {
if (key instanceof SymbolicTensor) {
if (this.id2Value[key.id] == null) {
throw new ValueError(`Nonexistent key: ${key.name}`);
}
else {
return this.id2Value[key.id];
}
}
else {
const id = this.name2Id[key];
if (id == null) {
throw new ValueError(`Feed dict has no SymbolicTensor name: ${key}`);
}
return this.id2Value[id];
}
}
/**
* Get the feed mask for given key.
* @param key The SymbolicTensor, or its name (as a string), of which the
* value is sought.
* @returns If `key` exists, the corresponding feed mask.
* @throws ValueError: If `key` does not exist in this `FeedDict`.
*/
getMask(key) {
if (key instanceof SymbolicTensor) {
if (this.id2Value[key.id] == null) {
throw new ValueError(`Nonexistent key: ${key.name}`);
}
else {
return this.id2Mask[key.id];
}
}
else {
const id = this.name2Id[key];
if (id == null) {
throw new ValueError(`Feed dict has no SymbolicTensor name: ${key}`);
}
return this.id2Mask[id];
}
}
/** Dispose all mask Tensors held by this object. */
disposeMasks() {
if (this.id2Mask != null) {
dispose(this.id2Mask);
}
}
}
// Cache for topologically sorted SymbolicTensors for given execution
// targets (i.e., fetches).
const cachedSorted = new LruCache();
// Cache for recipient count maps for given execution targets (i.e., fetches).
const cachedRecipientCounts = new LruCache();
/**
* Execute a SymbolicTensor by using concrete feed values.
*
* A `SymbolicTensor` object is a node in a computation graph of TF.js
* Layers. The object is backed by a source layer and input
* `SymbolicTensor`s to the source layer. This method evaluates
* the `call()` method of the source layer, using concrete values of the
* inputs obtained from either
* * `feedDict`, if the input key exists in `feedDict`, or else,
* * a recursive call to `execute()` itself.
*
* @param x: The `SymbolicTensor` to execute.
* @param feedDict: The feed values, as base condition of the recursion.
* execution.
* @param kwargs: Optional keyword arguments.
* @param probe: A probe object (of interface `ExecutionProbe`) used for
* testing memory footprint of `execute` calls.
* @returns Result of the execution.
* @throws ValueError: If any `SymbolicTensor`s from `InputLayer`s
* encountered during the execution lacks a feed value in `feedDict`.
*/
function execute(fetches, feedDict, kwargs, probe) {
const training = kwargs == null ? false : kwargs['training'];
const arrayFetches = Array.isArray(fetches);
const fetchArray = arrayFetches ? fetches : [fetches];
const outputNames = fetchArray.map(t => t.name);
const finalOutputs = [];
const feedNames = feedDict.names();
for (const outputName of outputNames) {
if (feedNames.indexOf(outputName) !== -1) {
finalOutputs.push(feedDict.getValue(outputName));
}
else {
finalOutputs.push(null);
}
}
// Check cache.
const fetchAndFeedKey = outputNames.join(',') + '|' + feedDict.names().sort().join(',');
let sorted = cachedSorted.get(fetchAndFeedKey);
let recipientCounts;
if (sorted == null) {
// Cache doesn't contain the desired combination of fetches. Compute
// topological sort for the combination for the first time.
const out = getTopologicalSortAndRecipientCounts(fetchArray, feedDict);
sorted = out.sorted;
recipientCounts = out.recipientCounts;
// Store results in cache for future use.
cachedSorted.put(fetchAndFeedKey, sorted);
cachedRecipientCounts.put(fetchAndFeedKey, recipientCounts);
}
recipientCounts = {};
if (!training) {
Object.assign(recipientCounts, cachedRecipientCounts.get(fetchAndFeedKey));
}
const internalFeedDict = new FeedDict(feedDict);
// Start iterative execution on the topologically-sorted SymbolicTensors.
for (let i = 0; i < sorted.length; ++i) {
const symbolic = sorted[i];
const srcLayer = symbolic.sourceLayer;
if (srcLayer instanceof InputLayer) {
continue;
}
const inputValues = [];
const inputMasks = [];
const tensorsToDispose = [];
let maskExists = false;
for (const input of symbolic.inputs) {
const value = internalFeedDict.getValue(input);
const mask = internalFeedDict.getMask(input);
inputValues.push(value);
inputMasks.push(mask);
if (mask != null) {
maskExists = true;
}
if (!training) {
recipientCounts[input.name]--;
if (recipientCounts[input.name] === 0 && !feedDict.hasKey(input) &&
outputNames.indexOf(input.name) === -1 && !value.isDisposed &&
input.sourceLayer.stateful !== true) {
tensorsToDispose.push(value);
}
}
}
if (maskExists) {
kwargs = kwargs || {};
kwargs['mask'] = inputMasks[0];
}
const outputTensors = toList(srcLayer.apply(inputValues, kwargs));
let outputMask = null;
if (srcLayer.supportsMasking) {
outputMask = srcLayer.computeMask(inputValues, inputMasks);
}
const layerOutputs = getNodeOutputs(symbolic);
const outputSymbolicTensors = Array.isArray(layerOutputs) ? layerOutputs : [layerOutputs];
for (let i = 0; i < outputSymbolicTensors.length; ++i) {
if (!internalFeedDict.hasKey(outputSymbolicTensors[i])) {
internalFeedDict.add(outputSymbolicTensors[i], outputTensors[i], Array.isArray(outputMask) ? outputMask[0] : outputMask);
}
const index = outputNames.indexOf(outputSymbolicTensors[i].name);
if (index !== -1) {
finalOutputs[index] = outputTensors[i];
}
}
if (!training) {
// Clean up Tensors that are no longer needed.
dispose(tensorsToDispose);
}
}
// NOTE(cais): Unlike intermediate tensors, we don't discard mask
// tensors as we go, because these tensors are sometimes passed over a
// series of mutliple layers, i.e., not obeying the immediate input
// relations in the graph. If this becomes a memory-usage concern,
// we can improve this in the future.
internalFeedDict.disposeMasks();
return arrayFetches ? finalOutputs : finalOutputs[0];
}
/**
* Sort the `SymbolicTensor`s topologically, for an array of fetches.
*
* This function calls getTopologicalSortAndRecipientCountsForOneFetch and
* merges their results.
*
* @param fetch The array of fetches requested. Must be a non-empty array.
* @param feedDict The dictionary of fed values.
* @returns sorted: Topologically-sorted array of SymbolicTensors.
* recipientCounts: Recipient counts for all SymbolicTensors in `sorted`.
*/
function getTopologicalSortAndRecipientCounts(fetches, feedDict) {
assert$1(fetches != null && fetches.length > 0, () => `Expected at least one fetch, got none`);
let finalSorted = [];
let finalRecipientMap = {};
if (fetches.length === 1) {
// Special-casing 1 fetch for efficiency.
const out = getTopologicalSortAndRecipientCountsForOneFetch(fetches[0], feedDict);
finalSorted = out.sorted;
finalRecipientMap = out.recipientMap;
}
else {
const visited = new Set();
for (const fetch of fetches) {
const { sorted, recipientMap } = getTopologicalSortAndRecipientCountsForOneFetch(fetch, feedDict);
// Merge sorted SymbolicTensor Arrays.
for (const symbolicTensor of sorted) {
if (!visited.has(symbolicTensor.name)) {
finalSorted.push(symbolicTensor);
visited.add(symbolicTensor.name);
}
}
// Merge recipient maps.
for (const name in recipientMap) {
if (finalRecipientMap[name] == null) {
finalRecipientMap[name] = new Set();
}
recipientMap[name].forEach(recipient => finalRecipientMap[name].add(recipient));
}
}
}
return {
sorted: finalSorted,
recipientCounts: recipientMap2Counts(finalRecipientMap)
};
}
function recipientMap2Counts(recipientMap) {
const recipientCounts = {};
for (const name in recipientMap) {
recipientCounts[name] = recipientMap[name].size;
}
return recipientCounts;
}
/**
* Sort the `SymbolicTensor`s topologically, for a single fetch.
*
* This helper function processes the upstream SymbolicTensors of a single
* fetch.
*
* @param fetch The single fetch requested.
* @param feedDict The dictionary of fed values.
* @returns sorted: Topologically-sorted array of SymbolicTensors.
* recipientMap: Recipient names for all SymbolicTensors in `sorted`.
*/
function getTopologicalSortAndRecipientCountsForOneFetch(fetch, feedDict) {
const visited = new Set();
const sorted = [];
const recipientMap = {};
// Put keys of the feedDict into visited first, so they don't have to be
// walked. This is needed in case where there are feeds for intermediate
// SymbolicTensors of the graph.
for (const key of feedDict.names()) {
visited.add(key);
}
const stack = [];
const marks = [];
// Initial population of stack and marks.
stack.push(fetch);
while (stack.length > 0) {
const top = stack[stack.length - 1];
if (visited.has(top.name)) {
stack.pop();
continue;
}
const topIsMarked = marks[marks.length - 1] === stack.length - 1;
if (top.inputs.length === 0 || topIsMarked) {
// Input SymbolicTensor or all children have been visited.
stack.pop();
sorted.push(top);
visited.add(top.name);
if (topIsMarked) {
marks.pop();
}
}
else {
// A non-input SymbolicTensor whose upstream SymbolicTensors haven't
// been visited yet. Push them onto the stack.
marks.push(stack.length - 1);
for (const input of top.inputs) {
// Increment the recipient count. Note that this needs to happen
// regardless of whether the SymbolicTensor has been visited before.
if (recipientMap[input.name] == null) {
recipientMap[input.name] = new Set();
}
recipientMap[input.name].add(top.name);
if (visited.has(input.name)) {
continue; // Avoid repeated visits to the same SymbolicTensor.
}
stack.push(input);
}
}
}
return { sorted, recipientMap };
}
/**
* Get the symbolic output tensors of the node to which a given fetch belongs.
* @param fetch The fetched symbolic tensor.
* @returns The Array of symbolic tensors output by the node to which `fetch`
* belongs.
*/
function getNodeOutputs(fetch) {
let layerOutputs;
if (fetch.sourceLayer.inboundNodes.length === 1) {
layerOutputs = fetch.sourceLayer.output;
}
else {
let nodeIndex = null;
for (let i = 0; i < fetch.sourceLayer.inboundNodes.length; ++i) {
for (const outputTensor of fetch.sourceLayer.inboundNodes[i]
.outputTensors) {
if (outputTensor.id === fetch.id) {
nodeIndex = i;
break;
}
}
}
layerOutputs = fetch.sourceLayer.getOutputAt(nodeIndex);
}
return layerOutputs;
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original source: keras/contraints.py */
/**
* Helper function used by many of the Constraints to find the L2Norms.
*/
function calcL2Norms(w, axis) {
return tidy(() => sqrt$2(sum$2(mul(w, w), axis, true)));
}
/**
* Base class for functions that impose constraints on weight values
*
* @doc {
* heading: 'Constraints',
* subheading: 'Classes',
* namespace: 'constraints'
* }
*/
class Constraint extends Serializable {
getConfig() {
return {};
}
}
class MaxNorm extends Constraint {
constructor(args) {
super();
this.defaultMaxValue = 2;
this.defaultAxis = 0;
this.maxValue =
args.maxValue != null ? args.maxValue : this.defaultMaxValue;
this.axis = args.axis != null ? args.axis : this.defaultAxis;
}
apply(w) {
return tidy(() => {
const norms = calcL2Norms(w, this.axis);
const desired = clipByValue$2(norms, 0, this.maxValue);
return mul(w, div$1(desired, add$1(epsilon(), norms)));
});
}
getConfig() {
return { maxValue: this.maxValue, axis: this.axis };
}
}
/** @nocollapse */
MaxNorm.className = 'MaxNorm';
registerClass(MaxNorm);
class UnitNorm extends Constraint {
constructor(args) {
super();
this.defaultAxis = 0;
this.axis = args.axis != null ? args.axis : this.defaultAxis;
}
apply(w) {
return tidy(() => div$1(w, add$1(epsilon(), calcL2Norms(w, this.axis))));
}
getConfig() {
return { axis: this.axis };
}
}
/** @nocollapse */
UnitNorm.className = 'UnitNorm';
registerClass(UnitNorm);
class NonNeg extends Constraint {
apply(w) {
return relu$2(w);
}
}
/** @nocollapse */
NonNeg.className = 'NonNeg';
registerClass(NonNeg);
class MinMaxNorm extends Constraint {
constructor(args) {
super();
this.defaultMinValue = 0.0;
this.defaultMaxValue = 1.0;
this.defaultRate = 1.0;
this.defaultAxis = 0;
this.minValue =
args.minValue != null ? args.minValue : this.defaultMinValue;
this.maxValue =
args.maxValue != null ? args.maxValue : this.defaultMaxValue;
this.rate = args.rate != null ? args.rate : this.defaultRate;
this.axis = args.axis != null ? args.axis : this.defaultAxis;
}
apply(w) {
return tidy(() => {
const norms = calcL2Norms(w, this.axis);
const desired = add$1(mul(this.rate, clipByValue$2(norms, this.minValue, this.maxValue)), mul(1.0 - this.rate, norms));
return mul(w, div$1(desired, add$1(epsilon(), norms)));
});
}
getConfig() {
return {
minValue: this.minValue,
maxValue: this.maxValue,
rate: this.rate,
axis: this.axis
};
}
}
/** @nocollapse */
MinMaxNorm.className = 'MinMaxNorm';
registerClass(MinMaxNorm);
// Maps the JavaScript-like identifier keys to the corresponding registry
// symbols.
const CONSTRAINT_IDENTIFIER_REGISTRY_SYMBOL_MAP = {
'maxNorm': 'MaxNorm',
'minMaxNorm': 'MinMaxNorm',
'nonNeg': 'NonNeg',
'unitNorm': 'UnitNorm'
};
function serializeConstraint(constraint) {
return serializeKerasObject(constraint);
}
function deserializeConstraint(config, customObjects = {}) {
return deserializeKerasObject(config, SerializationMap.getMap().classNameMap, customObjects, 'constraint');
}
function getConstraint(identifier) {
if (identifier == null) {
return null;
}
if (typeof identifier === 'string') {
const className = identifier in CONSTRAINT_IDENTIFIER_REGISTRY_SYMBOL_MAP ?
CONSTRAINT_IDENTIFIER_REGISTRY_SYMBOL_MAP[identifier] :
identifier;
const config = { className, config: {} };
return deserializeConstraint(config);
}
else if (identifier instanceof Constraint) {
return identifier;
}
else {
return deserializeConstraint(identifier);
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
// tslint:disable-next-line:max-line-length
/**
* Glorot uniform initializer, also called Xavier uniform initializer.
* It draws samples from a uniform distribution within [-limit, limit]
* where `limit` is `sqrt(6 / (fan_in + fan_out))`
* where `fan_in` is the number of input units in the weight tensor
* and `fan_out` is the number of output units in the weight tensor
*
* Reference:
* Glorot & Bengio, AISTATS 2010
* http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf.
*
* @doc {heading: 'Initializers', namespace: 'initializers'}
*/
function glorotUniform(args) {
return new GlorotUniform(args);
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Turn any Scalar values in a Logs object into actual number values.
*
* @param logs The `Logs` object to be resolved in place.
*/
async function resolveScalarsInLogs(logs) {
if (logs == null) {
return;
}
const promises = [];
const keys = [];
const scalarsToDispose = [];
for (const key in logs) {
const value = logs[key];
if (typeof value !== 'number') {
const valueScalar = value;
promises.push(valueScalar.data());
keys.push(key);
scalarsToDispose.push(valueScalar);
}
}
if (promises.length > 0) {
const values = await Promise.all(promises);
for (let i = 0; i < values.length; ++i) {
logs[keys[i]] = values[i][0];
}
// Dispose the original scalar tensors.
dispose(scalarsToDispose);
}
}
/**
* Dispose all Tensors in an UnresolvedLogs object.
*
* @param logs An `UnresolvedLogs` object potentially containing `tf.Tensor`s in
* places where the values can be `tf.Tensor` or `number`.
*/
function disposeTensorsInLogs(logs) {
if (logs == null) {
return;
}
for (const key in logs) {
const value = logs[key];
if (typeof value !== 'number') {
value.dispose();
}
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original source: keras/callbacks.py */
/** Verbosity logging level when fitting a model. */
var ModelLoggingVerbosity;
(function (ModelLoggingVerbosity) {
ModelLoggingVerbosity[ModelLoggingVerbosity["SILENT"] = 0] = "SILENT";
ModelLoggingVerbosity[ModelLoggingVerbosity["VERBOSE"] = 1] = "VERBOSE";
})(ModelLoggingVerbosity || (ModelLoggingVerbosity = {}));
/** How often to yield to the main thread when training (in ms). */
const DEFAULT_YIELD_EVERY_MS = 125;
/**
* Abstract base class used to build new callbacks.
*
* The `logs` dictionary that callback methods take as argument will contain
* keys for quantities relevant to the current batch or epoch.
*
* Currently, the `.fit()` method of the `Sequential` model class
* will include the following quantities in the `logs` that
* it passes to its callbacks:
*
* onEpochEnd: Logs include `acc` and `loss`, and optionally include `valLoss`
* (if validation is enabled in `fit`), and `valAcc` (if validation and
* accuracy monitoring are enabled).
* onBatchBegin: Logs include `size`, the number of samples in the current
* batch.
* onBatchEnd: Logs include `loss`, and optionally `acc` (if accuracy monitoring
* is enabled).
*/
class BaseCallback {
constructor() {
// TODO(michaelterry): This type is a best guess.
this.validationData = null;
}
setParams(params) {
this.params = params;
}
async onEpochBegin(epoch, logs) { }
async onEpochEnd(epoch, logs) { }
async onBatchBegin(batch, logs) { }
async onBatchEnd(batch, logs) { }
async onTrainBegin(logs) { }
async onTrainEnd(logs) { }
// LayersModel needs to call Callback.setModel(), but cannot actually depend
// on Callback because that creates a cyclic dependency. Providing this no-op
// method on BaseCallback breaks the cycle: this way LayersModel can depend on
// BaseCallback but not on Callback. The argument is typed as `Container`
// (the superclass of LayersModel) to avoid recapitulating the cycle. Callback
// overrides this method and enforces that the argument is really a
// LayersModel.
setModel(model) {
// Do nothing. Use Callback instead of BaseCallback to track the model.
}
}
/**
* Container abstracting a list of callbacks.
*/
class CallbackList {
// TODO(cais): When the need arises, uncomment the following lines and
// implement the queue for time values.
// private deltaTBatch: number;
// private deltaTsBatchBegin: Array<number>;
// private deltaTsBatchEnd: Array<number>;
/**
* Constructor of CallbackList.
* @param callbacks Array of `Callback` instances.
* @param queueLength Queue length for keeping running statistics over
* callback execution time.
*/
constructor(callbacks, queueLength = 10) {
// TODO(cais): Make use of queueLength when implementing the queue for time
// values.
if (callbacks == null) {
callbacks = [];
}
this.callbacks = callbacks;
this.queueLength = queueLength;
}
append(callback) {
this.callbacks.push(callback);
}
setParams(params) {
for (const callback of this.callbacks) {
callback.setParams(params);
}
}
setModel(model) {
for (const callback of this.callbacks) {
callback.setModel(model);
}
}
/**
* Called at the start of an epoch.
* @param epoch Index of epoch.
* @param logs Dictionary of logs.
*/
async onEpochBegin(epoch, logs) {
if (logs == null) {
logs = {};
}
for (const callback of this.callbacks) {
await callback.onEpochBegin(epoch, logs);
}
}
/**
* Called at the end of an epoch.
* @param epoch Index of epoch.
* @param logs Dictionary of logs.
*/
async onEpochEnd(epoch, logs) {
if (logs == null) {
logs = {};
}
for (const callback of this.callbacks) {
await callback.onEpochEnd(epoch, logs);
}
}
/**
* Called right before processing a batch.
* @param batch Index of batch within the current epoch.
* @param logs Dictionary of logs.
*/
async onBatchBegin(batch, logs) {
if (logs == null) {
logs = {};
}
for (const callback of this.callbacks) {
await callback.onBatchBegin(batch, logs);
}
}
/**
* Called at the end of a batch.
* @param batch Index of batch within the current epoch.
* @param logs Dictionary of logs.
*/
async onBatchEnd(batch, logs) {
if (logs == null) {
logs = {};
}
for (const callback of this.callbacks) {
await callback.onBatchEnd(batch, logs);
}
}
/**
* Called at the beginning of training.
* @param logs Dictionary of logs.
*/
async onTrainBegin(logs) {
if (logs == null) {
logs = {};
}
for (const callback of this.callbacks) {
await callback.onTrainBegin(logs);
}
}
/**
* Called at the end of training.
* @param logs Dictionary of logs.
*/
async onTrainEnd(logs) {
if (logs == null) {
logs = {};
}
for (const callback of this.callbacks) {
await callback.onTrainEnd(logs);
}
}
}
/**
* Callback that accumulates epoch averages of metrics.
*
* This callback is automatically applied to every LayersModel.
*/
class BaseLogger extends BaseCallback {
constructor() {
super();
}
async onEpochBegin(epoch) {
this.seen = 0;
this.totals = {};
}
async onBatchEnd(batch, logs) {
if (logs == null) {
logs = {};
}
const batchSize = logs['size'] == null ? 0 : logs['size'];
this.seen += batchSize;
for (const key in logs) {
const value = logs[key];
if (typeof value === 'number') {
if (!this.totals.hasOwnProperty(key)) {
this.totals[key] = 0;
}
this.totals[key] = this.totals[key] + value * batchSize;
}
else {
let oldTotalsToDispose;
if (key in this.totals) {
oldTotalsToDispose = this.totals[key];
}
else {
this.totals[key] = 0;
}
const total = tidy(() => add$1((this.totals[key]), mul(value, batchSize)));
this.totals[key] = total;
if (oldTotalsToDispose != null) {
oldTotalsToDispose.dispose();
}
}
}
}
async onEpochEnd(epoch, logs) {
if (logs != null) {
for (const key of this.params['metrics']) {
if (this.totals[key] == null) {
continue;
}
if (typeof this.totals[key] === 'number') {
logs[key] = this.totals[key] / this.seen;
}
else {
tidy(() => {
const log = mul(div$1(1, this.seen), this.totals[key]);
logs[key] = log;
this.totals[key].dispose();
keep(logs[key]);
});
}
}
}
}
}
/**
* Callback that records events into a `History` object. This callback is
* automatically applied to every TF.js Layers model. The `History` object
* gets returned by the `fit` method of models.
*/
class History extends BaseCallback {
async onTrainBegin(logs) {
this.epoch = [];
this.history = {};
}
async onEpochEnd(epoch, logs) {
if (logs == null) {
logs = {};
}
this.epoch.push(epoch);
for (const key in logs) {
if (this.history[key] == null) {
this.history[key] = [];
}
this.history[key].push(logs[key]);
}
}
/**
* Await the values of all losses and metrics.
*/
async syncData() {
const promises = [];
const keys = [];
const indices = [];
for (const key in this.history) {
const valueArray = this.history[key];
for (let i = 0; i < valueArray.length; ++i) {
if (typeof valueArray[i] !== 'number') {
const valueScalar = valueArray[i];
promises.push(valueScalar.data());
keys.push(key);
indices.push(i);
}
}
}
const values = await Promise.all(promises);
for (let n = 0; n < values.length; ++n) {
const tensorToDispose = this.history[keys[n]][indices[n]];
tensorToDispose.dispose();
this.history[keys[n]][indices[n]] = values[n][0];
}
}
}
/**
* Custom callback for training.
*/
class CustomCallback extends BaseCallback {
constructor(args, yieldEvery) {
super();
this.currentEpoch = 0;
this.nowFunc = args.nowFunc;
this.nextFrameFunc = args.nextFrameFunc || nextFrame;
this.yieldEvery = yieldEvery || 'auto';
if (this.yieldEvery === 'auto') {
this.yieldEvery = DEFAULT_YIELD_EVERY_MS;
}
if (this.yieldEvery === 'never' && args.onYield != null) {
throw new Error('yieldEvery is `never` but you provided an `onYield` callback. ' +
'Either change `yieldEvery` or remove the callback');
}
if (isNumber(this.yieldEvery)) {
// Decorate `maybeWait` so it will be called at most once every
// `yieldEvery` ms.
this.maybeWait = debounce(this.maybeWait.bind(this), this.yieldEvery, this.nowFunc);
}
this.trainBegin = args.onTrainBegin;
this.trainEnd = args.onTrainEnd;
this.epochBegin = args.onEpochBegin;
this.epochEnd = args.onEpochEnd;
this.batchBegin = args.onBatchBegin;
this.batchEnd = args.onBatchEnd;
this.yield = args.onYield;
}
async maybeWait(epoch, batch, logs) {
const ps = [];
if (this.yield != null) {
await resolveScalarsInLogs(logs);
ps.push(this.yield(epoch, batch, logs));
}
ps.push(this.nextFrameFunc());
await Promise.all(ps);
}
async onEpochBegin(epoch, logs) {
this.currentEpoch = epoch;
if (this.epochBegin != null) {
await resolveScalarsInLogs(logs);
await this.epochBegin(epoch, logs);
}
}
async onEpochEnd(epoch, logs) {
const ps = [];
if (this.epochEnd != null) {
await resolveScalarsInLogs(logs);
ps.push(this.epochEnd(epoch, logs));
}
if (this.yieldEvery === 'epoch') {
ps.push(this.nextFrameFunc());
}
await Promise.all(ps);
}
async onBatchBegin(batch, logs) {
if (this.batchBegin != null) {
await resolveScalarsInLogs(logs);
await this.batchBegin(batch, logs);
}
}
async onBatchEnd(batch, logs) {
const ps = [];
if (this.batchEnd != null) {
await resolveScalarsInLogs(logs);
ps.push(this.batchEnd(batch, logs));
}
if (this.yieldEvery === 'batch') {
ps.push(this.nextFrameFunc());
}
else if (isNumber(this.yieldEvery)) {
ps.push(this.maybeWait(this.currentEpoch, batch, logs));
}
await Promise.all(ps);
}
async onTrainBegin(logs) {
if (this.trainBegin != null) {
await resolveScalarsInLogs(logs);
await this.trainBegin(logs);
}
}
async onTrainEnd(logs) {
if (this.trainEnd != null) {
await resolveScalarsInLogs(logs);
await this.trainEnd(logs);
}
}
}
/**
* Standardize callbacks or configurations of them to an Array of callbacks.
*/
function standardizeCallbacks(callbacks, yieldEvery) {
if (callbacks == null) {
callbacks = {};
}
if (callbacks instanceof BaseCallback) {
return [callbacks];
}
if (Array.isArray(callbacks) && callbacks[0] instanceof BaseCallback) {
return callbacks;
}
// Convert custom callback configs to custom callback objects.
const callbackConfigs = toList(callbacks);
return callbackConfigs.map(callbackConfig => new CustomCallback(callbackConfig, yieldEvery));
}
/**
* A global registry for callback constructors to be used during
* LayersModel.fit().
*/
class CallbackConstructorRegistry {
/**
* Blocks public access to constructor.
*/
constructor() { }
/**
* Register a tf.LayersModel.fit() callback constructor.
*
* The registered callback constructor will be used to instantiate
* callbacks for every tf.LayersModel.fit() call afterwards.
*
* @param verbosityLevel Level of verbosity at which the `callbackConstructor`
* is to be reigstered.
* @param callbackConstructor A no-arg constructor for `tf.Callback`.
* @throws Error, if the same callbackConstructor has been registered before,
* either at the same or a different `verbosityLevel`.
*/
static registerCallbackConstructor(verbosityLevel, callbackConstructor) {
assert$1(verbosityLevel >= 0 && Number.isInteger(verbosityLevel), () => `Verbosity level is expected to be an integer >= 0, ` +
`but got ${verbosityLevel}`);
CallbackConstructorRegistry.checkForDuplicate(callbackConstructor);
if (CallbackConstructorRegistry.constructors[verbosityLevel] == null) {
CallbackConstructorRegistry.constructors[verbosityLevel] = [];
}
CallbackConstructorRegistry.constructors[verbosityLevel].push(callbackConstructor);
}
static checkForDuplicate(callbackConstructor) {
for (const levelName in CallbackConstructorRegistry.constructors) {
const constructors = CallbackConstructorRegistry.constructors[+levelName];
constructors.forEach(ctor => {
if (ctor === callbackConstructor) {
throw new ValueError('Duplicate callback constructor.');
}
});
}
}
/**
* Clear all registered callback constructors.
*/
static clear() {
CallbackConstructorRegistry.constructors = {};
}
/**
* Create callbacks using the registered callback constructors.
*
* Given `verbosityLevel`, all constructors registered at that level or above
* will be called and the instantiated callbacks will be used.
*
* @param verbosityLevel: Level of verbosity.
*/
static createCallbacks(verbosityLevel) {
const constructors = [];
for (const levelName in CallbackConstructorRegistry.constructors) {
const level = +levelName;
if (verbosityLevel >= level) {
constructors.push(...CallbackConstructorRegistry.constructors[level]);
}
}
return constructors.map(ctor => new ctor());
}
}
CallbackConstructorRegistry.constructors = {};
function configureCallbacks(callbacks, verbose, epochs, initialEpoch, numTrainSamples, stepsPerEpoch, batchSize, doValidation, callbackMetrics) {
const history = new History();
const actualCallbacks = [
new BaseLogger(), ...CallbackConstructorRegistry.createCallbacks(verbose)
];
if (callbacks != null) {
actualCallbacks.push(...callbacks);
}
actualCallbacks.push(history);
const callbackList = new CallbackList(actualCallbacks);
// TODO(cais): Figure out when this LayersModel instance can have a
// dynamically
// set property called 'callback_model' as in PyKeras.
callbackList.setParams({
epochs,
initialEpoch,
samples: numTrainSamples,
steps: stepsPerEpoch,
batchSize,
verbose,
doValidation,
metrics: callbackMetrics,
});
return { callbackList, history };
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original Source layers/__init__.py */
/**
* Instantiate a layer from a config dictionary.
* @param config dict of the form {class_name: str, config: dict}
* @param customObjects dict mapping class names (or function names)
* of custom (non-Keras) objects to class/functions
* @param fastWeightInit Optional flag to use fast weight initialization
* during deserialization. This is applicable to cases in which
* the initialization will be immediately overwritten by loaded weight
* values. Default: `false`.
* @returns Layer instance (may be LayersModel, Sequential, Layer...)
*/
function deserialize(config, customObjects = {}, fastWeightInit = false) {
return deserializeKerasObject(config, SerializationMap.getMap().classNameMap, customObjects, 'layer', fastWeightInit);
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original Source: losses.py */
/**
* Normalizes a tensor wrt the L2 norm alongside the specified axis.
* @param x
* @param axis Axis along which to perform normalization.
*/
function l2Normalize(x, axis) {
return tidy(() => {
if (x.dtype !== 'float32') {
x = cast$3(x, 'float32');
}
const squareSum = sum$2(square(x), axis, true);
const epsilonTensor = fill$2(squareSum.shape, epsilon());
const norm = sqrt$2(maximum$2(squareSum, epsilonTensor));
return div$1(x, norm);
});
}
function meanSquaredError(yTrue, yPred) {
return tidy(() => mean$1(square(sub$2(yPred, yTrue)), -1));
}
function meanAbsoluteError(yTrue, yPred) {
return tidy(() => mean$1(abs$2(sub$2(yPred, yTrue)), -1));
}
function meanAbsolutePercentageError(yTrue, yPred) {
return tidy(() => {
const diff = sub$2(yTrue, yPred);
const clippedTrue = clipByValue$2(abs$2(yTrue), epsilon(), Number.MAX_VALUE);
const absResult = abs$2(div$1(diff, clippedTrue));
return mul(100, mean$1(absResult, -1));
});
}
function meanSquaredLogarithmicError(yTrue, yPred) {
return tidy(() => {
const clippedPred = clipByValue$2(yPred, epsilon(), Number.MAX_VALUE);
const firstLog = log$2(add$1(1, clippedPred));
const clippedTrue = clipByValue$2(yTrue, epsilon(), Number.MAX_VALUE);
const secondLog = log$2(add$1(1, clippedTrue));
return mean$1(square(sub$2(firstLog, secondLog)), -1);
});
}
function squaredHinge(yTrue, yPred) {
return tidy(() => {
const maxResult = maximum$2(0, sub$2(1, mul(yTrue, yPred)));
return mean$1(square(maxResult), -1);
});
}
function hinge(yTrue, yPred) {
return tidy(() => {
const maxResult = maximum$2(0, sub$2(1, mul(yTrue, yPred)));
return mean$1(maxResult, -1);
});
}
function categoricalHinge(yTrue, yPred) {
return tidy(() => {
const pos = sum$2(mul(yTrue, yPred), -1);
const neg = max$2(mul(sub$2(1, yTrue), yPred), -1);
return maximum$2(0, add$1(1, sub$2(neg, pos)));
});
}
/**
* Logarithm of the hyperbolic cosine of the prediction error.
*
* `log(cosh(x))` is approximately equal to `(x ** 2) / 2` for small `x` and
* to `abs(x) - log(2)` for large `x`. This means that 'logcosh' works mostly
* like the mean squared error, but will not be so strongly affected by the
* occasional wildly incorrect prediction.
*/
function logcosh(yTrue, yPred) {
return tidy(() => {
const log2 = Math.log(2);
const predictionDiff = sub$2(yPred, yTrue);
const logcoshResult = sub$2(add$1(predictionDiff, softplus$2(mul(-2, predictionDiff))), log2);
return mean$1(logcoshResult, -1);
});
}
function categoricalCrossentropy$1(target, output, fromLogits = false) {
return tidy(() => {
if (fromLogits) {
output = softmax$2(output);
}
else {
// scale preds so that the class probabilities of each sample sum to 1.
const outputSum = sum$2(output, output.shape.length - 1, true);
output = div$1(output, outputSum);
}
output = clipByValue$2(output, epsilon(), 1 - epsilon());
return neg$2(sum$2(mul(cast$3(target, 'float32'), log$2(output)), output.shape.length - 1));
});
}
/**
* Categorical crossentropy with integer targets.
*
* @param target An integer tensor.
* @param output A tensor resulting from a softmax (unless `fromLogits` is
* `true`, in which case `output` is expected to be the logits).
* @param fromLogits Boolean, whether `output` is the result of a softmax, or is
* a tensor of logits.
*/
function sparseCategoricalCrossentropy$1(target, output, fromLogits = false) {
return tidy(() => {
const flatTarget = cast$3(floor$2(flatten(target)), 'int32');
output = clipByValue$2(output, epsilon(), 1 - epsilon());
const outputShape = output.shape;
const oneHotTarget = reshape$2(oneHot$2(flatTarget, outputShape[outputShape.length - 1]), outputShape);
return categoricalCrossentropy$1(oneHotTarget, output, fromLogits);
});
}
/**
* From TensorFlow's implementation in nn_impl.py:
*
* For brevity, let `x = logits`, `z = labels`. The logistic loss is
* z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
* = z * -log(1 / (1 + exp(-x))) + (1 - z) * -log(exp(-x) / (1 + exp(-x)))
* = z * log(1 + exp(-x)) + (1 - z) * (-log(exp(-x)) + log(1 + exp(-x)))
* = z * log(1 + exp(-x)) + (1 - z) * (x + log(1 + exp(-x))
* = (1 - z) * x + log(1 + exp(-x))
* = x - x * z + log(1 + exp(-x))
* For x < 0, to avoid overflow in exp(-x), we reformulate the above
* x - x * z + log(1 + exp(-x))
* = log(exp(x)) - x * z + log(1 + exp(-x))
* = - x * z + log(1 + exp(x))
* Hence, to ensure stability and avoid overflow, the implementation uses this
* equivalent formulation
* max(x, 0) - x * z + log(1 + exp(-abs(x)))
*
* @param labels The labels.
* @param logits The logits.
*/
function sigmoidCrossEntropyWithLogits(labels, logits) {
if (!arraysEqual(labels.shape, logits.shape)) {
throw new ValueError(`logits and labels must have the same shape, but got shapes ` +
`${JSON.stringify(labels.shape)} and ${JSON.stringify(logits.shape)}`);
}
return tidy(() => {
// The logistic loss formula from above is
// x - x * z + log(1 + exp(-x))
// For x < 0, a more numerically stable formula is
// -x * z + log(1 + exp(x))
// Note that these two expressions can be combined into the following:
// max(x, 0) - x * z + log(1 + exp(-abs(x)))
const reluLogits = relu$2(logits);
const negAbsLogits = neg$2(abs$2(logits));
return add$1(sub$2(reluLogits, mul(logits, labels)), log1p$2(exp$2(negAbsLogits)));
});
}
function binaryCrossentropy$1(yTrue, yPred) {
return tidy(() => {
let y;
y = clipByValue$2(yPred, epsilon(), 1 - epsilon());
y = log$2(div$1(y, sub$2(1, y)));
return mean$1(sigmoidCrossEntropyWithLogits(yTrue, y), -1);
});
}
function kullbackLeiblerDivergence(yTrue, yPred) {
return tidy(() => {
const clippedTrue = clipByValue$2(yTrue, epsilon(), 1);
const clippedPred = clipByValue$2(yPred, epsilon(), 1);
return sum$2(mul(yTrue, log$2(div$1(clippedTrue, clippedPred))), -1);
});
}
function poisson(yTrue, yPred) {
return tidy(() => {
const logPred = log$2(add$1(epsilon(), yPred));
return mean$1(sub$2(yPred, mul(yTrue, logPred)), -1);
});
}
function cosineProximity(yTrue, yPred) {
return tidy(() => {
const trueNormalized = l2Normalize(yTrue, -1);
const predNormalized = l2Normalize(yPred, -1);
const trueXPred = mul(trueNormalized, predNormalized);
return neg$2(sum$2(trueXPred, -1));
});
}
// TODO(michaelterry): Add deserialize() function.
const lossesMap = {
meanSquaredError,
meanAbsoluteError,
meanAbsolutePercentageError,
meanSquaredLogarithmicError,
squaredHinge,
hinge,
categoricalHinge,
logcosh,
categoricalCrossentropy: categoricalCrossentropy$1,
sparseCategoricalCrossentropy: sparseCategoricalCrossentropy$1,
binaryCrossentropy: binaryCrossentropy$1,
kullbackLeiblerDivergence,
poisson,
cosineProximity
};
// Porting note: This diverges from the PyKeras implementation and may need to
// change based on (de)serialization requirements.
function get$1(identifierOrFn) {
if (typeof identifierOrFn === 'string') {
if (identifierOrFn in lossesMap) {
return lossesMap[identifierOrFn];
}
let errMsg = `Unknown loss ${identifierOrFn}`;
if (identifierOrFn.toLowerCase().includes('softmaxcrossentropy')) {
errMsg = `Unknown loss ${identifierOrFn}. ` +
'Use "categoricalCrossentropy" as the string name for ' +
'tf.losses.softmaxCrossEntropy';
}
throw new ValueError(errMsg);
}
else {
return identifierOrFn;
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Built-in metrics.
*/
function binaryAccuracy(yTrue, yPred) {
return tidy(() => {
const threshold = mul(.5, onesLike$2(yPred));
const yPredThresholded = cast(greater$2(yPred, threshold), yTrue.dtype);
return mean$1(equal$2(yTrue, yPredThresholded), -1);
});
}
function categoricalAccuracy(yTrue, yPred) {
return tidy(() => cast(equal$2(argMax$2(yTrue, -1), argMax$2(yPred, -1)), 'float32'));
}
function truePositives(yTrue, yPred) {
return tidy(() => {
return cast$3(sum$2(logicalAnd$2(equal$2(yTrue, 1), equal$2(yPred, 1))), 'float32');
});
}
function falsePositives(yTrue, yPred) {
return tidy(() => {
return cast$3(sum$2(logicalAnd$2(equal$2(yTrue, 0), equal$2(yPred, 1))), 'float32');
});
}
function precision(yTrue, yPred) {
return tidy(() => {
const tp = truePositives(yTrue, yPred);
const fp = falsePositives(yTrue, yPred);
const denominator = add$1(tp, fp);
return cast$3(where(greater$2(denominator, 0), div$1(tp, denominator), 0), 'float32');
});
}
function binaryCrossentropy(yTrue, yPred) {
return binaryCrossentropy$1(yTrue, yPred);
}
function sparseCategoricalAccuracy(yTrue, yPred) {
if (yTrue.rank === yPred.rank) {
yTrue = squeeze(yTrue, [yTrue.rank - 1]);
}
yPred = argMax$2(yPred, -1);
if (yPred.dtype !== yTrue.dtype) {
yPred = cast$3(yPred, yTrue.dtype);
}
return cast$3(equal$2(yTrue, yPred), 'float32');
}
// Aliases.
const mse = meanSquaredError;
const MSE = meanSquaredError;
const mae = meanAbsoluteError;
const MAE = meanAbsoluteError;
const mape = meanAbsolutePercentageError;
const MAPE = meanAbsolutePercentageError;
const categoricalCrossentropy = categoricalCrossentropy$1;
const cosine = cosineProximity;
const sparseCategoricalCrossentropy = sparseCategoricalCrossentropy$1;
// TODO(cais, nielsene): Add serialize().
const metricsMap = {
binaryAccuracy,
categoricalAccuracy,
precision,
categoricalCrossentropy,
sparseCategoricalCrossentropy,
mse,
MSE,
mae,
MAE,
mape,
MAPE,
cosine
};
function get(identifier) {
if (typeof identifier === 'string' && identifier in metricsMap) {
return metricsMap[identifier];
}
else if (typeof identifier !== 'string' && identifier != null) {
return identifier;
}
else {
throw new ValueError(`Unknown metric ${identifier}`);
}
}
/**
* Get the shortcut function name.
*
* If the fn name is a string,
* directly return the string name.
* If the function is included in metricsMap or lossesMap,
* return key of the map.
* - If the function relative to multiple keys,
* return the first found key as the function name.
* - If the function exists in both lossesMap and metricsMap,
* search lossesMap first.
* If the function is not included in metricsMap or lossesMap,
* return the function name.
*
* @param fn loss function, metric function, or short cut name.
* @returns Loss or Metric name in string.
*/
function getLossOrMetricName(fn) {
assert(fn !== null, `Unknown LossOrMetricFn ${fn}`);
if (typeof fn === 'string') {
return fn;
}
else {
let fnName;
for (const key of Object.keys(lossesMap)) {
if (lossesMap[key] === fn) {
fnName = key;
break;
}
}
if (fnName !== undefined) {
return fnName;
}
for (const key of Object.keys(metricsMap)) {
if (metricsMap[key] === fn) {
fnName = key;
break;
}
}
if (fnName !== undefined) {
return fnName;
}
return fn.name;
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Optimizers.
*/
// Add (de)serialize()
// Porting note: This diverges from the PyKeras implementation and may need to
// change based on (de)serialization requirements.
function getOptimizer(identifier) {
const optimizerMap = {
'Adagrad': () => train.adagrad(0.01),
'Adadelta': () => train.adadelta(1, 0.95, epsilon()),
'Adam': () => train.adam(0.001, 0.9, 0.999, epsilon()),
'Adamax': () => train.adamax(0.002, 0.9, 0.999, epsilon(), 0),
'RMSProp': () => train.rmsprop(0.001, 0.9, 0, epsilon()),
'SGD': () => train.sgd(0.01)
};
optimizerMap['adagrad'] = optimizerMap['Adagrad'];
optimizerMap['adadelta'] = optimizerMap['Adadelta'];
optimizerMap['adam'] = optimizerMap['Adam'];
optimizerMap['adamax'] = optimizerMap['Adamax'];
optimizerMap['rmsprop'] = optimizerMap['RMSProp'];
optimizerMap['sgd'] = optimizerMap['SGD'];
if (identifier in optimizerMap) {
return optimizerMap[identifier]();
}
throw new ValueError(`Unknown Optimizer ${identifier}`);
}
/**
* @license
* Copyright 2019 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/** Utility functions related to user-defined metadata. */
// Maximum recommended serialized size for user-defined metadata.
// Beyond this limit, a warning message will be printed during model loading and
// saving.
const MAX_USER_DEFINED_METADATA_SERIALIZED_LENGTH = 1 * 1024 * 1024;
/**
* Check validity of user-defined metadata.
*
* @param userDefinedMetadata
* @param modelName Name of the model that the user-defined metadata belongs to.
* Used during construction of error messages.
* @param checkSize Whether to check the size of the metadata is under
* recommended limit. Default: `false`. If `true`, will try stringify the
* JSON object and print a console warning if the serialzied size is above the
* limit.
* @throws Error if `userDefinedMetadata` is not a plain JSON object.
*/
function checkUserDefinedMetadata(userDefinedMetadata, modelName, checkSize = false) {
if (userDefinedMetadata == null ||
typeof userDefinedMetadata !== 'object' ||
Object.getPrototypeOf(userDefinedMetadata) !== Object.prototype ||
!plainObjectCheck(userDefinedMetadata)) {
throw new Error('User-defined metadata is expected to be a JSON object, but is not.');
}
if (checkSize) {
const out = JSON.stringify(userDefinedMetadata);
if (out.length > MAX_USER_DEFINED_METADATA_SERIALIZED_LENGTH) {
console.warn(`User-defined metadata of model "${modelName}" is too large in ` +
`size (length=${out.length} when serialized). It is not ` +
`recommended to store such large objects in user-defined metadata. ` +
`Please make sure its serialized length is <= ` +
`${MAX_USER_DEFINED_METADATA_SERIALIZED_LENGTH}.`);
}
}
}
/**
* Check if an input is plain JSON object or any valid subfield of it.
*
* @param x The input to be checked.
* @param assertObject Whether to assert `x` is a JSON object, i.e., reject
* cases of arrays and primitives.
* @return Returns `true` if and only if `x` is a plain JSON object,
* a JSON-valid primitive including string, number, boolean and null,
* or an array of the said types.
*/
// tslint:disable-next-line:no-any
function plainObjectCheck(x) {
if (x === null) {
// Note: typeof `null` is 'object', and `null` is valid in JSON.
return true;
}
else if (typeof x === 'object') {
if (Object.getPrototypeOf(x) === Object.prototype) {
// `x` is a JavaScript object and its prototype is Object.
const keys = Object.keys(x);
for (const key of keys) {
if (typeof key !== 'string') {
// JSON keys must be strings.
return false;
}
if (!plainObjectCheck(x[key])) { // Recursive call.
return false;
}
}
return true;
}
else {
// `x` is a JavaScript object but its prototype is not Object.
if (Array.isArray(x)) {
// `x` is a JavaScript array.
for (const item of x) {
if (!plainObjectCheck(item)) { // Recursive call.
return false;
}
}
return true;
}
else {
// `x` is a JavaScript object and its prototype is not Object,
// and it's not an Array. I.e., it's a complex object such as
// `Error` and `Date`.
return false;
}
}
}
else {
// `x` is not a JavaScript object or `null`.
const xType = typeof x;
return xType === 'string' || xType === 'number' || xType === 'boolean';
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Print the summary of a LayersModel object.
*
* @param model tf.LayersModel instance.
* @param lineLength Total length of printed lines. Set this to adapt to the
* display to different terminal or console sizes.
* @param positions Relative or absolute positions of log elements in each
* line. Each number corresponds to right-most (i.e., ending) position of a
* column.
* If not provided, defaults to `[0.45, 0.85, 1]` for sequential-like
* models and `[0.33, 0.55, 0.67, 1]` for non-sequential like models.
* @param printFn Print function to use.
* It will be called on each line of the summary. You can provide a custom
* function in order to capture the string summary. Defaults to `console.log`.
*/
function printSummary(model, lineLength, positions,
// tslint:disable-next-line:no-any
printFn = console.log) {
const sequentialLike = isModelSequentialLike(model);
// Header names for different log elements.
const toDisplay = ['Layer (type)', 'Input Shape', 'Output shape', 'Param #'];
if (sequentialLike) {
lineLength = lineLength || 90;
positions = positions || [0.32, 0.61, 0.89, 1];
}
else {
lineLength = lineLength || 115;
positions = positions || [0.24, 0.48, 0.70, 0.80, 1];
// Header names for different log elements.
}
if (positions[positions.length - 1] <= 1) {
// `positions` is relative. Convert it to absolute positioning.
positions = positions.map(p => Math.floor(lineLength * p));
}
let relevantNodes;
if (!sequentialLike) {
toDisplay.push('Receives inputs');
relevantNodes = [];
for (const depth in model.nodesByDepth) {
relevantNodes.push(...model.nodesByDepth[depth]);
}
}
printFn('_'.repeat(lineLength));
printRow(toDisplay, positions, printFn);
printFn('='.repeat(lineLength));
const layers = model.layers;
for (let i = 0; i < layers.length; ++i) {
if (sequentialLike) {
printLayerSummary(layers[i], positions, printFn);
}
else {
printLayerSummaryWithConnections(layers[i], positions, relevantNodes, printFn);
}
printFn((i === layers.length - 1 ? '=' : '_').repeat(lineLength));
}
// tslint:disable-next-line:no-any
model.checkTrainableWeightsConsistency();
const trainableCount = countTrainableParams(model);
const nonTrainableCount = countParamsInWeights(model.nonTrainableWeights);
printFn(`Total params: ${trainableCount + nonTrainableCount}`);
printFn(`Trainable params: ${trainableCount}`);
printFn(`Non-trainable params: ${nonTrainableCount}`);
printFn('_'.repeat(lineLength));
}
function countTrainableParams(model) {
let trainableCount;
// tslint:disable:no-any
if (model.collectedTrainableWeights != null) {
trainableCount =
countParamsInWeights(model.collectedTrainableWeights);
}
else {
trainableCount = countParamsInWeights(model.trainableWeights);
}
// tslint:enable:no-any
return trainableCount;
}
function isModelSequentialLike(model) {
let sequentialLike = true;
const nodesByDepth = [];
const nodes = [];
for (const depth in model.nodesByDepth) {
nodesByDepth.push(model.nodesByDepth[depth]);
}
for (const depthNodes of nodesByDepth) {
if (depthNodes.length > 1 ||
depthNodes.length === 1 && depthNodes[0].inboundLayers.length > 1) {
sequentialLike = false;
break;
}
nodes.push(...depthNodes);
}
if (sequentialLike) {
// Search for shared layers.
for (const layer of model.layers) {
let flag = false;
for (const node of layer.inboundNodes) {
if (nodes.indexOf(node) !== -1) {
if (flag) {
sequentialLike = false;
break;
}
else {
flag = true;
}
}
}
if (!sequentialLike) {
break;
}
}
}
return sequentialLike;
}
function printRow(fields, positions,
// tslint:disable-next-line:no-any
printFn = console.log) {
let line = '';
for (let i = 0; i < fields.length; ++i) {
if (i > 0) {
line = line.slice(0, line.length - 1) + ' ';
}
line += fields[i];
line = line.slice(0, positions[i]);
line += ' '.repeat(positions[i] - line.length);
}
printFn(line);
}
/**
* Prints a summary for a single Layer, without connectivity information.
*
* @param layer: Layer instance to print.
*/
function printLayerSummary(layer, positions,
// tslint:disable-next-line:no-any
printFn) {
let outputShape;
let inputShape;
try {
inputShape = (layer.inboundNodes.map(x => JSON.stringify(x.inputShapes))).join(',');
}
catch (err) {
inputShape = 'multiple';
}
try {
outputShape = JSON.stringify(layer.outputShape);
}
catch (err) {
outputShape = 'multiple';
}
const name = layer.name;
const className = layer.getClassName();
const fields = [`${name} (${className})`, inputShape,
outputShape, layer.countParams().toString()];
printRow(fields, positions, printFn);
}
/**
* Prints a summary for a single Layer, with connectivity information.
*/
function printLayerSummaryWithConnections(layer, positions, relevantNodes,
// tslint:disable-next-line:no-any
printFn) {
let outputShape;
let inputShape;
try {
inputShape = (layer.inboundNodes.map(x => JSON.stringify(x.inputShapes))).join(',');
}
catch (err) {
inputShape = 'multiple';
}
try {
outputShape = JSON.stringify(layer.outputShape);
}
catch (err) {
outputShape = 'multiple';
}
const connections = [];
for (const node of layer.inboundNodes) {
if (relevantNodes != null && relevantNodes.length > 0 &&
relevantNodes.indexOf(node) === -1) {
continue;
}
for (let i = 0; i < node.inboundLayers.length; ++i) {
const inboundLayer = node.inboundLayers[i].name;
const inboundLayerIndex = node.nodeIndices[i];
const inboundTensorIndex = node.tensorIndices[i];
connections.push(`${inboundLayer}[${inboundLayerIndex}][${inboundTensorIndex}]`);
}
}
const name = layer.name;
const className = layer.getClassName();
const firstConnection = connections.length === 0 ? '' : connections[0];
const fields = [
`${name} (${className})`, inputShape,
outputShape, layer.countParams().toString(),
firstConnection
];
printRow(fields, positions, printFn);
for (let i = 1; i < connections.length; ++i) {
printRow(['', '', '', '', connections[i]], positions, printFn);
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
// tslint:enable
/**
* Test whether a value in an array is the name of a LayersModel or Layer.
* @param key The key name that the value is found under. Note that the key
* may not be at the level immediately above the value, if the value is in a
* nested array.
* @param index Index of the value in the Array that it is found in.
* @param value The value object.
* @returns A boolean indicating whether value is a name.
*/
function isArrayItemInputOrOutputName(key, index, value) {
return (key === 'inboundNodes' || key === 'outputLayers' ||
key === 'inputLayers') &&
index === 0 && typeof value === 'string';
}
/**
* Convert a Pythonic config object to TypeScript config object.
* @param pythonicConfig The config object to convert.
* @param key Optional key name of the object being converted.
* @returns Result of the conversion.
*/
function convertPythonicToTs(pythonicConfig, key) {
if (pythonicConfig === null) {
return null;
}
else if (typeof pythonicConfig === 'string') {
return toCamelCase(pythonicConfig);
}
else if ((typeof pythonicConfig === 'number') ||
(typeof pythonicConfig === 'boolean')) {
return pythonicConfig;
}
else if (pythonicConfig instanceof Array) {
const tsArray = [];
const arrayLength = pythonicConfig.length;
for (let i = 0; i < arrayLength; ++i) {
const item = pythonicConfig[i];
if (isArrayItemInputOrOutputName(key, i, item)) {
tsArray.push(item);
}
else {
tsArray.push(convertPythonicToTs(item, key));
}
}
return tsArray;
}
else {
const tsDict = {};
for (const pythonicKey of Object.keys(pythonicConfig)) {
const pythonicValue = pythonicConfig[pythonicKey];
if (pythonicKey === 'name' && typeof pythonicValue === 'string') {
// Special case the 'name' key with a string value. Name values, such as
// the names of LayersModel and Layer instances, should not undergo the
// camel-case conversion.
tsDict[pythonicKey] = pythonicValue;
}
else {
const tsKey = toCamelCase(pythonicKey);
tsDict[tsKey] = convertPythonicToTs(pythonicValue, tsKey);
}
}
return tsDict;
}
}
/**
* Convert a TypeScript config object to Python config object.
* @param tsConfig The config object to convert.
* @param key Optional key name of the object being converted.
* @returns Result of the conversion.
*/
function convertTsToPythonic(tsConfig, key) {
if (tsConfig === null || tsConfig === undefined) {
return null;
}
else if (typeof tsConfig === 'string') {
return toSnakeCase(tsConfig);
}
else if ((typeof tsConfig === 'number') || (typeof tsConfig === 'boolean')) {
return tsConfig;
}
else if (tsConfig instanceof Array) {
const pyArray = [];
const arrayLength = tsConfig.length;
for (let i = 0; i < arrayLength; ++i) {
const item = tsConfig[i];
if (isArrayItemInputOrOutputName(key, i, item)) {
pyArray.push(item);
}
else {
pyArray.push(convertTsToPythonic(item, key));
}
}
return pyArray;
}
else {
const pyDict = {};
for (const tsKey of Object.keys(tsConfig)) {
const tsValue = tsConfig[tsKey];
const pyKey = toSnakeCase(tsKey);
if ((tsKey === 'name' || tsKey === 'className') &&
typeof tsValue === 'string') {
// Special case the 'name' key with a string value. Name values, such as
// the names of LayersModel and Layer instances, should not undergo the
// snake-case conversion.
pyDict[pyKey] = tsValue;
}
else {
pyDict[pyKey] = convertTsToPythonic(tsValue, tsKey);
}
}
return pyDict;
}
}
/** @license See the LICENSE file. */
// This code is auto-generated, do not modify this file!
const version = '4.22.0';
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original source: keras/engine/topology.py */
// get weights key from tensor map in order to check if it is from keras v3.
// e.g. dense/0
const isKerasSavedModelFormat = (weights) => {
const keys = Object.keys(weights);
if (keys.length === 0) {
return false;
}
const key = keys[0].split('/');
return !isNaN(parseInt(key[key.length - 1], 10));
};
/**
* A Container is a directed acyclic graph of layers.
*
* It is the topological form of a "model". A LayersModel
* is simply a Container with added training routines.
*
*/
class Container extends Layer {
constructor(args) {
// No args passed to super's constructor.
super({});
this.containerNodes = new Set();
this.name = args.name;
if (this.name == null) {
const prefix = this.getClassName().toLowerCase();
this.name = getUid(prefix);
}
this.supportsMasking = false;
this.trainable_ = true;
// TODO(michaelterry): Initialize perInputLosses/Updates here.
// Container-specific properties.
if (Array.isArray(args.inputs)) {
this.inputs = args.inputs.slice();
}
else {
this.inputs = [args.inputs];
}
if (Array.isArray(args.outputs)) {
this.outputs = args.outputs.slice();
}
else {
this.outputs = [args.outputs];
}
// Check for redundancy in inputs.
if (unique(this.inputs).length !== this.inputs.length) {
throw new ValueError('The list of inputs passed to the model is ' +
'redundant. All inputs should only appear once. Found: ' +
`${this.inputs.map(x => x.name)}`);
}
// Check for redundancy in outputs.
if (unique(this.outputs).length !== this.outputs.length) {
console.warn('The list of outputs passed to the model is redundant. ' +
'All outputs should only appear once. Found: ' +
`${this.outputs.map(x => x.name)}`);
}
/*
List of initial layers (1 to 1 mapping with this.inputs, hence the same
layer might appear twice)
*/
this.inputLayers = [];
this.inputLayersNodeIndices = [];
this.inputLayersTensorIndices = [];
/*
List of layers (1 to 1 mapping with this.outputs, hence the same layer
might appear twice)
*/
this.outputLayers = [];
this.outputLayersNodeIndices = [];
this.outputLayersTensorIndices = [];
/*
All layers in order of horizontal graph traversal. Entries are unique.
Includes input and output layers.
*/
this.layers = [];
/*
References to container layers that were constructed internally. We need
these to properly dispose of tensors from nested containers.
*/
this.internalContainerRefs = [];
// TODO(michaelterry): Determine if caching still needed with eager
// backend.
/*
This is for performance optimization when calling the Container on new
inputs. Every time the Container is called on a set on input tensors,
we compute the output tensors, output masks and output shapes in one pass,
then cache them here. When one of these outputs is queried later,
we retrieve it from there instead of recomputing it.
*/
// this.outputTensorCache = {};
// this.outputShapeCache = {};
// Build this.outputLayers:
for (const x of this.outputs) {
const layer = x.sourceLayer;
const nodeIndex = x.nodeIndex;
const tensorIndex = x.tensorIndex;
this.outputLayers.push(layer);
this.outputLayersNodeIndices.push(nodeIndex);
this.outputLayersTensorIndices.push(tensorIndex);
}
// TODO(michaelterry): Add output mask cache code.
// Build this.inputLayers:
for (const x of this.inputs) {
const layer = x.sourceLayer;
const nodeIndex = x.nodeIndex;
const tensorIndex = x.tensorIndex;
/*
It's supposed to be an input layer, so only one node
and one tensor output.
*/
assert(nodeIndex === 0, 'input layer has >1 nodes');
assert(tensorIndex === 0, 'input layer has >1 tensors');
this.inputLayers.push(layer);
this.inputLayersNodeIndices.push(nodeIndex);
this.inputLayersTensorIndices.push(tensorIndex);
}
// Build this.inputNames and this.outputNames.
this.inputNames = [];
this.outputNames = [];
this.feedInputShapes = [];
this.feedInputNames = [];
this.feedOutputNames = [];
for (let i = 0; i < this.inputLayers.length; i++) {
const layer = this.inputLayers[i];
// Check that layer is an InputLayer.
if (!(layer instanceof InputLayer)) {
throw new TypeError('Input layers to a LayersModel must be InputLayer objects. ' +
`Received inputs: ${args.inputs}. ` +
`Input ${i} (0-based) originates ` +
`from layer type ${layer.getClassName()}.`);
}
this.inputNames.push(layer.name);
this.feedInputShapes.push(layer.batchInputShape);
this.feedInputNames.push(layer.name);
}
for (const layer of this.outputLayers) {
this.outputNames.push(layer.name);
}
this.internalInputShapes = this.inputs.map(x => x.shape);
this.internalOutputShapes = this.outputs.map(x => x.shape);
/*
Container_nodes: set of nodes included in the graph (not all nodes
included in the layers are relevant to the current graph).
*/
// ids of all nodes relevant to the Container:
const nodesDepths = {};
// To recover nodes from their ID.
const nodeIDToNode = {};
const layersDepths = {};
// To layers from their ID.
const layerIDToLayer = {};
const layerIndices = {};
const nodesInDecreasingDepth = [];
/**
* Builds a map of the graph of layers.
*
* This recursively updates the map `layerIndices`,
* the list `nodesInDecreasingDepth` and the set `containerNodes`.
*
* @param tensor Some tensor in a graph.
* @param finishedNodes Set of nodes whose subgraphs have been traversed
* completely. Useful to prevent duplicated work.
* @param nodesInProgress Set of nodes that are currently active on the
* recursion stack. Useful to detect cycles.
* @param layer Layer from which `tensor` comes from. If not provided,
* will be obtained from tensor.sourceLayer.
* @param nodeIndex Node index from which `tensor` comes from.
* @param tensorIndex TensorIndex from which `tensor` comes from.
*
* @exception RuntimeError if a cycle is detected.
*/
const buildMapOfGraph = (tensor, finishedNodes, nodesInProgress, layer, nodeIndex, tensorIndex) => {
if (layer == null || nodeIndex == null || tensorIndex == null) {
layer = tensor.sourceLayer;
nodeIndex = tensor.nodeIndex;
tensorIndex = tensor.tensorIndex;
}
const node = layer.inboundNodes[nodeIndex];
// Prevent cycles.
if (nodesInProgress.indexOf(node) !== -1) {
throw new RuntimeError(`The tensor ${tensor.name} at layer "${layer.name}" ` +
'is part of a cycle.');
}
// Don't repeat work for shared subgraphs
if (finishedNodes.indexOf(node) !== -1) {
return;
}
// Update containerNodes.
this.containerNodes.add(Container.nodeKey(layer, nodeIndex));
// Store the traversal order for layer sorting.
if (!(layer.id in layerIndices)) {
layerIndices[layer.id] = Object.keys(layerIndices).length;
}
if (nodesInProgress.indexOf(node) === -1) {
nodesInProgress.push(node);
}
// Propagate to all previous tensors connected to this node.
const numInboundLayers = node.inboundLayers.length;
for (let i = 0; i < numInboundLayers; i++) {
const x = node.inputTensors[i];
const layer = node.inboundLayers[i];
const nodeIndex = node.nodeIndices[i];
const tensorIndex = node.tensorIndices[i];
buildMapOfGraph(x, finishedNodes, nodesInProgress, layer, nodeIndex, tensorIndex);
}
finishedNodes.push(node);
while (nodesInProgress.indexOf(node) >= 0) {
nodesInProgress.splice(nodesInProgress.indexOf(node), 1);
}
nodesInDecreasingDepth.push(node);
};
const finishedNodes = [];
const nodesInProgress = [];
for (const x of this.outputs) {
buildMapOfGraph(x, finishedNodes, nodesInProgress);
}
const reversedNodesInDecreasingDepth = nodesInDecreasingDepth.slice().reverse();
for (const node of reversedNodesInDecreasingDepth) {
nodeIDToNode[node.id] = node;
// If the depth is not set, the node has no outbound nodes (depth 0).
if (!(node.id in nodesDepths)) {
nodesDepths[node.id] = 0;
}
let depth = nodesDepths[node.id];
// Update the depth of the corresponding layer
const previousDepth = (layersDepths[node.outboundLayer.id] == null ?
0 :
layersDepths[node.outboundLayer.id]);
/*
If we've seen this layer before at a higher depth, we should use that
depth instead of the node depth. This is necessary for shared layers
that have inputs at different depth levels in the graph.
*/
depth = Math.max(depth, previousDepth);
layersDepths[node.outboundLayer.id] = depth;
layerIDToLayer[node.outboundLayer.id] = node.outboundLayer;
nodesDepths[node.id] = depth;
// Update the depth of inbound nodes.
for (let i = 0; i < node.inboundLayers.length; i++) {
const inboundLayer = node.inboundLayers[i];
const nodeIndex = node.nodeIndices[i];
const inboundNode = inboundLayer.inboundNodes[nodeIndex];
const previousDepth = (nodesDepths[inboundNode.id] == null ? 0 :
nodesDepths[inboundNode.id]);
nodesDepths[inboundNode.id] = Math.max(depth + 1, previousDepth);
nodeIDToNode[inboundNode.id] = inboundNode;
}
}
// Build a dict {depth: list of nodes with this depth}
const nodesByDepth = {};
for (const nodeID in nodesDepths) {
const depth = nodesDepths[nodeID];
if (!(depth in nodesByDepth)) {
nodesByDepth[depth] = [];
}
nodesByDepth[depth].push(nodeIDToNode[nodeID]);
}
// Build a dict {depth: list of layers with this depth}
const layersByDepth = {};
for (const layerID in layersDepths) {
const depth = layersDepths[layerID];
if (!(depth in layersByDepth)) {
layersByDepth[depth] = [];
}
layersByDepth[depth].push(layerIDToLayer[layerID]);
}
// Get sorted list of layer depths.
let depthKeys = Object.keys(layersByDepth)
.map(x => parseInt(x, 10))
.sort(reverseNumberCompare);
// Set this.layers and this.layersByDepth.
this.layers = [];
for (const depth of depthKeys) {
const layersForDepth = layersByDepth[depth];
// Container.layers needs to have a deterministic order:
// here we order them by traversal order.
layersForDepth.sort((a, b) => {
const aIndex = layerIndices[a.id];
const bIndex = layerIndices[b.id];
if (aIndex < bIndex) {
return -1;
}
if (aIndex > bIndex) {
return 1;
}
return 0;
});
for (const layer of layersForDepth) {
if (layer instanceof Container) {
this.internalContainerRefs.push(layer);
}
this.layers.push(layer);
}
}
this.layersByDepth = layersByDepth;
// Get sorted list of node depths;
depthKeys = Object.keys(nodesByDepth)
.map(x => parseInt(x, 10))
.sort(reverseNumberCompare);
// Check that all tensors required are computable.
// computable_tensors: all tensors in the graph
// that can be computed from the inputs provided.
const computableTensors = this.inputs.slice();
// To provide a better error msg.
const layersWithCompleteInput = [];
for (const depth of depthKeys) {
for (const node of nodesByDepth[depth]) {
const layer = node.outboundLayer;
if (layer != null) {
for (const x of node.inputTensors) {
if (computableTensors.indexOf(x) === -1) {
throw new RuntimeError(`Graph disconnected: cannot obtain value for tensor ${x}` +
` at layer "${layer.name}". ` +
'The following previous layers were accessed without ' +
`issue: ${layersWithCompleteInput}`);
}
}
for (const x of node.outputTensors) {
computableTensors.push(x);
}
layersWithCompleteInput.push(layer.name);
}
}
}
// Set this.containerNodes and this.nodesByDepth.
this.nodesByDepth = nodesByDepth;
// Ensure name unicity, which will be crucial for serialization
// (since serialized nodes refer to layers by their name).
const allNames = this.layers.map(x => x.name);
for (const name of allNames) {
const numOccurrences = allNames.filter(x => x === name).length;
if (numOccurrences !== 1) {
throw new RuntimeError(`The name "${name}" is used ${numOccurrences} times ` +
'in the model. All layer names should be unique. Layer names: ' +
JSON.stringify(allNames));
}
}
// Layer parameters.
// The new container starts with a single inbound node
// for its inputs, and no outbound nodes.
// Will be appended to by future calls to apply().
this.outboundNodes = [];
// Will be appended to below, and by future calls to apply().
this.inboundNodes = [];
// Create the node linking internal inputs to internal outputs.
// (This call has side effects.)
// tslint:disable-next-line:no-unused-expression
new Node({
outboundLayer: this,
inboundLayers: [],
nodeIndices: [],
tensorIndices: [],
inputTensors: this.inputs,
outputTensors: this.outputs,
inputMasks: this.inputs.map(x => null),
outputMasks: this.outputs.map(x => null),
inputShapes: this.inputs.map(x => x.shape),
outputShapes: this.outputs.map(x => x.shape)
});
this.built = true;
this._refCount = 1; // The ref count of a container always start at 1.
}
assertNotDisposed() {
if (this._refCount === 0) {
throw new Error(`Container '${this.name}' is already disposed.`);
}
}
/**
* Attempt to dispose a LayersModel's weights.
*
* This method decrease the reference count of the LayersModel object by 1.
*
* A LayersModel is reference-counted. Its reference count is incremented by 1
* when it is first constructed and when it is used as a Layer of another
* LayersModel.
*
* If the reference count of a LayersModel becomes 0, the `dispose` method of
* all its constituent `Layer`s will be called.
*
* Note: If the reference count is greater than 0 after the decrement, the
* `dispose` method of its constituent `Layer`s will *not* be called.
*
* After a LayersModel is disposed, it cannot be used in calls such as
* 'predict`, `evaluate` or `fit` anymore.
*
* @returns A DisposeResult Object with the following fields:
* - refCountAfterDispose: The reference count of the LayersModel after this
* `dispose()` call.
* - numDisposedVariables: Number of `tf.Variable`s (i.e., weights) disposed
* during this `dispose()` call.
* @throws {Error} If the layer is not built yet, or if the LayersModel has
* already been disposed.
*/
dispose() {
this.assertNotDisposed();
const result = { refCountAfterDispose: null, numDisposedVariables: 0 };
if (--this._refCount === 0) {
for (const layer of this.layers) {
result.numDisposedVariables += layer.dispose().numDisposedVariables;
}
// Call dispose on each internally created container layer again to ensure
// their refCounts hit zero and their tensors are subsequently deleted.
for (const container of this.internalContainerRefs) {
result.numDisposedVariables += container.dispose().numDisposedVariables;
}
}
result.refCountAfterDispose = this._refCount;
return result;
}
get trainable() {
return this.trainable_;
}
set trainable(trainable) {
this.layers.forEach(layer => {
// tslint:disable-next-line:no-any
layer._trainableWeights
.forEach(w => w.trainable = trainable);
});
this.trainable_ = trainable;
}
get trainableWeights() {
// Porting Note: This check below is to prevent errors where the
// _trainableWeights inherited from the parent class (Layer) gets
// inadvertently used.
if (this._trainableWeights.length > 0) {
throw new ValueError('Container instance unexpectedly contains _trainableWeights.' +
'The trainable weights of a Container are a union of the ' +
'trainable weights of its consituent Layers. Its own ' +
'_trainableWeights must remain an empty Array.');
}
if (!this.trainable) {
return [];
}
let weights = [];
for (const layer of this.layers) {
weights = weights.concat(layer.trainableWeights);
}
return weights;
}
get nonTrainableWeights() {
const weights = [];
for (const layer of this.layers) {
weights.push(...layer.nonTrainableWeights);
}
if (!this.trainable) {
const trainableWeights = [];
for (const layer of this.layers) {
trainableWeights.push(...layer.trainableWeights);
}
return trainableWeights.concat(weights);
}
return weights;
}
get weights() {
return this.trainableWeights.concat(this.nonTrainableWeights);
}
/**
* Loads all layer weights from a JSON object.
*
* Porting Note: HDF5 weight files cannot be directly loaded in JavaScript /
* TypeScript. The utility script at `scripts/pykeras.py` offers means
* to convert them into JSON strings compatible with this method.
* Porting Note: TensorFlow.js Layers supports only loading by name currently.
*
* @param weights A JSON mapping weight names to weight values as nested
* arrays of numbers, or a `NamedTensorMap`, i.e., a JSON mapping weight
* names to `tf.Tensor` objects.
* @param strict Require that the provided weights exactly match those
* required by the container. Default: `true`. Passing `false` means that
* extra weights and missing weights will be silently ignored.
*/
loadWeights(weights, strict = true) {
const nameToWeight = {};
let totalWeightsCount = 0;
const modelIsKerasSavedModelFormat = isKerasSavedModelFormat(weights);
if (modelIsKerasSavedModelFormat) {
this.parseWeights(weights);
}
// Check if weights from keras v3.
for (const layer of this.layers) {
for (const [index, weight] of layer.weights.entries()) {
// Parse the name to layerName/index.
// e.g. dense/0, dense/1, dense_1/0, dense_1/1
const parsedName = modelIsKerasSavedModelFormat ?
`${weight.name.split('/').slice(0, -1).join('/') + '/'}${index}` :
weight.originalName;
if (nameToWeight[parsedName] != null) {
throw new ValueError(`Duplicate weight name: ${parsedName}`);
}
nameToWeight[parsedName] = weight;
totalWeightsCount++;
}
}
const weightValueTuples = [];
for (const name in weights) {
// TF 2.2.0 added cell name to the weight name in the format of
// layer_name/cell_name/weight_name, we need to remove
// the inner cell name.
let validatedName = name;
if (nameToWeight[name] == null) {
const tokens = name.split('/');
const shortenNameArray = tokens.slice(0, -2).concat([tokens[tokens.length - 1]]);
validatedName = shortenNameArray.join('/');
}
if (nameToWeight[validatedName] != null) {
weightValueTuples.push([nameToWeight[validatedName], weights[name]]);
}
else if (strict) {
throw new ValueError(`Provided weight data has no target variable: ${name}`);
}
delete nameToWeight[validatedName];
}
if (strict) {
// Check that all weights are set.
const unsetNames = [];
for (const name in nameToWeight) {
unsetNames.push(name);
}
if (unsetNames.length > 0) {
throw new ValueError(`${unsetNames.length} of ${totalWeightsCount} weights are not set: ` +
`${unsetNames}`);
}
}
batchSetValue(weightValueTuples);
}
parseWeights(weights) {
for (const key in Object.keys(weights)) {
const listParts = key.split('/');
const list = ['vars', 'layer_checkpoint_dependencies'];
// For keras v3, the weights name are saved based on the folder structure.
// e.g. _backbone/_layer_checkpoint_dependencies/transformer/_self../
// _output_dense/vars/0
// Therefore we discard the `vars` and `layer_checkpoint_depencies` within
// the saved name and only keeps the layer name and weights.
// This can help to mapping the actual name of the layers and load each
// weight accordingly.
const newKey = listParts
.map(str => {
if (str.startsWith('_')) {
return str.slice(1);
}
return str;
})
.filter(str => !list.includes(str))
.join('/');
if (newKey !== key) {
weights[newKey] = weights[key];
delete weights[key];
}
}
}
/**
* Util shared between different serialization methods.
* @returns LayersModel config with Keras version information added.
*/
updatedConfig() {
const theConfig = this.getConfig();
const modelConfig = {};
modelConfig['className'] = this.getClassName();
modelConfig['config'] = theConfig;
modelConfig['kerasVersion'] = `tfjs-layers ${version}`;
// TODO(nielsene): Replace something like K.backend() once
// possible.
modelConfig['backend'] = 'TensorFlow.js';
return modelConfig;
}
/**
* Returns a JSON string containing the network configuration.
*
* To load a network from a JSON save file, use
* models.modelFromJSON(jsonString);
* @param extraJsonArgs Unused in tfjs-layers, maintained for PyKeras
* @param returnString Whether the return value should be stringified
* (default: `true`).
* @returns a JSON string if `returnString` (default), or a JSON object if
* `!returnString`.
*/
// tslint:disable-next-line:no-any
toJSON(unused, returnString = true) {
const modelConfig = convertTsToPythonic(this.updatedConfig());
return returnString ? JSON.stringify(modelConfig) : modelConfig;
}
/**
* Call the model on new inputs.
*
* In this case `call` just reapplies all ops in the graph to the new inputs
* (e.g. build a new computational graph from the provided inputs).
*
* @param inputs A tensor or list of tensors.
* @param mask A mask or list of masks. A mask can be either a tensor or null
* (no mask).
*
* @return A tensor if there is a single output, or a list of tensors if there
* are more than one outputs.
*/
call(inputs, kwargs) {
return tidy(() => {
inputs = toList(inputs);
const feedDict = new FeedDict();
for (let i = 0; i < this.inputs.length; ++i) {
feedDict.add(this.inputs[i], inputs[i]);
}
return execute(this.outputs, feedDict, kwargs);
});
}
/**
* Computes an output mask tensor.
*
* @param inputs Tensor or list of tensors.
* @param mask Tensor or list of tensors.
*
* @return null or a tensor (or list of tensors, one per output tensor of the
* layer).
*/
computeMask(inputs, mask) {
return tidy(() => {
inputs = toList(inputs);
let masks;
if (mask == null) {
masks = pyListRepeat(null, inputs.length);
}
else {
masks = toList(mask);
}
// TODO(michaelterry): Add support for mask caching.
return this.runInternalGraph(inputs, masks)[1];
});
}
/**
* Computes the output shape of the layer.
*
* Assumes that the layer will be built to match that input shape provided.
*
* @param inputShape A shape (tuple of integers) or a list of shape tuples
* (one per output tensor of the layer). Shape tuples can include null for
* free dimensions, instead of an integer.
*/
computeOutputShape(inputShape) {
const inputShapes = normalizeShapeList(inputShape);
if (inputShapes.length !== this.inputLayers.length) {
throw new ValueError(`Invalid inputShape argument ${inputShape}: ` +
`model has ${this.inputLayers.length} tensor inputs.`);
}
// TODO(michaelterry): Add caching
const layersToOutputShapes = {};
for (let i = 0; i < inputShapes.length; i++) {
const layer = this.inputLayers[i];
const inputShape = inputShapes[i];
// It's an input layer: computeOutputShape is identity,
// and there is only one node and one tensor output.
const shapeKey = layer.name + '_0_0';
layersToOutputShapes[shapeKey] = inputShape;
}
const depthKeys = Object.keys(this.nodesByDepth)
.map(x => parseInt(x, 10))
.sort(reverseNumberCompare);
// Iterate over nodes, by depth level.
if (depthKeys.length > 1) {
for (const depth of depthKeys) {
const nodes = this.nodesByDepth[depth];
for (const node of nodes) {
// This is always a single layer, never a list.
const layer = node.outboundLayer;
if (this.inputLayers.map(x => x.id).indexOf(layer.id) !== -1) {
// We've already covered the input layers a few lines above.
continue;
}
// Potentially redundant list, same size of node.inputTensors.
const inputShapes = [];
for (let j = 0; j < node.inboundLayers.length; j++) {
const inboundLayer = node.inboundLayers[j];
const nodeIndex = node.nodeIndices[j];
const tensorIndex = node.tensorIndices[j];
const shapeKey = `${inboundLayer.name}_${nodeIndex}_${tensorIndex}`;
const inputShape = layersToOutputShapes[shapeKey];
inputShapes.push(inputShape);
}
const outputShape = layer.computeOutputShape(singletonOrArray(inputShapes));
const outputShapes = normalizeShapeList(outputShape);
const nodeIndex = layer.inboundNodes.indexOf(node);
for (let j = 0; j < outputShapes.length; j++) {
const shapeKey = `${layer.name}_${nodeIndex}_${j}`;
layersToOutputShapes[shapeKey] = outputShapes[j];
}
}
}
}
// Read final output shapes from layersToOutputShapes.
const outputShapes = [];
const outputShapeKeys = [];
for (let i = 0; i < this.outputLayers.length; i++) {
const layer = this.outputLayers[i];
const nodeIndex = this.outputLayersNodeIndices[i];
const tensorIndex = this.outputLayersTensorIndices[i];
const shapeKey = `${layer.name}_${nodeIndex}_${tensorIndex}`;
outputShapeKeys.push(shapeKey);
}
for (let i = 0; i < outputShapeKeys.length; i++) {
const key = outputShapeKeys[i];
assert(key in layersToOutputShapes);
outputShapes.push(layersToOutputShapes[key]);
}
// TODO(michaelterry): Update cache
return singletonOrArray(outputShapes);
}
/**
* Computes output tensors for new inputs.
*
* Note:
* - Expects `inputs` to be a list (potentially with 1 element).
*
* @param inputs List of tensors
* @param masks List of masks (tensors or null).
* @return Three lists: outputTensors, outputMasks, outputShapes
*/
runInternalGraph(inputs, masks) {
if (masks == null) {
masks = pyListRepeat(null, inputs.length);
}
// Dictionary mapping reference tensors to tuples
// (computed tensor, compute mask)
// we assume a 1:1 mapping from tensor to mask
// TODO: raise exception when a `.computeMask()` call
// does not return a list the same size as `call`
const tensorMap = {};
for (let i = 0; i < this.inputs.length; ++i) {
const x = this.inputs[i];
const y = inputs[i];
const mask = masks[i];
tensorMap[x.id] = [y, mask];
}
const depthKeys = Object.keys(this.nodesByDepth)
.map(x => parseInt(x, 10))
.sort(reverseNumberCompare);
for (const depth of depthKeys) {
const nodes = this.nodesByDepth[depth];
for (const node of nodes) {
// This is always a single layer, never a list.
const layer = node.outboundLayer;
const referenceInputTensors = node.inputTensors;
const referenceOutputTensors = node.outputTensors;
// If all previous input tensors are available in tensorMap,
// then call node.inboundLayer on them.
// List of tuples [input, mask]:
const computedData = new Array();
for (const x of referenceInputTensors) {
if (x.id in tensorMap) {
computedData.push(tensorMap[x.id]);
}
}
if (computedData.length === referenceInputTensors.length) {
// TODO(michaelterry): Add K.name_scope here, if we need it.
let kwargs = {};
let computedTensors;
let computedMasks;
let outputTensors;
let outputMasks;
// call layer
if (node.callArgs != null) {
kwargs = node.callArgs;
}
if (computedData.length === 1) {
const [computedTensor, computedMask] = computedData[0];
if (kwargs['mask'] == null) {
kwargs['mask'] = computedMask;
}
outputTensors =
toList(layer.call(computedTensor, kwargs));
outputMasks = toList(layer.computeMask(computedTensor, computedMask));
computedTensors = [computedTensor];
computedMasks = [computedMask];
}
else {
computedTensors = computedData.map(x => x[0]);
computedMasks = computedData.map(x => x[1]);
if (kwargs['mask'] == null) {
kwargs['mask'] = computedMasks;
}
outputTensors =
toList(layer.call(computedTensors, kwargs));
outputMasks = toList(layer.computeMask(computedTensors, computedMasks));
}
if (layer.activityRegularizer) {
throw new NotImplementedError('LayersModel invocation with concrete Tensor value(s) in the ' +
'presence of activity regularizer(s) is not supported yet.');
}
// TODO(michaelterry): Add model updates and losses
// Update tensor map.
for (let i = 0; i < referenceOutputTensors.length; ++i) {
const x = referenceOutputTensors[i];
const y = outputTensors[i];
const mask = outputMasks[i];
tensorMap[x.id] = [y, mask];
}
}
}
}
const outputTensors = [];
const outputMasks = [];
const outputShapes = [];
for (const x of this.outputs) {
assert(x.id in tensorMap, `Could not compute output ${x.name} : ${x.id}`);
const [tensor, mask] = tensorMap[x.id];
outputShapes.push(tensor.shape);
outputTensors.push(tensor);
outputMasks.push(mask);
}
// TODO(michaelterry): Add support for caches.
return [outputTensors, outputMasks, outputShapes];
}
/**
* Builds a map of internal node keys to node ordering.
* Used in serializaion a node orderings may change as unused nodes are
* dropped. Porting Note: This helper method was pulled out of getConfig to
* improve readability.
* @param layers An array of Layers in the model.
* @returns Map of Node Keys to index order within the layer.
*/
buildNodeConversionMap(layers) {
const nodeConversionMap = {};
let keptNodes;
for (const layer of this.layers) {
keptNodes = layer instanceof Container ? 1 : 0;
for (let originalNodeIndex = 0; originalNodeIndex < layer.inboundNodes.length; originalNodeIndex++) {
const nodeKey = Container.nodeKey(layer, originalNodeIndex);
if (this.containerNodes.has(nodeKey)) {
// i.e. we mark it to be saved
nodeConversionMap[nodeKey] = keptNodes;
keptNodes += 1;
}
}
}
return nodeConversionMap;
}
getLayer(nameOrIndex, index) {
if (index != null) {
return this.findLayer(index);
}
else {
if (nameOrIndex == null) {
throw new ValueError('Provide either a layer name or layer index');
}
if (typeof nameOrIndex === 'number') {
return this.findLayer(nameOrIndex);
}
}
for (const layer of this.layers) {
if (layer.name === nameOrIndex) {
return layer;
}
}
throw new ValueError(`No such layer: ${nameOrIndex}`);
}
findLayer(index) {
if (this.layers.length <= index) {
throw new ValueError(`Was asked to retrieve layer at index ${index}, but model only ` +
`has ${this.layers.length} layer(s).`);
}
else {
return this.layers[index];
}
}
/**
* Retrieves the Container's current loss values.
*
* Used for regularizers during training.
*/
calculateLosses() {
// Porting Node: This is an augmentation to Container.loss in PyKeras.
// In PyKeras, Container.loss returns symbolic tensors. Here a concrete
// Tensor (specifically Scalar) values are returned. This is due to the
// imperative backend.
return tidy(() => {
const losses = [];
for (const layer of this.layers) {
for (let nodeIndex = 0; nodeIndex < layer.inboundNodes.length; ++nodeIndex) {
const nodeKey = Container.nodeKey(layer, nodeIndex);
if (this.containerNodes.has(nodeKey)) {
losses.push(...layer.calculateLosses());
}
}
}
// TODO(cais): Add any unconditional model-level losses?
return losses;
});
}
getConfig() {
const config = { name: this.name };
// Build a map from layer unique name (self._node_key)
// to the index of the nodes that are saved in the config.
// Only nodes in container_nodes are saved.
const nodeConversionMap = this.buildNodeConversionMap(this.layers);
// Serialize and save the layers in layerConfigs
const layerConfigs = [];
for (const layer of this.layers) {
const layerClassName = layer.getClassName();
const layerConfig = layer.getConfig();
const filteredInboundNodes = [];
for (let originalNodeIndex = 0; originalNodeIndex < layer.inboundNodes.length; originalNodeIndex++) {
const node = layer.inboundNodes[originalNodeIndex];
const nodeKey = Container.nodeKey(layer, originalNodeIndex);
let kwargs = {};
if (this.containerNodes.has(nodeKey)) {
// The node is relevant to the model:
// add to filteredInboundNodes.
if (node.callArgs) {
try {
JSON.stringify(node.callArgs);
kwargs = node.callArgs;
}
catch (err) {
console.warn(`Layer ${layer.name} was passed ` +
`non-serializable keyword arguments: ` +
`${node.callArgs}. They will not be included ` +
`in the serialized model (and thus will be ` +
`missing at deserialization time).`);
kwargs = {};
}
}
if (node.inboundLayers.length > 0) {
const nodeData = [];
for (let i = 0; i < node.inboundLayers.length; i++) {
const inboundLayer = node.inboundLayers[i];
const nodeIndex = node.nodeIndices[i];
const tensorIndex = node.tensorIndices[i];
const nodeKey = Container.nodeKey(inboundLayer, nodeIndex);
let newNodeIndex = nodeConversionMap[nodeKey];
if (newNodeIndex == null) {
newNodeIndex = 0;
}
nodeData.push([inboundLayer.name, newNodeIndex, tensorIndex, kwargs]);
}
filteredInboundNodes.push(nodeData);
}
}
}
const dict = {};
dict['name'] = layer.name;
dict['className'] = layerClassName;
dict['config'] = layerConfig;
dict['inboundNodes'] = filteredInboundNodes;
layerConfigs.push(dict);
}
config['layers'] = layerConfigs;
// Gather info about inputs and outputs
const modelInputs = [];
for (let i = 0; i < this.inputLayers.length; i++) {
const layer = this.inputLayers[i];
const nodeIndex = this.inputLayersNodeIndices[i];
const nodeKey = Container.nodeKey(layer, nodeIndex);
if (!this.containerNodes.has(nodeKey)) {
continue;
}
let newNodeIndex = nodeConversionMap[nodeKey];
if (newNodeIndex === null || newNodeIndex === undefined) {
newNodeIndex = 0;
}
const tensorIndex = this.inputLayersTensorIndices[i];
modelInputs.push([layer.name, newNodeIndex, tensorIndex]);
}
config['inputLayers'] = modelInputs;
const modelOutputs = [];
for (let i = 0; i < this.outputLayers.length; i++) {
const layer = this.outputLayers[i];
const nodeIndex = this.outputLayersNodeIndices[i];
const nodeKey = Container.nodeKey(layer, nodeIndex);
if (!this.containerNodes.has(nodeKey)) {
continue;
}
let newNodeIndex = nodeConversionMap[nodeKey];
if (newNodeIndex === null || newNodeIndex === undefined) {
newNodeIndex = 0;
}
const tensorIndex = this.outputLayersTensorIndices[i];
modelOutputs.push([layer.name, newNodeIndex, tensorIndex]);
}
config['outputLayers'] = modelOutputs;
return config;
}
/**
* Instantiates a LayersModel from its config (output of `get_config()`).
* @param cls the class to create
* @param config LayersModel config dictionary.
* @param customObjects An optional dictionary of custom objects.
* @param fastWeightInit Optional flag to use fast weight initialization
* during deserialization. This is applicable to cases in which
* the initialization will be immediately overwritten by loaded weight
* values. Default: `false`.
* @returns A LayersModel instance.
* @throws ValueError: In case of improperly formatted config dict.
*/
/** @nocollapse */
static fromConfig(cls, config, customObjects = {}, fastWeightInit = false) {
// Layer instances created during
// the graph reconstruction process
const createdLayers = {};
// Dictionary mapping layer instances to
// node data that specifies a layer call.
// It acts as a queue that maintains any unprocessed
// layer call until it becomes possible to process it
// (i.e. until the input tensors to the call all exist).
const unprocessedNodes = {};
function addUnprocessedNode(layer, nodeData) {
if (!(layer.name in unprocessedNodes)) {
unprocessedNodes[layer.name] = [nodeData];
}
else {
unprocessedNodes[layer.name].push(nodeData);
}
}
function processNode(layer, nodeData) {
const inputTensors = [];
let kwargs;
for (const inputData of nodeData) {
const inboundLayerName = inputData[0];
const inboundNodeIndex = inputData[1];
const inboundTensorIndex = inputData[2];
kwargs = inputData[3] == null ?
{} :
inputData[3];
if (!(inboundLayerName in createdLayers)) {
addUnprocessedNode(layer, nodeData);
return;
}
const inboundLayer = createdLayers[inboundLayerName];
if (inboundLayer.inboundNodes.length <= inboundNodeIndex) {
addUnprocessedNode(layer, nodeData);
return;
}
const inboundNode = inboundLayer.inboundNodes[inboundNodeIndex];
inputTensors.push(inboundNode.outputTensors[inboundTensorIndex]);
}
// Call layer on its inputs, thus creating the node
// and building the layer if needed.
// Note: This has Eager vs Graph Implications.
if (inputTensors.length > 0) {
layer.apply(singletonOrArray(inputTensors), kwargs); // was ** kwargs
}
}
/**
* Deserialize a layer, then call it on appropriate inputs.
* @param layerData: layer config dict.
* @throws ValueError: In case of improperly formatted `layer_data`
* dict.
*/
function processLayer(layerData) {
const layerName = layerData['name'];
// Instantiate layer.
const layer = deserialize(layerData, config['customObjects'] != null ?
config['customObjects'] :
{});
layer.setFastWeightInitDuringBuild(fastWeightInit);
createdLayers[layerName] = layer;
// Gather layer inputs.
const inboundNodesData = layerData['inboundNodes'];
inboundNodesData.forEach(nodeData => {
if (!(nodeData instanceof Array)) {
throw new ValueError(`Corrupted configuration, expected array for nodeData: ${nodeData}`);
}
// We don't process nodes (i.e. make layer calls)
// on the fly because the inbound node may not yet exist,
// in case of layer shared at different topological depths
// (e.g.a model such as A(B(A(B(x)))))
addUnprocessedNode(layer, nodeData);
});
}
// First, we create all layers and enqueue nodes to be processed.
const name = config['name'];
const layersFromConfig = config['layers'];
for (const layerData of layersFromConfig) {
processLayer(layerData);
}
// Then we process nodes in order of layer depth.
// Nodes that cannot yet be processed(if the inbound node
// does not yet exist) are re - enqueued, and the process
// is repeated until all nodes are processed.
while (!isObjectEmpty(unprocessedNodes)) {
for (const layerData of layersFromConfig) {
const layer = createdLayers[layerData['name']];
if (layer.name in unprocessedNodes) {
const currentUnprocessedNodesForLayer = unprocessedNodes[layer.name];
delete unprocessedNodes[layer.name];
for (const nodeData of currentUnprocessedNodesForLayer) {
processNode(layer, nodeData);
}
}
}
}
const inputTensors = [];
const outputTensors = [];
const inputLayersFromConfig = config['inputLayers'];
for (const layerData of inputLayersFromConfig) {
const layerName = layerData[0];
const nodeIndex = layerData[1];
const tensorIndex = layerData[2];
assert(layerName in createdLayers);
const layer = createdLayers[layerName];
const layerOutputTensors = layer.inboundNodes[nodeIndex].outputTensors;
inputTensors.push(layerOutputTensors[tensorIndex]);
}
const outputLayersFromConfig = config['outputLayers'];
for (const layerData of outputLayersFromConfig) {
const layerName = layerData[0];
const nodeIndex = layerData[1];
const tensorIndex = layerData[2];
assert(layerName in createdLayers);
const layer = createdLayers[layerName];
const layerOutputTensors = layer.inboundNodes[nodeIndex].outputTensors;
outputTensors.push(layerOutputTensors[tensorIndex]);
}
return new cls({ inputs: inputTensors, outputs: outputTensors, name });
}
/**
* Determine whether the container is stateful.
*
* Porting Note: this is the equivalent of the stateful @property of
* the Container class in PyKeras.
*/
get stateful() {
// Porting Note: This check is to prevent inadvertent setting of the
// _stateful property of the Container instance.
if (this._stateful) {
throw new ValueError('Container instance unexpectedly has _stateful = true. The ' +
'statefulness of a Container is determined by the Layers it ' +
'contains. Its _stateful property must remain the default false.');
}
for (const layer of this.layers) {
if (layer.stateful) {
return true;
}
}
return false;
}
/**
* Reset the state of all stateful constituent layers (if any).
*
* Examples of stateful layers include RNN layers whose `stateful` property
* is set as `true`.
*/
resetStates() {
tidy(() => {
this.layers.forEach(layer => {
// tslint:disable:no-any
if (layer.stateful) {
layer.resetStates();
}
// tslint:enable:no-any
});
});
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
function standardizeSampleOrClassWeights(xWeight, outputNames, weightType) {
const numOutputs = outputNames.length;
if (xWeight == null || (Array.isArray(xWeight) && xWeight.length === 0)) {
return outputNames.map(name => null);
}
if (numOutputs === 1) {
if (Array.isArray(xWeight) && xWeight.length === 1) {
return xWeight;
}
else if (typeof xWeight === 'object' && outputNames[0] in xWeight) {
return [xWeight[outputNames[0]]];
}
else {
return [xWeight];
}
}
if (Array.isArray(xWeight)) {
if (xWeight.length !== numOutputs) {
throw new Error(`Provided ${weightType} is an array of ${xWeight.length} ` +
`element(s), but the model has ${numOutputs} outputs. ` +
`Make sure a set of weights is provided for each model output.`);
}
return xWeight;
}
else if (typeof xWeight === 'object' && Object.keys(xWeight).length > 0 &&
typeof xWeight[Object.keys(xWeight)[0]] ===
'object') {
const output = [];
outputNames.forEach(outputName => {
if (outputName in xWeight) {
output.push(xWeight[outputName]);
}
else {
output.push(null);
}
});
return output;
}
else {
throw new Error(`The model has multiple (${numOutputs}) outputs, ` +
`so ${weightType} must be either an array with ` +
`${numOutputs} elements or an object with ${outputNames} keys. ` +
`Provided ${weightType} not understood: ${JSON.stringify(xWeight)}`);
}
}
/**
* Standardize class weighting objects.
*
* This function takes a single class-weighting object, an array of them,
* or a map from output name to class-weighting object. It compares it to the
* output name(s) of the model, base on which it outputs an array of
* class-weighting objects of which the length matches the number of outputs.
*
* @param classWeight Input class-weighting object(s).
* @param outputNames All output name(s) of the model.
* @return An array of class-weighting objects. The length of the array matches
* the model's number of outputs.
*/
function standardizeClassWeights(classWeight, outputNames) {
return standardizeSampleOrClassWeights(classWeight, outputNames, 'classWeight');
}
/**
* Standardize by-sample and/or by-class weights for training.
*
* Note that this function operates on one model output at a time. For a model
* with multiple outputs, you must call this function multiple times.
*
* @param y The target tensor that the by-sample and/or by-class weight is for.
* The values of y are assumed to encode the classes, either directly
* as an integer index, or as one-hot encoding.
* @param sampleWeight By-sample weights.
* @param classWeight By-class weights: an object mapping class indices
* (integers) to a weight (float) to apply to the model's loss for the
* samples from this class during training. This can be useful to tell the
* model to "pay more attention" to samples from an under-represented class.
* @param sampleWeightMode The mode for the sample weights.
* @return A Promise of weight tensor, of which the size of the first dimension
* matches that of `y`.
*/
async function standardizeWeights(y, sampleWeight, classWeight, sampleWeightMode) {
if (classWeight != null) {
// Apply class weights per sample.
const yClasses = tidy(() => {
if (y.shape.length === 1) {
// Assume class indices.
return clone(y);
}
else if (y.shape.length === 2) {
if (y.shape[1] > 1) {
// Assume one-hot encoding of classes.
const axis = 1;
return argMax$2(y, axis);
}
else if (y.shape[1] === 1) {
// Class index.
return reshape$2(y, [y.shape[0]]);
}
else {
throw new Error(`Encountered unexpected last-dimension size (${y.shape[1]}) ` +
`during handling of class weights. The size is expected to be ` +
`>= 1.`);
}
}
else {
throw new Error(`Unexpected rank of target (y) tensor (${y.rank}) during ` +
`handling of class weights. The rank is expected to be 1 or 2.`);
}
});
const yClassIndices = Array.from(await yClasses.data());
dispose(yClasses);
const classSampleWeight = [];
yClassIndices.forEach(classIndex => {
if (classWeight[classIndex] == null) {
throw new Error(`classWeight must contain all classes in the training data. ` +
`The class ${classIndex} exists in the data but not in ` +
`classWeight`);
}
else {
classSampleWeight.push(classWeight[classIndex]);
}
});
return tensor1d(classSampleWeight, 'float32');
}
else {
return null;
}
}
/**
* Apply per-sample weights on the loss values from a number of samples.
*
* @param losses Loss tensor of shape `[batchSize]`.
* @param sampleWeights Per-sample weight tensor of shape `[batchSize]`.
* @returns Tensor of the same shape as`losses`.
*/
function computeWeightedLoss(losses, sampleWeights) {
return mul(losses, sampleWeights);
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Interfaces and methods for training models using TensorFlow.js datasets.
*/
// Default batch size used during tensor-based validation.
const DEFAULT_VALIDATION_BATCH_SIZE = 32;
/**
* Standardize the output of a dataset iterator for use by
* LayersModel.fitDataset().
*
* @param model: A `tf.LayersModel` object.
* @param iteratorOut The output of a dataset iterator. It is required to be
* an object of the form `{xs: TensorOrArrayOrMap, ys:
* TensorOrArrayOrMap}`, where `TensorOrArrayOrMap` is a single `tf.Tensor`,
* a `tf.Tensor[]`, or a flat map from string names to `tf.Tensor`s.
* @returns A flat array of `tf.Tensor` objects: the input `tf.Tensor`s
* followed by the target `tf.Tensor`s. When `tf.Tensor`s are provided
* as a map, the order in the resulting array is taken from the `inputNames`
* and `outputNames` of the model.
*/
function standardizeDataIteratorOutput(
// Type `model` as `any` here to avoid circular dependency w/
// training.ts.
// tslint:disable-next-line:no-any
model, iteratorOut) {
let xs;
let ys;
const iteratorOutObj = iteratorOut;
xs = iteratorOutObj['xs'];
ys = iteratorOutObj['ys'];
assert$1(xs != null && ys != null, () => 'A Dataset iterator for fitDataset() is expected to generate ' +
'objects of the form `{xs: xVal, ys: yVal}`, where the two ' +
'values may be `tf.Tensor`, an array of Tensors, or a map of ' +
'string to Tensor. The provided Dataset instead generates ' +
`${iteratorOut}`);
const flattenedXs = flattenTensorOrArrayOrMap('input', model.inputNames, xs);
const flattenedYs = flattenTensorOrArrayOrMap('output', model.outputNames, ys);
const batchSize = flattenedXs[0].shape[0];
assert$1(flattenedXs.length === model.inputs.length, () => `LayersModel has ${model.inputs.length} inputs, but the dataset ` +
`provides ${flattenedXs.length} inputs. (Expected input keys: ` +
`${JSON.stringify(model.inputNames)})`);
assert$1(flattenedYs.length === model.outputs.length, () => `LayersModel has ${model.outputs.length} outputs, but the dataset ` +
`provides ${flattenedYs.length} outputs. (Expected output keys: ` +
`${JSON.stringify(model.outputNames)})`);
for (let xIndex = 0; xIndex < flattenedXs.length; xIndex++) {
assert$1(flattenedXs[xIndex].shape[0] === batchSize, () => `Batch size mismatch: input ` +
`${model.inputNames[xIndex]} has ${flattenedXs[xIndex].shape[0]}; ` +
`expected ${batchSize} based on input ${model.inputNames[0]}.`);
}
for (let yIndex = 0; yIndex < flattenedYs.length; yIndex++) {
assert$1(flattenedYs[yIndex].shape[0] === batchSize, () => `Batch size mismatch: output ` +
`${model.outputNames[yIndex]} has ${flattenedYs[yIndex].shape[0]}; ` +
`expected ${batchSize} based on input ${model.inputNames[0]}.`);
}
return { xs: flattenedXs, ys: flattenedYs };
}
function flattenTensorOrArrayOrMap(inputOrOutput, names, values) {
if (values instanceof Tensor) {
return [values];
}
else if (Array.isArray(values)) {
assert$1(values.length === names.length, () => `Received an array of ${values.length} Tensors, but expected ${names.length} to match the ${inputOrOutput} keys ${names}.`);
return values;
}
else {
const result = [];
// Check that all the required keys are available.
for (const name of names) {
if (values[name] == null) {
throw new ValueError(`The feature data generated by the dataset lacks the required ` +
`${inputOrOutput} key '${name}'.`);
}
result.push(values[name]);
}
return result;
}
}
function standardizeTensorValidationData(data) {
if (data.length === 3) {
throw new NotImplementedError('Validation with sample weights is not implemented yet.');
}
return { xs: data[0], ys: data[1] };
}
async function fitDataset(
// Type `model` as `any` here to avoid circular dependency w/
// training.ts.
// tslint:disable-next-line:no-any
model, dataset, args) {
const hasBatchesPerEpoch = args.batchesPerEpoch != null;
assert$1(model.optimizer != null, () => 'You must compile a model before training/testing. Use ' +
'LayersModel.compile(modelCompileConfig).');
assert$1(args != null, () => `For fitDataset(), the 2nd argument (config) is required, ` +
`but it is not provided in this call.`);
assert$1(args.epochs != null && args.epochs > 0 && Number.isInteger(args.epochs), () => `For fitDataset(), config.epochs is expected to be a positive ` +
`integer, but got ${args.epochs}`);
assert$1(!hasBatchesPerEpoch ||
(args.batchesPerEpoch > 0 && Number.isInteger(args.batchesPerEpoch)), () => `For fitDataset(), config.batchesPerEpoch is expected to be a ` +
`positive integer if specified, but got ${args.batchesPerEpoch}`);
assert$1(
// tslint:disable-next-line:no-any
args['validationSplit'] == null, () => '`validationSplit` is not supported by `fitDataset()`. ' +
'Use validationData instead.');
if (model.isTraining) {
throw new Error('Cannot start training because another fit() call is ongoing.');
}
model.isTraining = true;
try {
const doValidation = args.validationData != null;
let valXs;
let valYs;
if (doValidation) {
if (isDatasetObject(args.validationData)) {
assert$1(args.validationBatches == null ||
(args.validationBatches > 0 &&
Number.isInteger(args.validationBatches)), () => `For fitDataset() with dataset-based validation, ` +
`config.validationBatches is expected not to be provided, ` +
`or to be a positive integer, ` +
`but got ${args.validationBatches}`);
}
else {
const validationData = standardizeTensorValidationData(args.validationData);
valXs = validationData.xs;
valYs = validationData.ys;
}
}
const trainFunction = model.makeTrainFunction();
const outLabels = model.getDedupedMetricsNames();
let callbackMetrics;
if (doValidation) {
callbackMetrics =
outLabels.slice().concat(outLabels.map(n => 'val_' + n));
}
else {
callbackMetrics = outLabels.slice();
}
const callbacks = standardizeCallbacks(args.callbacks, args.yieldEvery);
const verbose = args.verbose == null ? 1 : args.verbose;
const { callbackList, history } = configureCallbacks(callbacks, verbose, args.epochs, null, null, getStepsPerEpoch(dataset, args), null, // Batch size determined by the dataset itself.
doValidation, callbackMetrics);
callbackList.setModel(model);
model.history = history;
await callbackList.onTrainBegin();
model.stopTraining_ = false;
let epoch = args.initialEpoch == null ? 0 : args.initialEpoch;
let dataIterator = await dataset.iterator();
while (epoch < args.epochs) {
const epochLogs = {};
await callbackList.onEpochBegin(epoch);
let stepsDone = 0;
let batchIndex = 0;
if (!hasBatchesPerEpoch) {
dataIterator = await dataset.iterator();
}
while (hasBatchesPerEpoch ? stepsDone < args.batchesPerEpoch : true) {
const iteratorOut = await dataIterator.next();
// If `batchesPerEpoch` is specified, the dataset should not be
// exhausted until all epoches are done.
if (hasBatchesPerEpoch && iteratorOut.done) {
console.warn('You provided `batchesPerEpoch` as ' +
`${args.batchesPerEpoch}, ` +
'but your dataset iterator ran out of data after ' +
`${stepsDone} batches; ` +
'interrupting training. Make sure that your ' +
'dataset can generate at least `batchesPerEpoch * epochs` ' +
'batches (in this case, ' +
`${args.batchesPerEpoch * args.epochs} batches). ` +
'You may need to use the repeat() function when building ' +
'your dataset.');
break;
}
if (iteratorOut.value != null) {
const { xs, ys } = standardizeDataIteratorOutput(model, iteratorOut.value);
const batchLogs = {};
batchLogs['batch'] = batchIndex;
batchLogs['size'] = xs[0].shape[0];
await callbackList.onBatchBegin(batchIndex, batchLogs);
const sampleWeights = [];
if (args.classWeight != null) {
const standardClassWeights = standardizeClassWeights(args.classWeight, model.outputNames);
for (let i = 0; i < standardClassWeights.length; ++i) {
sampleWeights.push(await standardizeWeights(ys[i], null, standardClassWeights[i]));
}
}
// Train on batch.
const ins = xs.concat(ys).concat(sampleWeights);
const outs = trainFunction(ins);
dispose(ins);
for (let i = 0; i < outLabels.length; ++i) {
const label = outLabels[i];
const out = outs[i];
batchLogs[label] = out;
keep(out);
}
await callbackList.onBatchEnd(batchIndex, batchLogs);
disposeTensorsInLogs(batchLogs);
batchIndex++;
stepsDone++;
}
if (hasBatchesPerEpoch ? stepsDone >= args.batchesPerEpoch :
iteratorOut.done) {
// Epoch finished. Perform validation.
if (doValidation) {
let valOuts;
if (isDatasetObject(args.validationData)) {
valOuts = toList(await model.evaluateDataset(args.validationData, { batches: args.validationBatches }));
}
else {
valOuts = toList(model.evaluate(valXs, valYs, {
batchSize: args.validationBatchSize == null ?
DEFAULT_VALIDATION_BATCH_SIZE :
args.validationBatchSize,
verbose: 0
}));
}
for (let i = 0; i < model.metricsNames.length; ++i) {
epochLogs[`val_${model.metricsNames[i]}`] = valOuts[i];
}
}
// Call `break` to exit one epoch lopp after validation is done. If
// config.batchesPerEpoch is specified, an epoch while loop will
// stop when `stepsDone >= config.batchesPerEpoch`. When
// config.batchesPerEpoch is not provided, the following `break` is
// required to exit the while lopp after dataset is exhausted.
break;
}
if (model.stopTraining_) {
break;
}
}
await callbackList.onEpochEnd(epoch, epochLogs);
epoch++;
if (model.stopTraining_) {
break;
}
}
await callbackList.onTrainEnd();
await model.history.syncData();
return model.history;
}
finally {
model.isTraining = false;
}
}
/** Helper function that determines number of steps (batches) per epoch. */
function getStepsPerEpoch(dataset, args) {
// Attempt to determine # of batches in an epoch.
let stepsPerEpoch = null;
if (args.batchesPerEpoch != null) {
stepsPerEpoch = args.batchesPerEpoch;
}
else if (Number.isFinite(dataset.size)) {
stepsPerEpoch = dataset.size;
}
return stepsPerEpoch;
}
// Check if provided object is a Dataset object by checking its .iterator
// element.
function isDatasetObject(dataset) {
return (typeof dataset.iterator === 'function');
}
// Check if provided object is a LazyIterator object by checking it's .next
// element.
function isLazyIteratorObject(iterator) {
return (typeof iterator.next === 'function');
}
async function evaluateDataset(
// Type `model` as `any` here to avoid circular dependency w/
// training.ts.
// tslint:disable-next-line:no-any
model, dataset, args) {
args = args || {};
const hasBatches = args.batches != null;
const f = model.testFunction;
let outs = [];
if (args.verbose > 0) {
throw new NotImplementedError('Verbose mode is not implemented yet.');
}
assert$1(!hasBatches || (args.batches > 0 && Number.isInteger(args.batches)), () => 'Test loop expects `batches` to be a positive integer, but ' +
`received ${JSON.stringify(args.batches)}`);
const dataIterator = isLazyIteratorObject(dataset) ?
dataset :
await dataset.iterator();
// Keeps track of number of examples used in this evaluation.
let numExamples = 0;
let batch = 0;
while (hasBatches ? batch < args.batches : true) {
const iteratorOut = await dataIterator.next();
outs = tidy(() => {
if (iteratorOut.value) {
// TODO(cais): Once real dataset is available, use
// `map(x => standardizeDataIteratorOutput(model, x).map(f)`.
const { xs, ys } = standardizeDataIteratorOutput(model, iteratorOut.value);
const xsAndYs = xs.concat(ys);
const batchOuts = tidy(() => f(xsAndYs));
dispose(xsAndYs);
if (batch === 0) {
for (let i = 0; i < batchOuts.length; ++i) {
outs.push(scalar(0));
}
}
const batchSize = xsAndYs[0].shape[0];
for (let i = 0; i < batchOuts.length; ++i) {
const batchOut = batchOuts[i];
const oldScalar = outs[i];
outs[i] =
tidy(() => add$1(outs[i], mul(batchSize, batchOut)));
if (batch > 0) {
dispose(oldScalar);
}
}
dispose(batchOuts);
numExamples += batchSize;
++batch;
}
return outs;
});
if (iteratorOut.done) {
if (hasBatches) {
console.warn('Your dataset iterator ran out of data during evaluateDataset(). ' +
'Interrupting evalution. Make sure that your ' +
'dataset can generate at least `batches` ' +
`batches (in this case, ${args.batches} batches). ` +
'You may need to use the repeat() function when building ' +
'your dataset.');
}
break;
}
}
for (let i = 0; i < outs.length; ++i) {
const oldScalar = outs[i];
outs[i] = div$1(outs[i], numExamples);
dispose(oldScalar);
}
return singletonOrArray(outs);
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Interfaces and methods for training models using tf.Tensor objects.
*/
function checkBatchSize(batchSize) {
assert$1(batchSize > 0 && Number.isInteger(batchSize), () => `batchSize is required to be a positive integer, but got ${batchSize}`);
}
/**
* Slice a Tensor or an Array of Tensors, by start and stop indices.
*
* Porting Note: The `_slice_arrays` function in PyKeras is covered by this
* function and `sliceArraysByIndices()` together.
*
* @param arrays: the input.
* @param start: the starting index (inclusive).
* @param stop: the stopping index (exclusive).
* @returns The result of the slicing. If `arrays` is an `Array` of
* `tf.Tensor`s, the slicing will be applied to all elements of the `Array`
* in the same way.
*/
function sliceArrays(arrays, start, stop) {
if (arrays == null) {
return [null];
}
else if (Array.isArray(arrays)) {
return arrays.map(array => sliceAlongFirstAxis(array, start, stop - start));
}
else { // Tensor.
return sliceAlongFirstAxis(arrays, start, stop - start);
}
}
/**
* Slice a Tensor or an Array of Tensors, by random-order indices.
*
* Porting Note: The `_slice_arrays` function in PyKeras is covered by this
* function and `sliceArrays()` together.
*
* @param arrays The input `tf.Tensor` or `Array` of `tf.Tensor`s to slice.
* If an `Array` of `tf.Tensor`s, all `tf.Tensor`s will be sliced in the
* same fashion.
* @param indices The indices to use for slicing along the first (batch)
* dimension.
* @returns Result(s) of the slicing.
*/
function sliceArraysByIndices(arrays, indices) {
return tidy(() => {
if (arrays == null) {
return null;
}
else if (Array.isArray(arrays)) {
return arrays.map(array => sliceArraysByIndices(array, indices));
}
else {
// TODO(cais): indices should be a pre-constructed Tensor1D to avoid
// tensor1d() calls.
return gather(arrays, indices.dtype === 'int32' ? indices : cast$3(indices, 'int32'));
}
});
}
/**
* Returns a list of batch indices (tuples of indices).
* @param size: Integer, total size of the data to slice into batches.
* @param batchSize: Integer, batch size.
* @returns An Array of [batchStart, batchEnd] tuples. batchStart is
* inclusive; batchEnd is exclusive. I.e., each batch consists of indices x
* that satisfy batchStart <= x < batchEnd.
*/
function makeBatches(size, batchSize) {
const output = [];
let batchStart = 0;
let batchEnd = null;
while (batchStart < size) {
batchEnd = batchStart + batchSize;
if (batchEnd >= size) {
batchEnd = size;
}
output.push([batchStart, batchEnd]);
batchStart = batchEnd;
}
return output;
}
/**
* Ensure tensors all have a rank of at least 2.
*
* If a tensor has a rank of 1, it is dimension-expanded to rank 2.
* If any tensor has a rank of 0 (i.e., is a scalar), an error will be thrown.
*/
function ensureTensorsRank2OrHigher(tensors) {
const outs = [];
if (tensors instanceof Tensor) {
tensors = [tensors];
}
// Make Tensors at least 2D.
for (let i = 0; i < tensors.length; ++i) {
const tensor = tensors[i];
if (tensor.rank === 1) {
outs.push(expandDims(tensor, 1));
}
else if (tensor.rank === 0) {
throw new Error('Expected tensor to be at least 1D, but received a 0D tensor ' +
'(scalar).');
}
else {
outs.push(tensor);
}
}
return outs;
}
/**
* Compare a set of tensors with a reference (old) set, discard the ones
* in the new set that are not present in the reference set.
*
* This method is used for memory clenaup during calls such as
* LayersModel.fit().
*
* @param tensors New set which may contain Tensors not present in
* `refTensors`.
* @param refTensors Reference Tensor set.
*/
// TODO(cais, kangyizhang): Deduplicate with tfjs-data.
function disposeNewTensors(tensors, refTensors) {
if (tensors == null) {
return;
}
const oldTensorIds = [];
if (refTensors instanceof Tensor) {
oldTensorIds.push(refTensors.id);
}
else if (Array.isArray(refTensors)) {
refTensors.forEach(t => oldTensorIds.push(t.id));
}
else if (refTensors != null) {
// `oldTensors` is a map from string name to Tensor.
for (const name in refTensors) {
const oldTensor = refTensors[name];
oldTensorIds.push(oldTensor.id);
}
}
const tensorsToDispose = [];
if (tensors instanceof Tensor) {
if (oldTensorIds.indexOf(tensors.id) === -1) {
tensorsToDispose.push(tensors);
}
}
else if (Array.isArray(tensors)) {
tensors.forEach(t => {
if (oldTensorIds.indexOf(t.id) === -1) {
tensorsToDispose.push(t);
}
});
}
else if (tensors != null) {
// `oldTensors` is a map from string name to Tensor.
for (const name in tensors) {
const tensor = tensors[name];
if (oldTensorIds.indexOf(tensor.id) === -1) {
tensorsToDispose.push(tensor);
}
}
}
tensorsToDispose.forEach(t => {
if (!t.isDisposed) {
t.dispose();
}
});
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original Source: engine/training.py */
/**
* Helper function for polymorphic input data: 1. singleton Tensor.
*/
function isDataTensor(x) {
return x instanceof Tensor;
}
/**
* Helper function for polymorphic input data: 2. Array of Tensor.
*/
function isDataArray(x) {
return Array.isArray(x);
}
/**
* Helper function for polymorphic input data: 3. "dict" of Tensor.
*/
function isDataDict(x) {
return !isDataTensor(x) && !isDataArray(x);
}
/**
* Normalizes inputs and targets provided by users.
* @param data User-provided input data (polymorphic).
* @param names An Array of expected Tensor names.
* @param shapes Optional Array of expected Tensor shapes.
* @param checkBatchAxis Whether to check that the batch axis of the arrays
* match the expected value found in `shapes`.
* @param exceptionPrefix String prefix used for exception formatting.
* @returns List of standardized input Tensors (one Tensor per model input).
* @throws ValueError: in case of improperly formatted user data.
*/
function standardizeInputData(data, names, shapes, checkBatchAxis = true, exceptionPrefix = '') {
if (names == null || names.length === 0) {
// Check for the case where the model expected no data, but some data got
// sent.
if (data != null) {
let gotUnexpectedData = false;
if (isDataArray(data) && data.length > 0) {
gotUnexpectedData = true;
}
else if (isDataDict(data)) {
for (const key in data) {
if (data.hasOwnProperty(key)) {
gotUnexpectedData = true;
break;
}
}
}
else {
// `data` is a singleton Tensor in this case.
gotUnexpectedData = true;
}
if (gotUnexpectedData) {
throw new ValueError(`Error when checking model ${exceptionPrefix} expected no data, ` +
`but got ${data}`);
}
}
return [];
}
if (data == null) {
return names.map(name => null);
}
let arrays;
if (isDataDict(data)) {
data = data;
arrays = [];
for (const name of names) {
if (data[name] == null) {
throw new ValueError(`No data provided for "${name}". Need data for each key in: ` +
`${names}`);
}
arrays.push(data[name]);
}
}
else if (isDataArray(data)) {
data = data;
if (data.length !== names.length) {
throw new ValueError(`Error when checking model ${exceptionPrefix}: the Array of ` +
`Tensors that you are passing to your model is not the size the ` +
`model expected. Expected to see ${names.length} Tensor(s), but ` +
`instead got the following list of Tensor(s): ${data}`);
}
arrays = data;
}
else {
data = data;
if (names.length > 1) {
throw new ValueError(`The model ${exceptionPrefix} expects ${names.length} Tensor(s), ` +
`but only received one Tensor. Found: Tensor with shape ${data.shape}`);
}
arrays = [data];
}
arrays = ensureTensorsRank2OrHigher(arrays);
// Check shape compatibility.
if (shapes != null) {
for (let i = 0; i < names.length; ++i) {
if (shapes[i] == null) {
continue;
}
const array = arrays[i];
if (array.shape.length !== shapes[i].length) {
throw new ValueError(`Error when checking ${exceptionPrefix}: expected ${names[i]} ` +
`to have ${shapes[i].length} dimension(s). but got array with ` +
`shape ${array.shape}`);
}
for (let j = 0; j < shapes[i].length; ++j) {
if (j === 0 && !checkBatchAxis) {
// Skip the first (batch) axis.
continue;
}
const dim = array.shape[j];
const refDim = shapes[i][j];
if (refDim != null && refDim >= 0 && dim !== refDim) {
throw new ValueError(`${exceptionPrefix} expected a batch of elements where each ` +
`example has shape [${shapes[i].slice(1, shapes[i].length)}] ` +
`(i.e.,tensor shape [*,${shapes[i].slice(1, shapes[i].length)}])` +
` but the ${exceptionPrefix} received an input with ${array.shape[0]}` +
` examples, each with shape [${array.shape.slice(1, array.shape.length)}]` +
` (tensor shape [${array.shape}])`);
}
}
}
}
return arrays;
}
/**
* User input validation for Tensors.
* @param inputs `Array` of `tf.Tensor`s for inputs.
* @param targets `Array` of `tf.Tensor`s for targets.
* @param weights Optional `Array` of `tf.Tensor`s for sample weights.
* @throws ValueError: in case of incorrectly formatted data.
*/
function checkArrayLengths(inputs, targets, weights) {
const setX = unique(inputs.map(input => input.shape[0]));
setX.sort();
const setY = unique(targets.map(target => target.shape[0]));
setY.sort();
// TODO(cais): Check `weights` as well.
if (setX.length > 1) {
throw new ValueError(`All input Tensors (x) should have the same number of samples. ` +
`Got array shapes: ` +
`${JSON.stringify(inputs.map(input => input.shape))}`);
}
if (setY.length > 1) {
throw new ValueError(`All target Tensors (y) should have the same number of samples. ` +
`Got array shapes: ` +
`${JSON.stringify(targets.map(target => target.shape))}`);
}
if (setX.length > 0 && setY.length > 0 && !arraysEqual(setX, setY)) {
throw new ValueError(`Input Tensors should have the same number of samples as target ` +
`Tensors. Found ${setX[0]} input sample(s) and ${setY[0]} target ` +
`sample(s).`);
}
}
/**
* Validation on the compatibility of targes and loss functions.
*
* This helps prevent users from using loss functions incorrectly.
*
* @param targets `Array` of `tf.Tensor`s of targets.
* @param lossFns `Array` of loss functions.
* @param outputShapes `Array` of shapes of model outputs.
*/
function checkLossAndTargetCompatibility(targets, lossFns, outputShapes) {
// TODO(cais): Dedicated test coverage?
const keyLosses = [
meanSquaredError, binaryCrossentropy$1,
categoricalCrossentropy$1
];
for (let i = 0; i < targets.length; ++i) {
const y = targets[i];
const loss = lossFns[i];
const shape = outputShapes[i];
if (loss == null) {
continue;
}
if (loss === categoricalCrossentropy$1) {
if (y.shape[y.shape.length - 1] === 1) {
throw new ValueError(`You are passing a target array of shape ${y.shape} while using ` +
`a loss 'categorical_crossentropy'. 'categorical_crossentropy'` +
`expects targets to be binary matrices (1s and 0s) of shape ` +
`[samples, classes].`);
// TODO(cais): Example code in error message.
}
}
if (keyLosses.indexOf(loss) !== -1) {
const slicedYShape = y.shape.slice(1);
const slicedShape = shape.slice(1);
for (let j = 0; j < slicedYShape.length; ++j) {
const targetDim = slicedYShape[j];
const outDim = slicedShape[j];
if (outDim != null && targetDim !== outDim) {
throw new ValueError(`A target Tensor with shape ${y.shape} was passed for an ` +
`output of shape ${shape}, while using a loss function that ` +
`expects targets to have the same shape as the output.`);
}
}
}
}
}
/**
* Check inputs provided by the user.
*
* Porting Note: This corresponds to _standardize_input_data() in Python
* Keras. Because of the strong typing in TF.js, we do not need to convert
* the data. Specifically:
* 1) in PyKeras, `data` can be `DataFrame` instances from pandas, for
* example. We don't need to worry about that here because there is no
* widely popular javascript/typesdcript equivalent of pandas (so far).
* If one becomes available in the future, we can add support.
* 2) in PyKeras, inputs can be Python dict. But here we are stipulating
* that the data is either a single `tf.Tensor` or an Array of `tf.Tensor`s. We
* may add support for `Object` data inputs in the future when the need
* arises.
*
* Instead, we perform basic checks for number of parameters and shapes.
*
* @param data: The input data.
* @param names: Name for the inputs, from the model.
* @param shapes: Expected shapes for the input data, from the model.
* @param checkBatchAxis: Whether the size along the batch axis (i.e., the
* first dimension) will be checked for matching.
* @param exceptionPrefix: Execption prefix message, used in generating error
* messages.
* @throws ValueError: on incorrect number of inputs or mismatches in shapes.
*/
function checkInputData(data, names, shapes, checkBatchAxis = true, exceptionPrefix = '') {
let arrays;
if (Array.isArray(data)) {
if (data.length !== names.length) {
throw new ValueError(`Error when checking model ${exceptionPrefix}: the Array of ` +
`Tensors that you are passing to your model is not the size the ` +
`the model expected. Expected to see ${names.length} Tensor(s),` +
` but instead got ${data.length} Tensors(s).`);
}
arrays = data;
}
else {
if (names.length > 1) {
throw new ValueError(`The model expects ${names.length} ${exceptionPrefix} Tensors, ` +
`but only received one Tensor. Found: array with shape ` +
`${JSON.stringify(data.shape)}.`);
}
arrays = [data];
}
if (shapes != null) {
for (let i = 0; i < names.length; ++i) {
if (shapes[i] == null) {
continue;
}
const array = arrays[i];
if (array.shape.length !== shapes[i].length) {
throw new ValueError(`Error when checking ${exceptionPrefix}: expected ${names[i]} ` +
`to have ${shapes[i].length} dimension(s), but got array with ` +
`shape ${JSON.stringify(array.shape)}`);
}
for (let j = 0; j < shapes[i].length; ++j) {
if (j === 0 && !checkBatchAxis) {
continue;
}
const dim = array.shape[j];
const refDim = shapes[i][j];
if (refDim != null) {
if (refDim !== dim) {
throw new ValueError(`Error when checking ${exceptionPrefix}: expected ` +
`${names[i]} to have shape ${JSON.stringify(shapes[i])} but ` +
`got array with shape ${JSON.stringify(array.shape)}.`);
}
}
}
}
}
}
/**
* Maps metric functions to model outputs.
* @param metrics An shortcut strings name, metric function, `Array` or dict
* (`Object`) of metric functions.
* @param outputNames An `Array` of the names of model outputs.
* @returns An `Array` (one entry per model output) of `Array` of metric
* functions. For instance, if the model has 2 outputs, and for the first
* output we want to compute `binaryAccuracy` and `binaryCrossentropy`,
* and just `binaryAccuracy` for the second output, the `Array` would look
* like:
* `[[binaryAccuracy, binaryCrossentropy], [binaryAccuracy]]`
* @throws TypeError: incompatible metrics format.
*/
function collectMetrics(metrics, outputNames) {
if (metrics == null || Array.isArray(metrics) && metrics.length === 0) {
return outputNames.map(name => []);
}
let wrappedMetrics;
if (typeof metrics === 'string' || typeof metrics === 'function') {
wrappedMetrics = [metrics];
}
else if (Array.isArray(metrics) || typeof metrics === 'object') {
wrappedMetrics = metrics;
}
else {
throw new TypeError('Type of metrics argument not understood. Expected an string,' +
`function, Array, or Object, found: ${metrics}`);
}
if (Array.isArray(wrappedMetrics)) {
// We then apply all metrics to all outputs.
return outputNames.map(name => wrappedMetrics);
}
else {
// In this case, metrics is a dict.
const nestedMetrics = [];
for (const name of outputNames) {
let outputMetrics = wrappedMetrics.hasOwnProperty(name) ? wrappedMetrics[name] : [];
if (!Array.isArray(outputMetrics)) {
outputMetrics = [outputMetrics];
}
nestedMetrics.push(outputMetrics);
}
return nestedMetrics;
}
}
const LAYERS_MODEL_FORMAT_NAME = 'layers-model';
/**
* A `tf.LayersModel` is a directed, acyclic graph of `tf.Layer`s plus methods
* for training, evaluation, prediction and saving.
*
* `tf.LayersModel` is the basic unit of training, inference and evaluation in
* TensorFlow.js. To create a `tf.LayersModel`, use `tf.LayersModel`.
*
* See also:
* `tf.Sequential`, `tf.loadLayersModel`.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
class LayersModel extends Container {
constructor(args) {
super(args);
this.isTraining = false;
}
/**
* Print a text summary of the model's layers.
*
* The summary includes
* - Name and type of all layers that comprise the model.
* - Output shape(s) of the layers
* - Number of weight parameters of each layer
* - If the model has non-sequential-like topology, the inputs each layer
* receives
* - The total number of trainable and non-trainable parameters of the model.
*
* ```js
* const input1 = tf.input({shape: [10]});
* const input2 = tf.input({shape: [20]});
* const dense1 = tf.layers.dense({units: 4}).apply(input1);
* const dense2 = tf.layers.dense({units: 8}).apply(input2);
* const concat = tf.layers.concatenate().apply([dense1, dense2]);
* const output =
* tf.layers.dense({units: 3, activation: 'softmax'}).apply(concat);
*
* const model = tf.model({inputs: [input1, input2], outputs: output});
* model.summary();
* ```
*
* @param lineLength Custom line length, in number of characters.
* @param positions Custom widths of each of the columns, as either
* fractions of `lineLength` (e.g., `[0.5, 0.75, 1]`) or absolute number
* of characters (e.g., `[30, 50, 65]`). Each number corresponds to
* right-most (i.e., ending) position of a column.
* @param printFn Custom print function. Can be used to replace the default
* `console.log`. For example, you can use `x => {}` to mute the printed
* messages in the console.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
summary(lineLength, positions, printFn = console.log) {
if (!this.built) {
throw new ValueError(`This model has never been called, thus its weights have not been ` +
`created yet. So no summary can be displayed. Build the model ` +
`first (e.g., by calling it on some test data).`);
}
printSummary(this, lineLength, positions, printFn);
}
/**
* Configures and prepares the model for training and evaluation. Compiling
* outfits the model with an optimizer, loss, and/or metrics. Calling `fit`
* or `evaluate` on an un-compiled model will throw an error.
*
* @param args a `ModelCompileArgs` specifying the loss, optimizer, and
* metrics to be used for fitting and evaluating this model.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
compile(args) {
if (args.loss == null) {
args.loss = [];
}
this.loss = args.loss;
if (typeof args.optimizer === 'string') {
this.optimizer_ = getOptimizer(args.optimizer);
this.isOptimizerOwned = true;
}
else {
if (!(args.optimizer instanceof Optimizer)) {
throw new ValueError(`User-defined optimizer must be an instance of tf.Optimizer.`);
}
this.optimizer_ = args.optimizer;
this.isOptimizerOwned = false;
}
// TODO(cais): Add lossWeights.
// TODO(cais): Add sampleWeightMode.
// Prepare loss functions.
let lossFunctions = [];
if (!Array.isArray(args.loss) && typeof args.loss !== 'string' &&
typeof args.loss !== 'function') {
args.loss = args.loss;
for (const name in args.loss) {
if (this.outputNames.indexOf(name) === -1) {
throw new ValueError(`Unknown entry in loss dictionary: "${name}". ` +
`Only expected the following keys: ${this.outputNames}`);
}
}
for (const name of this.outputNames) {
if (args.loss[name] == null) {
console.warn(`Output "${name}" is missing from loss dictionary. We assume ` +
`this was done on purpose, and we will not be expecting data ` +
`to be passed to ${name} during training`);
}
lossFunctions.push(get$1(args.loss[name]));
}
}
else if (Array.isArray(args.loss)) {
if (args.loss.length !== this.outputs.length) {
throw new ValueError(`When passing an Array as loss, it should have one entry per ` +
`model output. The model has ${this.outputs.length} output(s), ` +
`but you passed loss=${args.loss}.`);
}
const theLosses = args.loss;
lossFunctions = theLosses.map(l => get$1(l));
}
else {
const lossFunction = get$1(args.loss);
this.outputs.forEach(_ => {
lossFunctions.push(lossFunction);
});
}
this.lossFunctions = lossFunctions;
this.feedOutputNames = [];
this.feedOutputShapes = [];
this.feedLossFns = [];
for (let i = 0; i < this.outputs.length; ++i) {
// TODO(cais): Logic for skipping target(s).
const shape = this.internalOutputShapes[i];
const name = this.outputNames[i];
this.feedOutputNames.push(name);
this.feedOutputShapes.push(shape);
this.feedLossFns.push(this.lossFunctions[i]);
}
// TODO(cais): Add logic for output masks.
// TODO(cais): Add logic for sample weights.
const skipTargetIndices = [];
// Prepare metrics.
this.metrics = args.metrics;
// TODO(cais): Add weightedMetrics.
this.metricsNames = ['loss'];
this.metricsTensors = [];
// Compute total loss.
// Porting Note: In PyKeras, metrics_tensors are symbolic tensor objects.
// Here, metricsTensors are TypeScript functions. This difference is due
// to the difference in symbolic/imperative property of the backends.
nameScope('loss', () => {
for (let i = 0; i < this.outputs.length; ++i) {
if (skipTargetIndices.indexOf(i) !== -1) {
continue;
}
// TODO(cais): Add weightedLoss, sampleWeight and mask.
// The following line should be weightedLoss
const weightedLoss = this.lossFunctions[i];
if (this.outputs.length > 1) {
this.metricsTensors.push([weightedLoss, i]);
this.metricsNames.push(this.outputNames[i] + '_loss');
}
}
// Porting Note: Due to the imperative nature of the backend, we calculate
// the regularizer penalties in the totalLossFunction, instead of here.
});
const nestedMetrics = collectMetrics(args.metrics, this.outputNames);
// TODO(cais): Add nestedWeightedMetrics.
/**
* Helper function used in loop below.
*/
const appendMetric = (outputIndex, metricName, metricTensor) => {
if (this.outputNames.length > 1) {
metricName = this.outputNames[outputIndex] + '_' + metricName;
}
this.metricsNames.push(metricName);
this.metricsTensors.push([metricTensor, outputIndex]);
};
nameScope('metric', () => {
for (let i = 0; i < this.outputs.length; ++i) {
if (skipTargetIndices.indexOf(i) !== -1) {
continue;
}
const outputMetrics = nestedMetrics[i];
// TODO(cais): Add weights and outputWeightedMetrics.
// TODO(cais): Add optional arg `weights` to the following function.
const handleMetrics = (metrics) => {
const metricNamePrefix = '';
let metricName;
let accFn;
let weightedMetricFn;
// TODO(cais): Use 'weights_' for weighted metrics.
for (const metric of metrics) {
if (typeof metric === 'string' &&
['accuracy', 'acc', 'crossentropy', 'ce'].indexOf(metric) !==
-1) {
const outputShape = this.internalOutputShapes[i];
if (outputShape[outputShape.length - 1] === 1 ||
this.lossFunctions[i] === binaryCrossentropy$1) {
// case: binary accuracy/crossentropy.
if (['accuracy', 'acc'].indexOf(metric) !== -1) {
accFn = binaryAccuracy;
}
else if (['crossentropy', 'ce'].indexOf(metric) !== -1) {
accFn = binaryCrossentropy;
}
}
else if (this.lossFunctions[i] ===
sparseCategoricalCrossentropy$1) {
// case: categorical accuracy / crossentropy with sparse
// targets.
if (['accuracy', 'acc'].indexOf(metric) !== -1) {
accFn = sparseCategoricalAccuracy;
}
else if (['crossentropy', 'ce'].indexOf(metric) !== -1) {
accFn = sparseCategoricalCrossentropy;
}
}
else {
// case: categorical accuracy / crossentropy.
if (['accuracy', 'acc'].indexOf(metric) !== -1) {
accFn = categoricalAccuracy;
}
else if (['crossentropy', 'ce'].indexOf(metric) !== -1) {
accFn = categoricalCrossentropy;
}
}
let suffix;
if (['accuracy', 'acc'].indexOf(metric) !== -1) {
suffix = 'acc';
}
else if (['crossentropy', 'ce'].indexOf(metric) !== -1) {
suffix = 'ce';
}
// TODO(cais): Add weighting actually.
weightedMetricFn = accFn;
metricName = metricNamePrefix + suffix;
}
else {
const metricFn = get(metric);
// TODO(cais): Add weighting actually.
weightedMetricFn = metricFn;
metricName =
metricNamePrefix + getLossOrMetricName(metric);
}
// TODO(cais): Add weighting and masking to metricResult.
let metricResult;
nameScope(metricName, () => {
metricResult = weightedMetricFn;
});
appendMetric(i, metricName, metricResult);
}
};
handleMetrics(outputMetrics);
// TODO(cais): Call handleMetrics with weights.
}
});
// Porting Notes: Given the imperative backend of tfjs-core,
// there is no need for constructing the symbolic graph and placeholders.
this.collectedTrainableWeights = this.trainableWeights;
}
/**
* Check trainable weights count consistency.
*
* This will raise a warning if `this.trainableWeights` and
* `this.collectedTrainableWeights` are inconsistent (i.e., have different
* numbers of parameters).
* Inconsistency will typically arise when one modifies `model.trainable`
* without calling `model.compile()` again.
*/
checkTrainableWeightsConsistency() {
if (this.collectedTrainableWeights == null) {
return;
}
if (this.trainableWeights.length !==
this.collectedTrainableWeights.length) {
console.warn('Discrepancy between trainableweights and collected trainable ' +
'weights. Did you set `model.trainable` without calling ' +
'`model.compile()` afterwards?');
}
}
/**
* Returns the loss value & metrics values for the model in test mode.
*
* Loss and metrics are specified during `compile()`, which needs to happen
* before calls to `evaluate()`.
*
* Computation is done in batches.
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [10]})]
* });
* model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
* const result = model.evaluate(
* tf.ones([8, 10]), tf.ones([8, 1]), {batchSize: 4});
* result.print();
* ```
*
* @param x `tf.Tensor` of test data, or an `Array` of `tf.Tensor`s if the
* model has multiple inputs.
* @param y `tf.Tensor` of target data, or an `Array` of `tf.Tensor`s if the
* model has multiple outputs.
* @param args A `ModelEvaluateArgs`, containing optional fields.
*
* @return `Scalar` test loss (if the model has a single output and no
* metrics) or `Array` of `Scalar`s (if the model has multiple outputs
* and/or metrics). The attribute `model.metricsNames`
* will give you the display labels for the scalar outputs.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
evaluate(x, y, args = {}) {
const batchSize = args.batchSize == null ? 32 : args.batchSize;
checkBatchSize(batchSize);
// TODO(cais): Standardize `config.sampleWeights` as well.
// Validate user data.
const checkBatchAxis = true;
const standardizedOuts = this.standardizeUserDataXY(x, y, checkBatchAxis, batchSize);
try {
// TODO(cais): If uses `useLearningPhase`, set the corresponding element
// of the input to 0.
const ins = standardizedOuts[0].concat(standardizedOuts[1]);
this.makeTestFunction();
const f = this.testFunction;
const testOuts = this.testLoop(f, ins, batchSize, args.verbose, args.steps);
return singletonOrArray(testOuts);
}
finally {
disposeNewTensors(standardizedOuts[0], x);
disposeNewTensors(standardizedOuts[1], y);
}
}
// TODO(cais): Add code snippet below once real dataset objects are
// available.
/**
* Evaluate model using a dataset object.
*
* Note: Unlike `evaluate()`, this method is asynchronous (`async`).
*
* @param dataset A dataset object. Its `iterator()` method is expected
* to generate a dataset iterator object, the `next()` method of which
* is expected to produce data batches for evaluation. The return value
* of the `next()` call ought to contain a boolean `done` field and a
* `value` field. The `value` field is expected to be an array of two
* `tf.Tensor`s or an array of two nested `tf.Tensor` structures. The former
* case is for models with exactly one input and one output (e.g.
* a sequential model). The latter case is for models with multiple
* inputs and/or multiple outputs. Of the two items in the array, the
* first is the input feature(s) and the second is the output target(s).
* @param args A configuration object for the dataset-based evaluation.
* @returns Loss and metric values as an Array of `Scalar` objects.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
async evaluateDataset(dataset, args) {
this.makeTestFunction();
return evaluateDataset(this, dataset, args);
}
/**
* Get number of samples provided for training, evaluation or prediction.
*
* @param ins Input `tf.Tensor`.
* @param batchSize Integer batch size, optional.
* @param steps Total number of steps (batches of samples) before
* declaring loop finished. Optional.
* @param stepsName The public API's parameter name for `steps`.
* @returns Number of samples provided.
*/
checkNumSamples(ins, batchSize, steps, stepsName = 'steps') {
let numSamples;
if (steps != null) {
numSamples = null;
if (batchSize != null) {
throw new ValueError(`If ${stepsName} is set, batchSize must be null or undefined.` +
`Got batchSize = ${batchSize}`);
}
}
else if (ins != null) {
if (Array.isArray(ins)) {
numSamples = ins[0].shape[0];
}
else {
numSamples = ins.shape[0];
}
}
else {
throw new ValueError(`Either the input data should have a defined shape, or ` +
`${stepsName} shoud be specified.`);
}
return numSamples;
}
/**
* Execute internal tensors of the model with input data feed.
* @param inputs Input data feed. Must match the inputs of the model.
* @param outputs Names of the output tensors to be fetched. Must match
* names of the SymbolicTensors that belong to the graph.
* @returns Fetched values for `outputs`.
*/
execute(inputs, outputs) {
if (Array.isArray(outputs) && outputs.length === 0) {
throw new ValueError('`outputs` is an empty Array, which is not allowed.');
}
const outputsIsArray = Array.isArray(outputs);
const outputNames = (outputsIsArray ? outputs : [outputs]);
const outputSymbolicTensors = this.retrieveSymbolicTensors(outputNames);
// Format the input into a FeedDict.
const feedDict = new FeedDict();
if (inputs instanceof Tensor) {
inputs = [inputs];
}
if (Array.isArray(inputs)) {
if (inputs.length !== this.inputs.length) {
throw new ValueError(`The number of inputs provided (${inputs.length}) ` +
`does not match the number of inputs of this model ` +
`(${this.inputs.length}).`);
}
for (let i = 0; i < this.inputs.length; ++i) {
feedDict.add(this.inputs[i], inputs[i]);
}
}
else {
for (const input of this.inputs) {
const tensorValue = inputs[input.name];
if (tensorValue == null) {
throw new ValueError(`No value is provided for the model's input ${input.name}`);
}
feedDict.add(input, tensorValue);
}
}
// Run execution.
const executeOutputs = execute(outputSymbolicTensors, feedDict);
return outputsIsArray ? executeOutputs : executeOutputs[0];
}
/**
* Retrieve the model's internal symbolic tensors from symbolic-tensor names.
*/
retrieveSymbolicTensors(symbolicTensorNames) {
const outputSymbolicTensors = pyListRepeat(null, symbolicTensorNames.length);
let outputsRemaining = symbolicTensorNames.length;
for (const layer of this.layers) {
const layerOutputs = Array.isArray(layer.output) ? layer.output : [layer.output];
const layerOutputNames = layerOutputs.map(output => output.name);
for (let i = 0; i < symbolicTensorNames.length; ++i) {
const index = layerOutputNames.indexOf(symbolicTensorNames[i]);
if (index !== -1) {
outputSymbolicTensors[i] = layerOutputs[index];
outputsRemaining--;
}
if (outputsRemaining === 0) {
break;
}
}
if (outputsRemaining === 0) {
break;
}
}
if (outputsRemaining > 0) {
const remainingNames = [];
outputSymbolicTensors.forEach((tensor, i) => {
if (tensor == null) {
remainingNames.push(symbolicTensorNames[i]);
}
});
throw new ValueError(`Cannot find SymbolicTensors for output name(s): ` +
`${JSON.stringify(remainingNames)}`);
}
return outputSymbolicTensors;
}
/**
* Helper method to loop over some data in batches.
*
* Porting Note: Not using the functional approach in the Python equivalent
* due to the imperative backend.
* Porting Note: Does not support step mode currently.
*
* @param ins: input data
* @param batchSize: integer batch size.
* @param verbose: verbosity model
* @returns: Predictions as `tf.Tensor` (if a single output) or an `Array` of
* `tf.Tensor` (if multipe outputs).
*/
predictLoop(ins, batchSize = 32, verbose = false) {
return tidy(() => {
const numSamples = this.checkNumSamples(ins);
if (verbose) {
throw new NotImplementedError('Verbose predictLoop() is not implemented yet.');
}
// Sample-based predictions.
// Porting Note: Tensor currently does not support sliced assignments as
// in numpy, e.g., x[1:3] = y. Therefore we use concatenation while
// iterating over the batches.
const batches = makeBatches(numSamples, batchSize);
const outsBatches = this.outputs.map(output => []);
// TODO(cais): Can the scope() be pushed down inside the for loop?
for (let batchIndex = 0; batchIndex < batches.length; ++batchIndex) {
const batchOuts = tidy(() => {
const batchStart = batches[batchIndex][0];
const batchEnd = batches[batchIndex][1];
// TODO(cais): Take care of the case of the last element is a flag for
// training/test.
const insBatch = sliceArrays(ins, batchStart, batchEnd);
// Construct the feeds for execute();
const feeds = [];
if (Array.isArray(insBatch)) {
for (let i = 0; i < insBatch.length; ++i) {
feeds.push({ key: this.inputs[i], value: insBatch[i] });
}
}
else {
feeds.push({ key: this.inputs[0], value: insBatch });
}
const feedDict = new FeedDict(feeds);
return execute(this.outputs, feedDict);
});
batchOuts.forEach((batchOut, i) => outsBatches[i].push(batchOut));
}
return singletonOrArray(outsBatches.map(batches => concat$2(batches, 0)));
});
}
/**
* Generates output predictions for the input samples.
*
* Computation is done in batches.
*
* Note: the "step" mode of predict() is currently not supported.
* This is because the TensorFlow.js core backend is imperative only.
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [10]})]
* });
* model.predict(tf.ones([8, 10]), {batchSize: 4}).print();
* ```
*
* @param x The input data, as a Tensor, or an `Array` of `tf.Tensor`s if
* the model has multiple inputs.
* @param args A `ModelPredictArgs` object containing optional fields.
*
* @return Prediction results as a `tf.Tensor`(s).
*
* @exception ValueError In case of mismatch between the provided input data
* and the model's expectations, or in case a stateful model receives a
* number of samples that is not a multiple of the batch size.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
predict(x, args = {}) {
const xsRank2OrHigher = ensureTensorsRank2OrHigher(x);
checkInputData(xsRank2OrHigher, this.inputNames, this.feedInputShapes, false);
try {
// TODO(cais): Take care of stateful models.
// if (this.stateful) ...
// TODO(cais): Take care of the learning_phase boolean flag.
// if (this.useLearningPhase) ...
const batchSize = args.batchSize == null ? 32 : args.batchSize;
checkBatchSize(batchSize);
return this.predictLoop(xsRank2OrHigher, batchSize);
}
finally {
disposeNewTensors(xsRank2OrHigher, x);
}
}
/**
* Returns predictions for a single batch of samples.
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [10]})]
* });
* model.predictOnBatch(tf.ones([8, 10])).print();
* ```
* @param x: Input samples, as a Tensor (for models with exactly one
* input) or an array of Tensors (for models with more than one input).
* @return Tensor(s) of predictions
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
predictOnBatch(x) {
checkInputData(x, this.inputNames, this.feedInputShapes, true);
// TODO(cais): Take care of the learning_phase boolean flag.
// if (this.useLearningPhase) ...
const batchSize = (Array.isArray(x) ? x[0] : x).shape[0];
return this.predictLoop(x, batchSize);
}
standardizeUserDataXY(x, y, checkBatchAxis = true, batchSize) {
// TODO(cais): Add sampleWeight, classWeight
if (this.optimizer_ == null) {
throw new RuntimeError('You must compile a model before training/testing. Use ' +
'LayersModel.compile(modelCompileArgs).');
}
const outputShapes = [];
for (let i = 0; i < this.feedOutputShapes.length; ++i) {
const outputShape = this.feedOutputShapes[i];
const lossFn = this.feedLossFns[i];
if (lossFn === sparseCategoricalCrossentropy$1) {
outputShapes.push(outputShape.slice(0, outputShape.length - 1).concat([1]));
}
else {
// Porting Note: Because of strong typing `lossFn` must be a function.
outputShapes.push(outputShape);
}
}
x = standardizeInputData(x, this.feedInputNames, this.feedInputShapes, false, 'input');
y = standardizeInputData(y, this.feedOutputNames, outputShapes, false, 'target');
// TODO(cais): Standardize sampleWeights & classWeights.
checkArrayLengths(x, y);
// TODO(cais): Check sampleWeights as well.
checkLossAndTargetCompatibility(y, this.feedLossFns, this.feedOutputShapes);
if (this.stateful && batchSize != null && batchSize > 0) {
if (x[0].shape[0] % batchSize !== 0) {
throw new ValueError(`In a stateful network, you should only pass inputs with a ` +
`number of samples that is divisible by the batch size ` +
`${batchSize}. Found: ${x[0].shape[0]} sample(s).`);
}
}
return [x, y];
}
async standardizeUserData(x, y, sampleWeight, classWeight, checkBatchAxis = true, batchSize) {
const [standardXs, standardYs] = this.standardizeUserDataXY(x, y, checkBatchAxis, batchSize);
// TODO(cais): Handle sampleWeights.
if (sampleWeight != null) {
throw new Error('sample weight is not supported yet.');
}
let standardSampleWeights = null;
if (classWeight != null) {
const classWeights = standardizeClassWeights(classWeight, this.outputNames);
standardSampleWeights = [];
for (let i = 0; i < classWeights.length; ++i) {
standardSampleWeights.push(await standardizeWeights(standardYs[i], null, classWeights[i]));
}
}
// TODO(cais): Deal with the case of model.stateful == true.
return [standardXs, standardYs, standardSampleWeights];
}
/**
* Loop over some test data in batches.
* @param f A Function returning a list of tensors.
* @param ins Array of tensors to be fed to `f`.
* @param batchSize Integer batch size or `null` / `undefined`.
* @param verbose verbosity mode.
* @param steps Total number of steps (batches of samples) before
* declaring test finished. Ignored with the default value of `null` /
* `undefined`.
* @returns Array of Scalars.
*/
testLoop(f, ins, batchSize, verbose = 0, steps) {
return tidy(() => {
const numSamples = this.checkNumSamples(ins, batchSize, steps, 'steps');
const outs = [];
if (verbose > 0) {
throw new NotImplementedError('Verbose mode is not implemented yet.');
}
// TODO(cais): Use `indicesForConversionToDense' to prevent slow down.
if (steps != null) {
throw new NotImplementedError('steps mode in testLoop() is not implemented yet');
}
else {
const batches = makeBatches(numSamples, batchSize);
const indexArray = tensor1d(range(0, numSamples));
for (let batchIndex = 0; batchIndex < batches.length; ++batchIndex) {
const batchStart = batches[batchIndex][0];
const batchEnd = batches[batchIndex][1];
const batchIds = sliceAlongFirstAxis(indexArray, batchStart, batchEnd - batchStart);
// TODO(cais): In ins, train flag can be a number, instead of an
// Tensor? Do we need to handle this in tfjs-layers?
const insBatch = sliceArraysByIndices(ins, batchIds);
const batchOuts = f(insBatch);
if (batchIndex === 0) {
for (let i = 0; i < batchOuts.length; ++i) {
outs.push(scalar(0));
}
}
for (let i = 0; i < batchOuts.length; ++i) {
const batchOut = batchOuts[i];
outs[i] =
add$1(outs[i], mul(batchEnd - batchStart, batchOut));
}
}
for (let i = 0; i < outs.length; ++i) {
outs[i] = div$1(outs[i], numSamples);
}
}
return outs;
});
}
getDedupedMetricsNames() {
const outLabels = this.metricsNames;
// Rename duplicated metrics names (can happen with an output layer
// shared among multiple dataflows).
const dedupedOutLabels = [];
for (let i = 0; i < outLabels.length; ++i) {
const label = outLabels[i];
let newLabel = label;
if (count(outLabels, label) > 1) {
const dupIndex = count(outLabels.slice(0, i), label);
newLabel += `_${dupIndex}`;
}
dedupedOutLabels.push(newLabel);
}
return dedupedOutLabels;
}
/**
* Creates a function that performs the following actions:
*
* 1. computes the losses
* 2. sums them to get the total loss
* 3. call the optimizer computes the gradients of the LayersModel's
* trainable weights w.r.t. the total loss and update the variables
* 4. calculates the metrics
* 5. returns the values of the losses and metrics.
*/
makeTrainFunction() {
return (data) => {
const lossValues = [];
const inputs = data.slice(0, this.inputs.length);
const targets = data.slice(this.inputs.length, this.inputs.length + this.outputs.length);
const sampleWeights = data.slice(this.inputs.length + this.outputs.length, this.inputs.length + this.outputs.length * 2);
const metricsValues = [];
// Create a function that computes the total loss based on the
// inputs. This function is used for obtaining gradients through
// backprop.
const totalLossFunction = () => {
const feeds = [];
for (let i = 0; i < this.inputs.length; ++i) {
feeds.push({ key: this.inputs[i], value: inputs[i] });
}
const feedDict = new FeedDict(feeds);
const outputs = execute(this.outputs, feedDict, { 'training': true });
// TODO(cais): Take care of the case of multiple outputs from a
// single layer?
let totalLoss;
for (let i = 0; i < this.lossFunctions.length; ++i) {
const lossFunction = this.lossFunctions[i];
let loss = lossFunction(targets[i], outputs[i]);
if (sampleWeights[i] != null) {
loss = computeWeightedLoss(loss, sampleWeights[i]);
}
// TODO(cais): push Scalar instead.
const meanLoss = mean$1(loss);
// TODO(cais): Use a scope() instead, to avoid ownership.
lossValues.push(meanLoss);
if (i === 0) {
totalLoss = loss;
}
else {
totalLoss = add$1(totalLoss, loss);
}
}
// Compute the metrics.
// TODO(cais): These should probably be calculated outside
// totalLossFunction to benefit speed?
for (let i = 0; i < this.metricsTensors.length; ++i) {
let weightedMetric;
if (this.outputs.length > 1 && i < this.outputs.length) {
weightedMetric = lossValues[i];
}
else {
const metric = this.metricsTensors[i][0];
const outputIndex = this.metricsTensors[i][1];
weightedMetric =
mean$1(metric(targets[outputIndex], outputs[outputIndex]));
}
keep(weightedMetric);
// TODO(cais): Use a scope() instead, to avoid ownership.
metricsValues.push(weightedMetric);
}
totalLoss = mean$1(totalLoss);
// Add regularizer penalties.
this.calculateLosses().forEach(regularizerLoss => {
totalLoss = add$1(totalLoss, regularizerLoss);
});
return totalLoss;
};
const variables = this.collectedTrainableWeights.map(param => param.read());
const returnCost = true;
const totalLossValue = this.optimizer_.minimize(totalLossFunction, returnCost, variables);
return [totalLossValue].concat(metricsValues);
};
}
/**
* Create a function which, when invoked with an array of `tf.Tensor`s as a
* batch of inputs, returns the prespecified loss and metrics of the model
* under the batch of input data.
*/
makeTestFunction() {
this.testFunction = (data) => {
return tidy(() => {
const valOutputs = [];
let totalLoss;
const inputs = data.slice(0, this.inputs.length);
const targets = data.slice(this.inputs.length, this.inputs.length + this.outputs.length);
const feeds = [];
for (let i = 0; i < this.inputs.length; ++i) {
feeds.push({ key: this.inputs[i], value: inputs[i] });
}
const feedDict = new FeedDict(feeds);
const outputs = execute(this.outputs, feedDict);
// Compute total loss.
for (let i = 0; i < this.lossFunctions.length; ++i) {
const lossFunction = this.lossFunctions[i];
// TODO(cais): Add sample weighting and replace the simple
// averaging.
const loss = mean$1(lossFunction(targets[i], outputs[i]));
if (i === 0) {
totalLoss = loss;
}
else {
totalLoss = add$1(totalLoss, loss);
}
valOutputs.push(totalLoss);
}
// Compute the metrics.
for (let i = 0; i < this.metricsTensors.length; ++i) {
const metric = this.metricsTensors[i][0];
const outputIndex = this.metricsTensors[i][1];
// TODO(cais): Replace K.mean() with a proper weighting function.
const meanMetric = mean$1(metric(targets[outputIndex], outputs[outputIndex]));
valOutputs.push(meanMetric);
}
return valOutputs;
});
};
}
/**
* Trains the model for a fixed number of epochs (iterations on a
* dataset).
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [10]})]
* });
* model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
* for (let i = 1; i < 5 ; ++i) {
* const h = await model.fit(tf.ones([8, 10]), tf.ones([8, 1]), {
* batchSize: 4,
* epochs: 3
* });
* console.log("Loss after Epoch " + i + " : " + h.history.loss[0]);
* }
* ```
*
* @param x `tf.Tensor` of training data, or an array of `tf.Tensor`s if the
* model has multiple inputs. If all inputs in the model are named, you
* can also pass a dictionary mapping input names to `tf.Tensor`s.
* @param y `tf.Tensor` of target (label) data, or an array of `tf.Tensor`s if
* the model has multiple outputs. If all outputs in the model are named,
* you can also pass a dictionary mapping output names to `tf.Tensor`s.
* @param args A `ModelFitArgs`, containing optional fields.
*
* @return A `History` instance. Its `history` attribute contains all
* information collected during training.
*
* @exception ValueError In case of mismatch between the provided input
* data and what the model expects.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
async fit(x, y, args = {}) {
if (this.isTraining) {
throw new Error('Cannot start training because another fit() call is ongoing.');
}
this.isTraining = true;
let inputs;
let targets;
let originalInputs;
let originalTargets;
let inputValX;
let inputValY;
let valX;
let valY;
let sampleWeights;
try {
const batchSize = args.batchSize == null ? 32 : args.batchSize;
checkBatchSize(batchSize);
// Validate user data.
// TODO(cais): Support sampleWeight.
const checkBatchAxis = false;
const standardizedOuts = await this.standardizeUserData(x, y, args.sampleWeight, args.classWeight, checkBatchAxis, batchSize);
inputs = standardizedOuts[0];
targets = standardizedOuts[1];
sampleWeights = standardizedOuts[2];
// Prepare validation data.
let doValidation = false;
let valIns;
if (args.validationData != null && args.validationData.length > 0) {
doValidation = true;
if (args.validationData.length === 2) {
// config.validationData consists of valX and valY.
inputValX = args.validationData[0];
inputValY = args.validationData[1];
}
else if (args.validationData.length === 3) {
throw new NotImplementedError('validationData including sample weights is not supported yet.');
}
else {
throw new ValueError(`When passing validation data, it must contain 2 (valX, valY) ` +
`or 3 (valX, valY, valSampleWeight) items; ` +
`${args.validationData} is invalid.`);
}
const checkBatchAxis = true;
const valStandardized = await this.standardizeUserData(inputValX, inputValY, null, /** Unused sample weights. */ null, /** Unused class weights. */ checkBatchAxis, batchSize);
valX = valStandardized[0];
valY = valStandardized[1];
valIns = valX.concat(valY);
// TODO(cais): Add useLearningPhase data properly.
}
else if (args.validationSplit != null && args.validationSplit > 0 &&
args.validationSplit < 1) {
doValidation = true;
// Porting Note: In tfjs-layers, inputs[0] is always a Tensor.
const splitAt = Math.floor(inputs[0].shape[0] * (1 - args.validationSplit));
const originalBatchSize = inputs[0].shape[0];
valX = sliceArrays(inputs, splitAt, originalBatchSize);
originalInputs = inputs;
inputs = sliceArrays(inputs, 0, splitAt);
valY = sliceArrays(targets, splitAt, originalBatchSize);
originalTargets = targets;
targets = sliceArrays(targets, 0, splitAt);
// TODO(cais): Once sampleWeights becomes available, slice it to get
// valSampleWeights.
valIns = valX.concat(valY);
// TODO(cais): Add useLearningPhase data properly.
}
else if (args.validationSteps != null) {
doValidation = true;
// TODO(cais): Add useLearningPhase.
}
const ins = inputs.concat(targets).concat(sampleWeights);
this.checkTrainableWeightsConsistency();
// TODO(cais): Handle use_learning_phase and learning_phase?
// Porting Note: Here we see a key deviation of tfjs-layers from
// Keras.
// Due to the imperative nature of tfjs-layers' backend (tfjs-core),
// we do not construct symbolic computation graphs to embody the
// training process. Instead, we define a function that performs the
// training action. In PyKeras, the data (inputs and targets) are fed
// through graph placeholders. In tfjs-layers, the data are fed as
// function arguments. Since the function are defined below in the
// scope, we don't have equivalents of PyKeras's
// `_make_train_funciton`.
const trainFunction = this.makeTrainFunction();
const outLabels = this.getDedupedMetricsNames();
let valFunction;
let callbackMetrics;
if (doValidation) {
this.makeTestFunction();
valFunction = this.testFunction;
callbackMetrics =
outLabels.slice().concat(outLabels.map(n => 'val_' + n));
}
else {
valFunction = null;
valIns = [];
callbackMetrics = outLabels.slice();
}
const callbacks = standardizeCallbacks(args.callbacks, args.yieldEvery);
const out = await this.fitLoop(trainFunction, ins, outLabels, batchSize, args.epochs, args.verbose, callbacks, valFunction, valIns, args.shuffle, callbackMetrics, args.initialEpoch, null, null);
return out;
}
finally {
this.isTraining = false;
// Memory clean up.
disposeNewTensors(inputs, x);
disposeNewTensors(targets, y);
disposeNewTensors(originalInputs, x);
disposeNewTensors(originalTargets, y);
disposeNewTensors(valX, inputValX);
disposeNewTensors(valY, inputValY);
if (sampleWeights != null) {
dispose(sampleWeights);
}
}
// TODO(cais): Add value to outLabels.
}
/**
* Abstract fit function for `f(ins)`.
* @param f A Function returning a list of tensors. For training, this
* function is expected to perform the updates to the variables.
* @param ins List of tensors to be fed to `f`.
* @param outLabels List of strings, display names of the outputs of `f`.
* @param batchSize Integer batch size or `== null` if unknown. Default : 32.
* @param epochs Number of times to iterate over the data. Default : 1.
* @param verbose Verbosity mode: 0, 1, or 2. Default: 1.
* @param callbacks List of callbacks to be called during training.
* @param valF Function to call for validation.
* @param valIns List of tensors to be fed to `valF`.
* @param shuffle Whether to shuffle the data at the beginning of every
* epoch. Default : true.
* @param callbackMetrics List of strings, the display names of the metrics
* passed to the callbacks. They should be the concatenation of the
* display names of the outputs of `f` and the list of display names
* of the outputs of `valF`.
* @param initialEpoch Epoch at which to start training (useful for
* resuming a previous training run). Default : 0.
* @param stepsPerEpoch Total number of steps (batches on samples) before
* declaring one epoch finished and starting the next epoch. Ignored with
* the default value of `undefined` or `null`.
* @param validationSteps Number of steps to run validation for (only if
* doing validation from data tensors). Not applicable for tfjs-layers.
* @returns A `History` object.
*/
async fitLoop(f, ins, outLabels, batchSize, epochs, verbose, callbacks, valF, valIns, shuffle$1, callbackMetrics, initialEpoch, stepsPerEpoch, validationSteps) {
if (batchSize == null) {
batchSize = 32;
}
if (epochs == null) {
epochs = 1;
}
if (shuffle$1 == null) {
shuffle$1 = true;
}
if (initialEpoch == null) {
initialEpoch = 0;
}
// TODO(cais): Change const to let below when implementing validation.
let doValidation = false;
if (valF != null && valIns != null) {
doValidation = true;
// TODO(cais): verbose message.
}
if (validationSteps != null) {
doValidation = true;
if (stepsPerEpoch == null) {
throw new ValueError('Can only use `validationSteps` when doing step-wise training, ' +
'i.e., `stepsPerEpoch` must be set.');
}
}
const numTrainSamples = this.checkNumSamples(ins, batchSize, stepsPerEpoch, 'steps_per_epoch');
let indexArray;
if (numTrainSamples != null) {
indexArray = range(0, numTrainSamples);
}
if (verbose == null) {
verbose = 1;
}
const { callbackList, history } = configureCallbacks(callbacks, verbose, epochs, initialEpoch, numTrainSamples, stepsPerEpoch, batchSize, doValidation, callbackMetrics);
callbackList.setModel(this);
this.history = history;
await callbackList.onTrainBegin();
this.stopTraining_ = false;
// TODO(cais): Take care of callbacks.validation_data as in PyKeras.
// TODO(cais): Pre-convert feeds for performance as in PyKeras.
for (let epoch = initialEpoch; epoch < epochs; ++epoch) {
await callbackList.onEpochBegin(epoch);
const epochLogs = {};
if (stepsPerEpoch != null) {
throw new NotImplementedError('stepsPerEpoch mode is not implemented yet.');
}
else {
if (shuffle$1 === 'batch') {
throw new NotImplementedError('batch shuffling is not implemneted'
+ ' yet');
}
else if (shuffle$1) {
shuffle(indexArray);
}
// Convert the potentially shuffled indices to Tensor1D, to avoid the
// cost of repeated creation of Array1Ds later on.
const epochIndexArray1D = tensor1d(indexArray);
const batches = makeBatches(numTrainSamples, batchSize);
for (let batchIndex = 0; batchIndex < batches.length; ++batchIndex) {
const batchLogs = {};
await callbackList.onBatchBegin(batchIndex, batchLogs);
tidy(() => {
const batchStart = batches[batchIndex][0];
const batchEnd = batches[batchIndex][1];
const batchIds = sliceAlongFirstAxis(epochIndexArray1D, batchStart, batchEnd - batchStart);
batchLogs['batch'] = batchIndex;
batchLogs['size'] = batchEnd - batchStart;
// TODO(cais): In ins, train flag can be a number, instead of an
// Tensor? Do we need to handle this in tfjs-layers?
const insBatch = sliceArraysByIndices(ins, batchIds);
const outs = f(insBatch);
for (let i = 0; i < outLabels.length; ++i) {
const label = outLabels[i];
const out = outs[i];
batchLogs[label] = out;
keep(out);
// TODO(cais): Use scope() to avoid ownership.
}
if (batchIndex === batches.length - 1) { // Last batch.
if (doValidation) {
const valOuts = this.testLoop(valF, valIns, batchSize);
// Porting Notes: In tfjs-layers, valOuts is always an Array.
for (let i = 0; i < outLabels.length; ++i) {
const label = outLabels[i];
const out = valOuts[i];
keep(out);
// TODO(cais): Use scope() to avoid ownership.
epochLogs['val_' + label] = out;
}
}
}
});
await callbackList.onBatchEnd(batchIndex, batchLogs);
disposeTensorsInLogs(batchLogs);
if (this.stopTraining_) {
break;
}
// TODO(cais): return outs as list of Tensor.
}
epochIndexArray1D.dispose();
}
// TODO(cais): Run validation at the end of the epoch.
await callbackList.onEpochEnd(epoch, epochLogs);
if (this.stopTraining_) {
break;
}
}
await callbackList.onTrainEnd();
await this.history.syncData();
return this.history;
}
// TODO(cais): Add code snippet below when it's possible to instantiate
// actual dataset objects.
/**
* Trains the model using a dataset object.
*
* @param dataset A dataset object. Its `iterator()` method is expected
* to generate a dataset iterator object, the `next()` method of which
* is expected to produce data batches for training. The return value
* of the `next()` call ought to contain a boolean `done` field and a
* `value` field. The `value` field is expected to be an array of two
* `tf.Tensor`s or an array of two nested `tf.Tensor` structures. The former
* case is for models with exactly one input and one output (e.g.
* a sequential model). The latter case is for models with multiple
* inputs and/or multiple outputs.
* Of the two items in the array, the first is the input feature(s) and
* the second is the output target(s).
* @param args A `ModelFitDatasetArgs`, containing optional fields.
*
* @return A `History` instance. Its `history` attribute contains all
* information collected during training.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
async fitDataset(dataset, args) {
return fitDataset(this, dataset, args);
}
/**
* Runs a single gradient update on a single batch of data.
*
* This method differs from `fit()` and `fitDataset()` in the following
* regards:
* - It operates on exactly one batch of data.
* - It returns only the loss and metric values, instead of
* returning the batch-by-batch loss and metric values.
* - It doesn't support fine-grained options such as verbosity and
* callbacks.
*
* @param x Input data. It could be one of the following:
* - A `tf.Tensor`, or an Array of `tf.Tensor`s (in case the model has
* multiple inputs).
* - An Object mapping input names to corresponding `tf.Tensor` (if the
* model has named inputs).
* @param y Target data. It could be either a `tf.Tensor` or multiple
* `tf.Tensor`s. It should be consistent with `x`.
* @returns Training loss or losses (in case the model has
* multiple outputs), along with metrics (if any), as numbers.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
async trainOnBatch(x, y) {
// TODO(cais): Support sampleWeight and classWeight.
// TODO(cais): Support Dataset objects.
const standardizeOut = await this.standardizeUserData(x, y);
const inputs = standardizeOut[0];
const targets = standardizeOut[1];
const trainFunction = this.makeTrainFunction();
const losses = trainFunction(inputs.concat(targets));
const lossValues = [];
for (const loss of losses) {
const v = await loss.data();
lossValues.push(v[0]);
}
dispose(losses);
disposeNewTensors(standardizeOut[0], x);
disposeNewTensors(standardizeOut[1], y);
return singletonOrArray(lossValues);
}
/**
* Extract weight values of the model.
*
* @param config: An instance of `io.SaveConfig`, which specifies
* model-saving options such as whether only trainable weights are to be
* saved.
* @returns A `NamedTensorMap` mapping original weight names (i.e.,
* non-uniqueified weight names) to their values.
*/
getNamedWeights(config) {
const namedWeights = [];
const trainableOnly = config != null && config.trainableOnly;
const weights = trainableOnly ? this.trainableWeights : this.weights;
const weightValues = this.getWeights(trainableOnly);
for (let i = 0; i < weights.length; ++i) {
if (trainableOnly && !weights[i].trainable) {
// Optionally skip non-trainable weights.
continue;
}
namedWeights.push({ name: weights[i].originalName, tensor: weightValues[i] });
}
return namedWeights;
}
/**
* Setter used for force stopping of LayersModel.fit() (i.e., training).
*
* Example:
*
* ```js
* const input = tf.input({shape: [10]});
* const output = tf.layers.dense({units: 1}).apply(input);
* const model = tf.model({inputs: [input], outputs: [output]});
* model.compile({loss: 'meanSquaredError', optimizer: 'sgd'});
* const xs = tf.ones([8, 10]);
* const ys = tf.zeros([8, 1]);
*
* const history = await model.fit(xs, ys, {
* epochs: 10,
* callbacks: {
* onEpochEnd: async (epoch, logs) => {
* if (epoch === 2) {
* model.stopTraining = true;
* }
* }
* }
* });
*
* // There should be only 3 values in the loss array, instead of 10
* values,
* // due to the stopping after 3 epochs.
* console.log(history.history.loss);
* ```
*/
set stopTraining(stop) {
this.stopTraining_ = stop;
}
get stopTraining() {
return this.stopTraining_;
}
get optimizer() {
return this.optimizer_;
}
set optimizer(optimizer) {
if (this.optimizer_ !== optimizer) {
this.optimizer_ = optimizer;
this.isOptimizerOwned = false;
}
}
dispose() {
const result = super.dispose();
if (result.refCountAfterDispose === 0 && this.optimizer != null &&
this.isOptimizerOwned) {
const numTensorsBeforeOptmizerDisposal = memory().numTensors;
this.optimizer_.dispose();
result.numDisposedVariables +=
numTensorsBeforeOptmizerDisposal - memory().numTensors;
}
return result;
}
getLossIdentifiers() {
let lossNames;
if (typeof this.loss === 'string') {
lossNames = toSnakeCase(this.loss);
}
else if (Array.isArray(this.loss)) {
for (const loss of this.loss) {
if (typeof loss !== 'string') {
throw new Error('Serialization of non-string loss is not supported.');
}
}
lossNames = this.loss.map(name => toSnakeCase(name));
}
else {
const outputNames = Object.keys(this.loss);
lossNames = {};
const losses = this.loss;
for (const outputName of outputNames) {
if (typeof losses[outputName] === 'string') {
lossNames[outputName] =
toSnakeCase(losses[outputName]);
}
else {
throw new Error('Serialization of non-string loss is not supported.');
}
}
}
return lossNames;
}
getMetricIdentifiers() {
if (typeof this.metrics === 'string' ||
typeof this.metrics === 'function') {
return [toSnakeCase(getLossOrMetricName(this.metrics))];
}
else if (Array.isArray(this.metrics)) {
return this.metrics.map(metric => toSnakeCase(getLossOrMetricName(metric)));
}
else {
const metricsIdentifiers = {};
for (const key in this.metrics) {
metricsIdentifiers[key] =
toSnakeCase(getLossOrMetricName(this.metrics[key]));
}
return metricsIdentifiers;
}
}
getTrainingConfig() {
return {
loss: this.getLossIdentifiers(),
metrics: this.getMetricIdentifiers(),
optimizer_config: {
class_name: this.optimizer.getClassName(),
config: this.optimizer.getConfig()
}
};
// TODO(cais): Add weight_metrics when they are supported.
// TODO(cais): Add sample_weight_mode when it's supported.
// TODO(cais): Add loss_weights when it's supported.
}
loadTrainingConfig(trainingConfig) {
if (trainingConfig.weighted_metrics != null) {
throw new Error('Loading weight_metrics is not supported yet.');
}
if (trainingConfig.loss_weights != null) {
throw new Error('Loading loss_weights is not supported yet.');
}
if (trainingConfig.sample_weight_mode != null) {
throw new Error('Loading sample_weight_mode is not supported yet.');
}
const tsConfig = convertPythonicToTs(trainingConfig.optimizer_config);
const optimizer = deserialize(tsConfig);
let loss;
if (typeof trainingConfig.loss === 'string') {
loss = toCamelCase(trainingConfig.loss);
}
else if (Array.isArray(trainingConfig.loss)) {
loss = trainingConfig.loss.map(lossEntry => toCamelCase(lossEntry));
}
else if (trainingConfig.loss != null) {
loss = {};
for (const key in trainingConfig.loss) {
loss[key] = toCamelCase(trainingConfig.loss[key]);
}
}
let metrics;
if (Array.isArray(trainingConfig.metrics)) {
metrics = trainingConfig.metrics.map(metric => toCamelCase(metric));
}
else if (trainingConfig.metrics != null) {
metrics = {};
for (const key in trainingConfig.metrics) {
metrics[key] = toCamelCase(trainingConfig.metrics[key]);
}
}
this.compile({ loss, metrics, optimizer });
}
/**
* Save the configuration and/or weights of the LayersModel.
*
* An `IOHandler` is an object that has a `save` method of the proper
* signature defined. The `save` method manages the storing or
* transmission of serialized data ("artifacts") that represent the
* model's topology and weights onto or via a specific medium, such as
* file downloads, local storage, IndexedDB in the web browser and HTTP
* requests to a server. TensorFlow.js provides `IOHandler`
* implementations for a number of frequently used saving mediums, such as
* `tf.io.browserDownloads` and `tf.io.browserLocalStorage`. See `tf.io`
* for more details.
*
* This method also allows you to refer to certain types of `IOHandler`s
* as URL-like string shortcuts, such as 'localstorage://' and
* 'indexeddb://'.
*
* Example 1: Save `model`'s topology and weights to browser [local
* storage](https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage);
* then load it back.
*
* ```js
* const model = tf.sequential(
* {layers: [tf.layers.dense({units: 1, inputShape: [3]})]});
* console.log('Prediction from original model:');
* model.predict(tf.ones([1, 3])).print();
*
* const saveResults = await model.save('localstorage://my-model-1');
*
* const loadedModel = await tf.loadLayersModel('localstorage://my-model-1');
* console.log('Prediction from loaded model:');
* loadedModel.predict(tf.ones([1, 3])).print();
* ```
*
* Example 2. Saving `model`'s topology and weights to browser
* [IndexedDB](https://developer.mozilla.org/en-US/docs/Web/API/IndexedDB_API);
* then load it back.
*
* ```js
* const model = tf.sequential(
* {layers: [tf.layers.dense({units: 1, inputShape: [3]})]});
* console.log('Prediction from original model:');
* model.predict(tf.ones([1, 3])).print();
*
* const saveResults = await model.save('indexeddb://my-model-1');
*
* const loadedModel = await tf.loadLayersModel('indexeddb://my-model-1');
* console.log('Prediction from loaded model:');
* loadedModel.predict(tf.ones([1, 3])).print();
* ```
*
* Example 3. Saving `model`'s topology and weights as two files
* (`my-model-1.json` and `my-model-1.weights.bin`) downloaded from
* browser.
*
* ```js
* const model = tf.sequential(
* {layers: [tf.layers.dense({units: 1, inputShape: [3]})]});
* const saveResults = await model.save('downloads://my-model-1');
* ```
*
* Example 4. Send `model`'s topology and weights to an HTTP server.
* See the documentation of `tf.io.http` for more details
* including specifying request parameters and implementation of the
* server.
*
* ```js
* const model = tf.sequential(
* {layers: [tf.layers.dense({units: 1, inputShape: [3]})]});
* const saveResults = await model.save('http://my-server/model/upload');
* ```
*
* @param handlerOrURL An instance of `IOHandler` or a URL-like,
* scheme-based string shortcut for `IOHandler`.
* @param config Options for saving the model.
* @returns A `Promise` of `SaveResult`, which summarizes the result of
* the saving, such as byte sizes of the saved artifacts for the model's
* topology and weight values.
*
* @doc {heading: 'Models', subheading: 'Classes', ignoreCI: true}
*/
async save(handlerOrURL, config) {
if (typeof handlerOrURL === 'string') {
const handlers = getSaveHandlers(handlerOrURL);
if (handlers.length === 0) {
throw new ValueError(`Cannot find any save handlers for URL '${handlerOrURL}'`);
}
else if (handlers.length > 1) {
throw new ValueError(`Found more than one (${handlers.length}) save handlers for ` +
`URL '${handlerOrURL}'`);
}
handlerOrURL = handlers[0];
}
if (handlerOrURL.save == null) {
throw new ValueError('LayersModel.save() cannot proceed because the IOHandler ' +
'provided does not have the `save` attribute defined.');
}
const weightDataAndSpecs = await encodeWeights(this.getNamedWeights(config));
const returnString = false;
const unusedArg = null;
const modelConfig = this.toJSON(unusedArg, returnString);
const modelArtifacts = {
modelTopology: modelConfig,
format: LAYERS_MODEL_FORMAT_NAME,
generatedBy: `TensorFlow.js tfjs-layers v${version}`,
convertedBy: null,
};
const includeOptimizer = config == null ? false : config.includeOptimizer;
if (includeOptimizer && this.optimizer != null) {
modelArtifacts.trainingConfig = this.getTrainingConfig();
const weightType = 'optimizer';
const { data: optimizerWeightData, specs: optimizerWeightSpecs } = await encodeWeights(await this.optimizer.getWeights(), weightType);
weightDataAndSpecs.specs.push(...optimizerWeightSpecs);
weightDataAndSpecs.data = concatenateArrayBuffers([weightDataAndSpecs.data, optimizerWeightData]);
}
if (this.userDefinedMetadata != null) {
// Check serialized size of user-defined metadata.
const checkSize = true;
checkUserDefinedMetadata(this.userDefinedMetadata, this.name, checkSize);
modelArtifacts.userDefinedMetadata = this.userDefinedMetadata;
}
modelArtifacts.weightData = weightDataAndSpecs.data;
modelArtifacts.weightSpecs = weightDataAndSpecs.specs;
return handlerOrURL.save(modelArtifacts);
}
/**
* Set user-defined metadata.
*
* The set metadata will be serialized together with the topology
* and weights of the model during `save()` calls.
*
* @param setUserDefinedMetadata
*/
setUserDefinedMetadata(userDefinedMetadata) {
checkUserDefinedMetadata(userDefinedMetadata, this.name);
this.userDefinedMetadata = userDefinedMetadata;
}
/**
* Get user-defined metadata.
*
* The metadata is supplied via one of the two routes:
* 1. By calling `setUserDefinedMetadata()`.
* 2. Loaded during model loading (if the model is constructed
* via `tf.loadLayersModel()`.)
*
* If no user-defined metadata is available from either of the
* two routes, this function will return `undefined`.
*/
getUserDefinedMetadata() {
return this.userDefinedMetadata;
}
}
// The class name is 'Model' rather than 'LayersModel' for backwards
// compatibility since this class name shows up in the serialization format.
/** @nocollapse */
LayersModel.className = 'Model';
registerClass(LayersModel);
/**
* A `tf.Functional` is an alias to `tf.LayersModel`.
*
* See also:
* `tf.LayersModel`, `tf.Sequential`, `tf.loadLayersModel`.
*/
/** @doc {heading: 'Models', subheading: 'Classes'} */
class Functional extends LayersModel {
}
Functional.className = 'Functional';
registerClass(Functional);
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* Original source keras/models.py */
/**
* Load a model and optionally its weights, using an IOHandler object.
*
* @param handler The instance of `IOHandler` to be used during the model
* loading.
* @param customObjects Any optional custom objects to be used during model
* loading.
* @param strict Whether the weight loading will be done in strict mode.
* Default: `true`.
*/
async function loadLayersModelFromIOHandler(handler, customObjects, options) {
if (options == null) {
options = {};
}
if (handler.load == null) {
throw new ValueError('Cannot proceed with model loading because the IOHandler provided ' +
'does not have the `load` method implemented.');
}
const artifacts = await handler.load();
let modelTopology = artifacts.modelTopology;
if (modelTopology['model_config'] != null) {
modelTopology = modelTopology['model_config'];
}
const strict = options.strict == null ? true : options.strict;
// If weights are provided and the weight-loading mode is strict, use
// fast weight initialization. This skips costly initializers such as
// 'orthogonal' and saves unnecessary computation in cases where
// the initialized weight values will immediately be overwritten by
// loaded weight values.
const fastWeightInit = artifacts.weightData != null && artifacts.weightSpecs != null && strict;
const model = deserialize(convertPythonicToTs(modelTopology), customObjects, fastWeightInit);
const trainingConfig = artifacts.trainingConfig;
if (trainingConfig != null) {
model.loadTrainingConfig(trainingConfig);
}
if (artifacts.userDefinedMetadata != null) {
model.setUserDefinedMetadata(artifacts.userDefinedMetadata);
}
// If weightData is present, load the weights into the model.
if (artifacts.weightData != null) {
// Loading weights requires weightSpecs.
if (artifacts.weightSpecs == null) {
throw new ValueError('LayersModel artifacts contains weight data, but not weight specs. ' +
'Therefore loading of weights cannot proceed.');
}
const { modelWeights, optimizerWeights } = decodeModelAndOptimizerWeights(artifacts.weightData, artifacts.weightSpecs);
model.loadWeights(modelWeights, strict);
if (model.optimizer != null && optimizerWeights.length > 0) {
await model.optimizer.setWeights(optimizerWeights);
}
// Dispose temporary weight values.
dispose(modelWeights);
dispose(optimizerWeights.map(w => w.tensor));
}
return model;
}
function decodeModelAndOptimizerWeights(weightData, specs) {
const name2Tensor = decodeWeights(weightData, specs);
const modelWeights = {};
const optimizerWeights = [];
specs.forEach(spec => {
if (spec.group === 'optimizer') {
optimizerWeights.push({ name: spec.name, tensor: name2Tensor[spec.name] });
}
else {
modelWeights[spec.name] = name2Tensor[spec.name];
}
});
return { modelWeights, optimizerWeights };
}
/**
* A model with a stack of layers, feeding linearly from one to the next.
*
* `tf.sequential` is a factory function that creates an instance of
* `tf.Sequential`.
*
* ```js
* // Define a model for linear regression.
* const model = tf.sequential();
* model.add(tf.layers.dense({units: 1, inputShape: [1]}));
*
* // Prepare the model for training: Specify the loss and the optimizer.
* model.compile({loss: 'meanSquaredError', optimizer: 'sgd'});
*
* // Generate some synthetic data for training.
* const xs = tf.tensor2d([1, 2, 3, 4], [4, 1]);
* const ys = tf.tensor2d([1, 3, 5, 7], [4, 1]);
*
* // Train the model using the data then do inference on a data point the
* // model hasn't seen:
* await model.fit(xs, ys);
* model.predict(tf.tensor2d([5], [1, 1])).print();
* ```
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
class Sequential extends LayersModel {
constructor(args) {
super({ inputs: [], outputs: [] });
args = args || {};
this.trainable = true;
this.built = false;
// Set model name.
this.name = (args.name != null) ? args.name : getUid('sequential_');
// Add to the model any layers passed to the constructor.
if (args.layers != null) {
for (const layer of args.layers) {
this.add(layer);
}
}
}
// Helper function to Sequential.add Throws if the new output shape will be
// invalid.
checkShape(layer) {
const shape = layer.inboundNodes[0].outputTensors[0].shape;
if (shape.some(x => x < 0)) {
throw new ValueError('Negative dimension size caused by adding layer ' +
`${layer.name} with input shape [` +
`${layer.inboundNodes[0].inputTensors[0].shape}]`);
}
}
/**
* Adds a layer instance on top of the layer stack.
*
* ```js
* const model = tf.sequential();
* model.add(tf.layers.dense({units: 8, inputShape: [1]}));
* model.add(tf.layers.dense({units: 4, activation: 'relu6'}));
* model.add(tf.layers.dense({units: 1, activation: 'relu6'}));
* // Note that the untrained model is random at this point.
* model.predict(tf.randomNormal([10, 1])).print();
* ```
* @param layer Layer instance.
*
* @exception ValueError In case the `layer` argument does not know its
* input shape.
* @exception ValueError In case the `layer` argument has multiple output
* tensors, or is already connected somewhere else (forbidden in
* `Sequential` models).
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
add(layer) {
const isLayerModelInstance = layer instanceof Sequential || layer instanceof LayersModel;
let modelLayer;
if (isLayerModelInstance) {
modelLayer = layer;
if (modelLayer.outputs.length !== 1) {
throw new ValueError('All layers in a Sequential model ' +
'should have a single output tensor. ' +
'For multi-output layers, ' +
'use the functional API.');
}
if (modelLayer.inputs.length !== 1) {
throw new ValueError('All layers in a Sequential model ' +
'should have a single input tensor. ' +
'For multi-input layers, ' +
'use the functional API.');
}
}
if (this.outputs.length === 0) {
// first layer in model: check that it is an input layer
if (layer.inboundNodes.length === 0) {
// create an input layer
if (layer.batchInputShape == null) {
throw new ValueError('The first layer in a Sequential model must ' +
'get an `inputShape` or `batchInputShape` argument.');
}
// Instantiate the input layer.
const x = Input({
batchShape: layer.batchInputShape,
dtype: layer.dtype,
name: layer.name + '_input'
});
// This will build the current layer and create the node connecting
// the current layer to the input layer we just created.
layer.apply(x);
}
if (isLayerModelInstance) {
this.outputs = modelLayer.outputs;
this.inputs = modelLayer.inputs;
}
else {
if (layer.inboundNodes.length !== 1) {
throw new ValueError('A layer added to a Sequential model must not already be ' +
`connected somewhere else. LayersModel received layer ${layer.name} ` +
`which has ${layer.inboundNodes.length} pre-existing inbound ` +
'connections.');
}
if (layer.inboundNodes[0].outputTensors.length !== 1) {
throw new ValueError('All layers in a Sequential model ' +
'should have a single output tensor. ' +
'For multi-output layers, ' +
'use the functional API.');
}
this.checkShape(layer);
this.outputs = [layer.inboundNodes[0].outputTensors[0]];
this.inputs = getSourceInputs(this.outputs[0]);
}
this.inboundNodes = [];
// We create an input node, which we will keep updated
// as we add more layers.
// (This call has side effects.)
// tslint:disable-next-line:no-unused-expression
new Node({
outboundLayer: this,
inboundLayers: [],
nodeIndices: [],
tensorIndices: [],
inputTensors: this.inputs,
outputTensors: this.outputs,
// no model-level masking for now
inputMasks: pyListRepeat(null, this.inputs.length),
outputMasks: [null],
inputShapes: this.inputs.map(x => x.shape),
outputShapes: this.outputs[0].shape
});
}
else {
const outputTensor = layer.apply(this.outputs[0]);
if (Array.isArray(outputTensor)) {
throw new TypeError('All layers in a Sequential model ' +
'should have a single output tensor. ' +
'For multi-output layers, ' +
'use the functional API.');
}
this.checkShape(layer);
this.outputs = [outputTensor];
// update self.inbound_nodes
this.inboundNodes[0].outputTensors = this.outputs;
this.inboundNodes[0].outputShapes = [this.outputs[0].shape];
}
this.layers.push(layer);
this.built = false;
}
/**
* Removes the last layer in the model.
*
* @exception TypeError if there are no layers in the model.
*/
pop() {
if (this.layers.length === 0) {
throw new TypeError('There are no layers in the model.');
}
this.layers.pop();
if (this.layers.length === 0) {
this.outputs = [];
this.inboundNodes = [];
this.outboundNodes = [];
}
else {
const lastLayerIndex = this.layers.length - 1;
this.layers[lastLayerIndex].outboundNodes = [];
this.outputs = [this.layers[lastLayerIndex].output];
// update self.inbound_nodes
this.inboundNodes[0].outputTensors = this.outputs;
this.inboundNodes[0].outputShapes = [this.outputs[0].shape];
}
}
call(inputs, kwargs) {
if (this.model == null) {
this.build();
}
return this.model.call(inputs, kwargs);
}
build(inputShape) {
// Call `getExactlyOneShape` without using its return value,
// to verify that exactly one input shape is provided.
getExactlyOneShape(inputShape);
if (this.inputs.length === 0 || this.outputs.length === 0) {
throw new TypeError('Sequential model cannot be built: model is empty.' +
' Add some layers first.');
}
// actually create the model
this.model = new LayersModel({
inputs: this.inputs,
outputs: this.outputs[0],
name: this.name + '_model'
});
this.model.trainable = this.trainable;
// mirror model attributes
this.supportsMasking = this.model.supportsMasking;
// TODO(michaelterry): Add caches
this.inputLayers = this.model.inputLayers;
this.inputLayersNodeIndices = this.model.inputLayersNodeIndices;
this.inputLayersTensorIndices = this.model.inputLayersTensorIndices;
this.outputLayers = this.model.outputLayers;
this.outputLayersNodeIndices = this.model.outputLayersNodeIndices;
this.outputLayersTensorIndices = this.model.outputLayersTensorIndices;
this.nodesByDepth = this.model.nodesByDepth;
this.containerNodes = this.model.containerNodes;
this.outputNames = this.model.outputNames;
this.inputNames = this.model.inputNames;
// TODO(michaelterry): Add feedInputNames, feedInputs, if needed.
// TODO(michaelterry): Add callbackModel if needed.
this.built = true;
}
countParams() {
if (!this.built) {
this.build();
}
return super.countParams();
}
/**
* Print a text summary of the Sequential model's layers.
*
* The summary includes
* - Name and type of all layers that comprise the model.
* - Output shape(s) of the layers
* - Number of weight parameters of each layer
* - The total number of trainable and non-trainable parameters of the
* model.
*
* ```js
* const model = tf.sequential();
* model.add(
* tf.layers.dense({units: 100, inputShape: [10], activation: 'relu'}));
* model.add(tf.layers.dense({units: 1, activation: 'sigmoid'}));
*
* model.summary();
* ```
*
* @param lineLength Custom line length, in number of characters.
* @param positions Custom widths of each of the columns, as either
* fractions of `lineLength` (e.g., `[0.5, 0.75, 1]`) or absolute number
* of characters (e.g., `[30, 50, 65]`). Each number corresponds to
* right-most (i.e., ending) position of a column.
* @param printFn Custom print function. Can be used to replace the default
* `console.log`. For example, you can use `x => {}` to mute the printed
* messages in the console.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
summary(lineLength, positions, printFn = console.log) {
if (!this.built) {
this.build();
}
super.summary(lineLength, positions, printFn);
}
/**
* Sets the weights of the model.
*
* @param weights Should be a list of Tensors with shapes and types matching
* the output of `model.getWeights()`.
*/
setWeights(weights) {
if (this.model == null) {
this.build();
}
this.model.setWeights(weights);
}
/**
* Returns the loss value & metrics values for the model in test mode.
*
* Loss and metrics are specified during `compile()`, which needs to happen
* before calls to `evaluate()`.
*
* Computation is done in batches.
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [10]})]
* });
* model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
* const result = model.evaluate(tf.ones([8, 10]), tf.ones([8, 1]), {
* batchSize: 4,
* });
* result.print();
* ```
*
* @param x `tf.Tensor` of test data, or an `Array` of `tf.Tensor`s if the
* model has multiple inputs.
* @param y `tf.Tensor` of target data, or an `Array` of `tf.Tensor`s if the
* model has multiple outputs.
* @param args A `ModelEvaluateConfig`, containing optional fields.
*
* @return `Scalar` test loss (if the model has a single output and no
* metrics) or `Array` of `Scalar`s (if the model has multiple outputs
* and/or metrics). The attribute `model.metricsNames`
* will give you the display labels for the scalar outputs.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
evaluate(x, y, args = {}) {
if (!this.built) {
throw new RuntimeError('The model needs to be compiled before being used.');
}
return this.model.evaluate(x, y, args);
}
// TODO(cais): Add code snippet below once real dataset objects are
// available.
/**
* Evaluate model using a dataset object.
*
* Note: Unlike `evaluate()`, this method is asynchronous (`async`).
*
* @param dataset A dataset object. Its `iterator()` method is expected
* to generate a dataset iterator object, the `next()` method of which
* is expected to produce data batches for evaluation. The return value
* of the `next()` call ought to contain a boolean `done` field and a
* `value` field. The `value` field is expected to be an array of two
* `tf.Tensor`s or an array of two nested `tf.Tensor` structures. The former
* case is for models with exactly one input and one output (e.g.
* a sequential model). The latter case is for models with multiple
* inputs and/or multiple outputs. Of the two items in the array, the
* first is the input feature(s) and the second is the output target(s).
* @param args A configuration object for the dataset-based evaluation.
* @returns Loss and metric values as an Array of `Scalar` objects.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
async evaluateDataset(dataset, args) {
if (!this.built) {
throw new RuntimeError('The model needs to be compiled before being used.');
}
return this.model.evaluateDataset(dataset, args);
}
/**
* Generates output predictions for the input samples.
*
* Computation is done in batches.
*
* Note: the "step" mode of predict() is currently not supported.
* This is because the TensorFlow.js core backend is imperative only.
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [10]})]
* });
* model.predict(tf.ones([2, 10])).print();
* ```
*
* @param x The input data, as a Tensor, or an `Array` of `tf.Tensor`s if
* the model has multiple inputs.
* @param conifg A `ModelPredictConfig` object containing optional fields.
*
* @return `tf.Tensor`(s) of predictions.
*
* @exception ValueError In case of mismatch between the provided input data
* and the model's expectations, or in case a stateful model receives a
* number of samples that is not a multiple of the batch size.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
predict(x, args = {}) {
if (this.model == null) {
this.build();
}
return this.model.predict(x, args);
}
/**
* Returns predictions for a single batch of samples.
*
* @param x: Input samples, as a Tensor, or list of Tensors (if the model
* has multiple inputs).
* @return Tensor(s) of predictions
*/
predictOnBatch(x) {
if (this.model == null) {
this.build();
}
return this.model.predictOnBatch(x);
}
/**
* See `LayersModel.compile`.
*
* @param args
*/
compile(args) {
this.build();
this.model.compile(args);
this.optimizer_ = this.model.optimizer;
// tslint:disable-next-line:no-any
this.isOptimizerOwned = this.model.isOptimizerOwned;
this.loss = this.model.loss;
this.metrics = this.model.metrics;
// TODO(cais): Add this.lossWeights, this.sampleWeightMode,
// this.weightedMetrics, this.targets.
this.metricsTensors = this.model.metricsTensors;
this.metricsNames = this.model.metricsNames;
// TODO(cais): Add sampleWeights.
}
get optimizer() {
return this.model == null ? undefined : this.model.optimizer;
}
set optimizer(optimizer) {
this.model.optimizer = optimizer;
}
/**
* Trains the model for a fixed number of epochs (iterations on a dataset).
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [10]})]
* });
* model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
* const history = await model.fit(tf.ones([8, 10]), tf.ones([8, 1]), {
* batchSize: 4,
* epochs: 3
* });
* console.log(history.history.loss[0]);
* ```
*
* @param x `tf.Tensor` of training data, or an array of `tf.Tensor`s if the
* model has multiple inputs. If all inputs in the model are named, you can
* also pass a dictionary mapping input names to `tf.Tensor`s.
* @param y `tf.Tensor` of target (label) data, or an array of `tf.Tensor`s if
* the model has multiple outputs. If all outputs in the model are named, you
* can also pass a dictionary mapping output names to `tf.Tensor`s.
* @param args A `ModelFitConfig`, containing optional fields.
*
* @return A `History` instance. Its `history` attribute contains all
* information collected during training.
*
* @exception ValueError In case of mismatch between the provided input data
* and what the model expects.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
async fit(x, y, args = {}) {
if (!this.built) {
throw new RuntimeError('The model needs to be compiled before ' +
'being used.');
}
return this.model.fit(x, y, args);
}
/**
* Trains the model using a dataset object.
*
* ```js
* const xArray = [
* [1, 1, 1, 1, 1, 1, 1, 1, 1],
* [1, 1, 1, 1, 1, 1, 1, 1, 1],
* [1, 1, 1, 1, 1, 1, 1, 1, 1],
* [1, 1, 1, 1, 1, 1, 1, 1, 1],
* ];
* const yArray = [1, 1, 1, 1];
* // Create a dataset from the JavaScript array.
* const xDataset = tf.data.array(xArray);
* const yDataset = tf.data.array(yArray);
* // Zip combines the `x` and `y` Datasets into a single Dataset, the
* // iterator of which will return an object containing of two tensors,
* // corresponding to `x` and `y`. The call to `batch(4)` will bundle
* // four such samples into a single object, with the same keys now pointing
* // to tensors that hold 4 examples, organized along the batch dimension.
* // The call to `shuffle(4)` causes each iteration through the dataset to
* // happen in a different order. The size of the shuffle window is 4.
* const xyDataset = tf.data.zip({xs: xDataset, ys: yDataset})
* .batch(4)
* .shuffle(4);
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 1, inputShape: [9]})]
* });
* model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
* const history = await model.fitDataset(xyDataset, {
* epochs: 4,
* callbacks: {onEpochEnd: (epoch, logs) => console.log(logs.loss)}
* });
* ```
*
* @param dataset A dataset object. Its `iterator()` method is expected to
* generate a dataset iterator object, the `next()` method of which is
* expected to produce data batches for evaluation. The return value of the
* `next()` call ought to contain a boolean `done` field and a `value`
* field.
*
* The `value` field is expected to be an object of with fields
* `xs` and `ys`, which point to the feature tensor and the target tensor,
* respectively. This case is for models with exactly one input and one
* output (e.g. a sequential model). For example:
* ```js
* {value: {xs: xsTensor, ys: ysTensor}, done: false}
* ```
*
* If the model has multiple inputs, the `xs` field of `value` should
* be an object mapping input names to their respective feature tensors.
* For example:
* ```js
* {
* value: {
* xs: {
* input_1: xsTensor1,
* input_2: xsTensor2
* },
* ys: ysTensor
* },
* done: false
* }
* ```
* If the model has multiple outputs, the `ys` field of `value` should
* be an object mapping output names to their respective target tensors.
* For example:
* ```js
* {
* value: {
* xs: xsTensor,
* ys: {
* output_1: ysTensor1,
* output_2: ysTensor2
* },
* },
* done: false
* }
* ```
* @param args A `ModelFitDatasetArgs`, containing optional fields.
*
* @return A `History` instance. Its `history` attribute contains all
* information collected during training.
*
* @doc {heading: 'Models', subheading: 'Classes', ignoreCI: true}
*/
async fitDataset(dataset, args) {
if (!this.built) {
throw new RuntimeError('The model needs to be compiled before ' +
'being used.');
}
return this.model.fitDataset(dataset, args);
}
/**
* Runs a single gradient update on a single batch of data.
*
* This method differs from `fit()` and `fitDataset()` in the following
* regards:
* - It operates on exactly one batch of data.
* - It returns only the loss and metric values, instead of
* returning the batch-by-batch loss and metric values.
* - It doesn't support fine-grained options such as verbosity and
* callbacks.
*
* @param x Input data. It could be one of the following:
* - A `tf.Tensor`, or an Array of `tf.Tensor`s (in case the model has
* multiple inputs).
* - An Object mapping input names to corresponding `tf.Tensor` (if the
* model has named inputs).
* @param y Target data. It could be either a `tf.Tensor` or multiple
* `tf.Tensor`s. It should be consistent with `x`.
* @returns Training loss or losses (in case the model has
* multiple outputs), along with metrics (if any), as numbers.
*
* @doc {heading: 'Models', subheading: 'Classes'}
*/
async trainOnBatch(x, y) {
return this.model.trainOnBatch(x, y);
}
/* See parent class for JsDoc */
/** @nocollapse */
static fromConfig(cls, config, customObjects = {}, fastWeightInit = false) {
let configArray;
let extraModelConfig = {};
if (config instanceof Array) {
if (!(config[0].className != null) ||
config[0]['className'] === 'Merge') {
throw new ValueError('Legacy serialization format not supported yet.');
}
configArray = config;
}
else {
assert$1(config['layers'] != null, () => `When the config data for a Sequential model is not an Array, ` +
`it must be an Object that contains the 'layers' field.`);
configArray = config['layers'];
delete config['layers'];
extraModelConfig = config;
}
const model = new cls(extraModelConfig);
if (!(model instanceof Sequential)) {
throw new NotImplementedError(`Sequential.fromConfig called on non-Sequential input: ${model}`);
}
for (const conf of configArray) {
const customObjects = undefined;
const layer = deserialize(conf, customObjects, fastWeightInit);
if (fastWeightInit) {
layer.setFastWeightInitDuringBuild(true);
}
model.add(layer);
}
return model;
}
/**
* Setter used for force stopping of LayersModel.fit() (i.e., training).
*
* Example:
*
* ```js
* const model = tf.sequential();
* model.add(tf.layers.dense({units: 1, inputShape: [10]}));
* model.compile({loss: 'meanSquaredError', optimizer: 'sgd'});
* const xs = tf.ones([8, 10]);
* const ys = tf.zeros([8, 1]);
*
* const history = await model.fit(xs, ys, {
* epochs: 10,
* callbacks: {
* onEpochEnd: async (epoch, logs) => {
* if (epoch === 2) {
* model.stopTraining = true;
* }
* }
* }
* });
*
* // There should be only 3 values in the loss array, instead of 10 values,
* // due to the stopping after 3 epochs.
* console.log(history.history.loss);
* ```
*/
set stopTraining(stop) {
// TODO(cais): When refactoring to remove the composition pattern happens,
// remove this method overriding.
if (this.model == null) {
throw new ValueError('Cannot set the stopTraining property of a sequential model before ' +
'it is compiled.');
}
this.model.stopTraining = stop;
}
get stopTraining() {
if (this.model == null) {
throw new ValueError('Cannot get the stopTraining property of a sequential model before ' +
'it is compiled.');
}
return this.model.stopTraining;
}
// TODO(cais): Override get trainableWeights() here
// tslint:disable-next-line:no-any
getConfig() {
// NOTE(cais): We override the return type of getConfig() to `any` here,
// because the `Sequential` class is a special case among `Container`
// subtypes in that its getConfig() method returns an Array (not a
// dict).
const layers = [];
for (const layer of this.layers) {
const dict = {};
dict['className'] = layer.getClassName();
dict['config'] = layer.getConfig();
layers.push(dict);
}
return { name: this.name, layers };
}
}
/** @nocollapse */
Sequential.className = 'Sequential';
registerClass(Sequential);
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Exported functions.
*/
/**
* Creates a `tf.Sequential` model. A sequential model is any model where the
* outputs of one layer are the inputs to the next layer, i.e. the model
* topology is a simple 'stack' of layers, with no branching or skipping.
*
* This means that the first layer passed to a `tf.Sequential` model should have
* a defined input shape. What that means is that it should have received an
* `inputShape` or `batchInputShape` argument, or for some type of layers
* (recurrent, Dense...) an `inputDim` argument.
*
* The key difference between `tf.model` and `tf.sequential` is that
* `tf.sequential` is less generic, supporting only a linear stack of layers.
* `tf.model` is more generic and supports an arbitrary graph (without
* cycles) of layers.
*
* Examples:
*
* ```js
* const model = tf.sequential();
*
* // First layer must have an input shape defined.
* model.add(tf.layers.dense({units: 32, inputShape: [50]}));
* // Afterwards, TF.js does automatic shape inference.
* model.add(tf.layers.dense({units: 4}));
*
* // Inspect the inferred shape of the model's output, which equals
* // `[null, 4]`. The 1st dimension is the undetermined batch dimension; the
* // 2nd is the output size of the model's last layer.
* console.log(JSON.stringify(model.outputs[0].shape));
* ```
*
* It is also possible to specify a batch size (with potentially undetermined
* batch dimension, denoted by "null") for the first layer using the
* `batchInputShape` key. The following example is equivalent to the above:
*
* ```js
* const model = tf.sequential();
*
* // First layer must have a defined input shape
* model.add(tf.layers.dense({units: 32, batchInputShape: [null, 50]}));
* // Afterwards, TF.js does automatic shape inference.
* model.add(tf.layers.dense({units: 4}));
*
* // Inspect the inferred shape of the model's output.
* console.log(JSON.stringify(model.outputs[0].shape));
* ```
*
* You can also use an `Array` of already-constructed `Layer`s to create
* a `tf.Sequential` model:
*
* ```js
* const model = tf.sequential({
* layers: [tf.layers.dense({units: 32, inputShape: [50]}),
* tf.layers.dense({units: 4})]
* });
* console.log(JSON.stringify(model.outputs[0].shape));
* ```
*
* @doc {heading: 'Models', subheading: 'Creation'}
*/
function sequential(config) {
return new Sequential(config);
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
// Layer activation functions
/**
* Base class for Activations.
*
* Special note: due to cross-language compatibility reasons, the
* static readonly className field in this family of classes must be set to
* the initialLowerCamelCase name of the activation.
*/
let Activation$1 = class Activation extends Serializable {
getConfig() {
return {};
}
};
/**
* Exponential linear unit (ELU).
* Reference: https://arxiv.org/abs/1511.07289
*/
class Elu extends Activation$1 {
/**
* Calculate the activation function.
*
* @param x: Input.
* @param alpha: Scaling factor the negative section.
* @return Output of the ELU activation.
*/
apply(x, alpha = 1) {
return elu(x, alpha);
}
}
/** @nocollapse */
Elu.className = 'elu';
registerClass(Elu);
/**
* Scaled Exponential Linear Unit. (Klambauer et al., 2017).
* Reference: Self-Normalizing Neural Networks, https://arxiv.org/abs/1706.02515
* Notes:
* - To be used together with the initialization "lecunNormal".
* - To be used together with the dropout variant "AlphaDropout".
*/
class Selu extends Activation$1 {
apply(x) {
return selu$2(x);
}
}
/** @nocollapse */
Selu.className = 'selu';
registerClass(Selu);
/**
* Rectified linear unit
*/
class Relu extends Activation$1 {
apply(x) {
return relu$2(x);
}
}
/** @nocollapse */
Relu.className = 'relu';
registerClass(Relu);
/**
* Rectified linear unit activation maxing out at 6.0.
*/
class Relu6 extends Activation$1 {
apply(x) {
return tidy(() => minimum$2(6.0, relu$2(x)));
}
}
/** @nocollapse */
Relu6.className = 'relu6';
registerClass(Relu6);
//* Linear activation (no-op) */
class Linear extends Activation$1 {
apply(x) {
return x;
}
}
/** @nocollapse */
Linear.className = 'linear';
registerClass(Linear);
/**
* Sigmoid activation function.
*/
class Sigmoid extends Activation$1 {
apply(x) {
return sigmoid$2(x);
}
}
/** @nocollapse */
Sigmoid.className = 'sigmoid';
registerClass(Sigmoid);
/**
* Segment-wise linear approximation of sigmoid.
*/
class HardSigmoid extends Activation$1 {
apply(x) {
return hardSigmoid(x);
}
}
/** @nocollapse */
HardSigmoid.className = 'hardSigmoid';
registerClass(HardSigmoid);
/**
* Softplus activation function.
*/
class Softplus extends Activation$1 {
apply(x) {
return softplus$2(x);
}
}
/** @nocollapse */
Softplus.className = 'softplus';
registerClass(Softplus);
/**
* Softsign activation function.
*/
class Softsign extends Activation$1 {
apply(x) {
return softsign(x);
}
}
/** @nocollapse */
Softsign.className = 'softsign';
registerClass(Softsign);
/**
* Hyperbolic tangent function.
*/
class Tanh extends Activation$1 {
apply(x) {
return tanh$2(x);
}
}
/** @nocollapse */
Tanh.className = 'tanh';
registerClass(Tanh);
/**
* Softmax activation function
*/
class Softmax extends Activation$1 {
/**
* Calculate the activation function.
*
* @param x Tensor.
* @param axis Integer, axis along which the softmax normalization is applied.
* Invalid if < 2, as softmax across 1 (the batch dimension) is assumed to be
* an error.
*
* @returns a Tensor of the same shape as x
*
* @throws ValueError: In case `dim(x) < 2`.
*/
apply(x, axis = (-1)) {
return softmax$2(x, axis);
}
}
/** @nocollapse */
Softmax.className = 'softmax';
registerClass(Softmax);
/**
* Log softmax activation function
*/
class LogSoftmax extends Activation$1 {
/**
* Calculate the activation function of log softmax:
* log( exp(x_i) / sum(exp(x)) )
*
* @param x Tensor.
* @param axis Integer, axis along which the softmax normalization is applied.
* Invalid if < 2, as softmax across 1 (the batch dimension) is assumed to be
* an error.
*
* @returns a Tensor of the same shape as x
*
* @throws ValueError: In case `dim(x) < 2`.
*/
apply(x, axis = (-1)) {
return logSoftmax(x, axis);
}
}
/** @nocollapse */
LogSoftmax.className = 'logSoftmax';
registerClass(LogSoftmax);
/**
* Gelu activation function
*/
class Gelu extends Activation$1 {
/**
* Calculate the activation function.
*
* @param x Tensor.
* @returns a Tensor of the same shape as x
*/
apply(x) {
return tidy(() => {
return tidy(() => {
const sqrtTwo = Math.sqrt(2);
// Compute Φ(x) using the erf function
const cdf = mul(0.5, add$1(1, erf$2(div$1(x, sqrtTwo))));
// Compute GELU(x) = x * Φ(x)
return mul(x, cdf);
});
});
}
}
/** @nocollapse */
Gelu.className = 'gelu';
registerClass(Gelu);
/**
* GeluNew activation function
*/
class GeluNew extends Activation$1 {
/**
* Calculate the activation function.
*
* @param x Tensor.
* @returns a Tensor of the same shape as x
*/
apply(x) {
return tidy(() => {
return mul(0.5, mul(x, add$1(1, tanh$2(mul(sqrt$2(div$1(2, Math.PI)), add$1(x, mul(0.044715, pow$2(x, 3))))))));
});
}
}
/** @nocollapse */
GeluNew.className = 'gelu_new';
registerClass(GeluNew);
/**
* Mish activation function
*/
class Mish extends Activation$1 {
/**
* Calculate the activation function.
*
* @param x Tensor.
* @returns a Tensor of the same shape as x
*/
apply(x) {
return tidy(() => mul(x, tanh$2(softplus$2(x))));
}
}
/** @nocollapse */
Mish.className = 'mish';
registerClass(Mish);
/**
* Swish activation function
*/
class Swish extends Activation$1 {
/**
* Calculate the activation function.
*
* @param x Tensor.
* @param alpha Scaling factor for the sigmoid function.
* @returns a Tensor of the same shape as x
*/
apply(x, alpha = 1) {
return tidy(() => mul(sigmoid$2(mul(x, alpha)), x));
}
}
/** @nocollapse */
Swish.className = 'swish';
registerClass(Swish);
function serializeActivation(activation) {
return activation.getClassName();
}
function deserializeActivation(config, customObjects = {}) {
return deserializeKerasObject(config, SerializationMap.getMap().classNameMap, customObjects, 'activation');
}
function getActivation(identifier) {
if (identifier == null) {
const config = {};
config['className'] = 'linear';
config['config'] = {};
return deserializeActivation(config);
}
if (typeof identifier === 'string') {
const config = {};
config['className'] = identifier;
config['config'] = {};
return deserializeActivation(config);
}
else if (identifier instanceof Activation$1) {
return identifier;
}
else {
return deserializeActivation(identifier);
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/* original source: keras/regularizers.py */
function assertObjectArgs(args) {
if (args != null && typeof args !== 'object') {
throw new Error(`Argument to L1L2 regularizer's constructor is expected to be an ` +
`object, but received: ${args}`);
}
}
/**
* Regularizer base class.
*/
class Regularizer extends Serializable {
}
class L1L2 extends Regularizer {
constructor(args) {
super();
assertObjectArgs(args);
this.l1 = args == null || args.l1 == null ? 0.01 : args.l1;
this.l2 = args == null || args.l2 == null ? 0.01 : args.l2;
this.hasL1 = this.l1 !== 0;
this.hasL2 = this.l2 !== 0;
}
/**
* Porting note: Renamed from __call__.
* @param x Variable of which to calculate the regularization score.
*/
apply(x) {
return tidy(() => {
let regularization = zeros$1([1]);
if (this.hasL1) {
regularization = add$1(regularization, sum$2(mul(this.l1, abs$2(x))));
}
if (this.hasL2) {
regularization =
add$1(regularization, sum$2(mul(this.l2, square(x))));
}
return reshape$2(regularization, []);
});
}
getConfig() {
return { 'l1': this.l1, 'l2': this.l2 };
}
/** @nocollapse */
static fromConfig(cls, config) {
return new cls({ l1: config['l1'], l2: config['l2'] });
}
}
/** @nocollapse */
L1L2.className = 'L1L2';
registerClass(L1L2);
// Maps the JavaScript-like identifier keys to the corresponding keras symbols.
const REGULARIZER_IDENTIFIER_REGISTRY_SYMBOL_MAP = {
'l1l2': 'L1L2'
};
function serializeRegularizer(constraint) {
return serializeKerasObject(constraint);
}
function deserializeRegularizer(config, customObjects = {}) {
return deserializeKerasObject(config, SerializationMap.getMap().classNameMap, customObjects, 'regularizer');
}
function getRegularizer(identifier) {
if (identifier == null) {
return null;
}
if (typeof identifier === 'string') {
const className = identifier in REGULARIZER_IDENTIFIER_REGISTRY_SYMBOL_MAP ?
REGULARIZER_IDENTIFIER_REGISTRY_SYMBOL_MAP[identifier] :
identifier;
const config = { className, config: {} };
return deserializeRegularizer(config);
}
else if (identifier instanceof Regularizer) {
return identifier;
}
else {
return deserializeRegularizer(identifier);
}
}
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* TensorFlow.js Layers: Basic Layers.
*/
class Dropout extends Layer {
constructor(args) {
super(args);
this.rate = Math.max(Math.min(args.rate, 1), 0);
// So that the scalar doesn't get tidied up between executions.
this.noiseShape = args.noiseShape;
this.seed = args.seed;
this.supportsMasking = true;
}
getNoiseShape(input) {
if (this.noiseShape == null) {
return this.noiseShape;
}
const inputShape = input.shape;
const noiseShape = [];
for (let i = 0; i < this.noiseShape.length; ++i) {
noiseShape.push(this.noiseShape[i] == null ? inputShape[i] : this.noiseShape[i]);
}
return noiseShape;
}
call(inputs, kwargs) {
return tidy(() => {
this.invokeCallHook(inputs, kwargs);
const input = getExactlyOneTensor(inputs);
if (0 < this.rate && this.rate < 1) {
const training = kwargs['training'] == null ? false : kwargs['training'];
const noiseShape = this.getNoiseShape(input);
const output = inTrainPhase(() => dropout$1(input, this.rate, noiseShape, this.seed), () => input, training);
return output;
}
return inputs;
});
}
getConfig() {
const config = {
rate: this.rate,
noiseShape: this.noiseShape,
seed: this.seed,
};
const baseConfig = super.getConfig();
Object.assign(config, baseConfig);
return config;
}
dispose() {
return super.dispose();
}
}
/** @nocollapse */
Dropout.className = 'Dropout';
registerClass(Dropout);
class SpatialDropout1D extends Dropout {
constructor(args) {
super(args);
this.inputSpec = [{ ndim: 3 }];
}
getNoiseShape(input) {
const inputShape = input.shape;
return [inputShape[0], 1, inputShape[2]];
}
}
/** @nocollapse */
SpatialDropout1D.className = 'SpatialDropout1D';
registerClass(SpatialDropout1D);
class Dense extends Layer {
constructor(args) {
super(args);
// Default activation: Linear (none).
this.activation = null;
this.useBias = true;
this.kernel = null;
this.bias = null;
this.DEFAULT_KERNEL_INITIALIZER = 'glorotNormal';
this.DEFAULT_BIAS_INITIALIZER = 'zeros';
if (args.batchInputShape == null && args.inputShape == null &&
args.inputDim != null) {
// This logic is copied from Layer's constructor, since we can't
// do exactly what the Python constructor does for Dense().
let batchSize = null;
if (args.batchSize != null) {
batchSize = args.batchSize;
}
this.batchInputShape = [batchSize, args.inputDim];
}
this.units = args.units;
assertPositiveInteger(this.units, 'units');
this.activation = getActivation(args.activation);
if (args.useBias != null) {
this.useBias = args.useBias;
}
this.kernelInitializer = getInitializer(args.kernelInitializer || this.DEFAULT_KERNEL_INITIALIZER);
this.biasInitializer =
getInitializer(args.biasInitializer || this.DEFAULT_BIAS_INITIALIZER);
this.kernelConstraint = getConstraint(args.kernelConstraint);
this.biasConstraint = getConstraint(args.biasConstraint);
this.kernelRegularizer = getRegularizer(args.kernelRegularizer);
this.biasRegularizer = getRegularizer(args.biasRegularizer);
this.activityRegularizer = getRegularizer(args.activityRegularizer);
this.supportsMasking = true;
this.inputSpec = [{ minNDim: 2 }];
}
build(inputShape) {
inputShape = getExactlyOneShape(inputShape);
const inputLastDim = inputShape[inputShape.length - 1];
if (this.kernel == null) {
this.kernel = this.addWeight('kernel', [inputLastDim, this.units], null, this.kernelInitializer, this.kernelRegularizer, true, this.kernelConstraint);
if (this.useBias) {
this.bias = this.addWeight('bias', [this.units], null, this.biasInitializer, this.biasRegularizer, true, this.biasConstraint);
}
}
this.inputSpec = [{ minNDim: 2, axes: { [-1]: inputLastDim } }];
this.built = true;
}
computeOutputShape(inputShape) {
inputShape = getExactlyOneShape(inputShape);
const outputShape = inputShape.slice();
outputShape[outputShape.length - 1] = this.units;
return outputShape;
}
call(inputs, kwargs) {
return tidy(() => {
this.invokeCallHook(inputs, kwargs);
// Dense layer accepts only a single input.
const input = getExactlyOneTensor(inputs);
const fusedActivationName = mapActivationToFusedKernel(this.activation.getClassName());
let output;
if (fusedActivationName != null) {
output = dot(input, this.kernel.read(), fusedActivationName, this.bias ? this.bias.read() : null);
}
else {
output = dot(input, this.kernel.read());
if (this.bias != null) {
output = biasAdd(output, this.bias.read());
}
if (this.activation != null) {
output = this.activation.apply(output);
}
}
return output;
});
}
getConfig() {
const config = {
units: this.units,
activation: serializeActivation(this.activation),
useBias: this.useBias,
kernelInitializer: serializeInitializer(this.kernelInitializer),
biasInitializer: serializeInitializer(this.biasInitializer),
kernelRegularizer: serializeRegularizer(this.kernelRegularizer),
biasRegularizer: serializeRegularizer(this.biasRegularizer),
activityRegularizer: serializeRegularizer(this.activityRegularizer),
kernelConstraint: serializeConstraint(this.kernelConstraint),
biasConstraint: serializeConstraint(this.biasConstraint)
};
const baseConfig = super.getConfig();
Object.assign(config, baseConfig);
return config;
}
}
/** @nocollapse */
Dense.className = 'Dense';
registerClass(Dense);
class Flatten extends Layer {
constructor(args) {
args = args || {};
super(args);
this.inputSpec = [{ minNDim: 3 }];
this.dataFormat = args.dataFormat;
}
computeOutputShape(inputShape) {
inputShape = getExactlyOneShape(inputShape);
for (const dim of inputShape.slice(1)) {
if (dim == null) {
throw new ValueError(`The shape of the input to "Flatten" is not fully defined ` +
`(got ${inputShape.slice(1)}). Make sure to pass a complete ` +
`"input_shape" or "batch_input_shape" argument to the first ` +
`layer in your model.`);
}
}
return [inputShape[0], arrayProd(inputShape, 1)];
}
call(inputs, kwargs) {
return tidy(() => {
this.invokeCallHook(inputs, kwargs);
let input = getExactlyOneTensor(inputs);
if (this.dataFormat === 'channelsFirst' && input.rank > 1) {
const permutation = [0];
for (let i = 2; i < input.rank; ++i) {
permutation.push(i);
}
permutation.push(1);
input = transpose$2(input, permutation);
}
return batchFlatten(input);
});
}
getConfig() {
const config = {};
if (this.dataFormat != null) {
config['dataFormat'] = this.dataFormat;
}
const baseConfig = super.getConfig();
Object.assign(config, baseConfig);
return config;
}
}
/** @nocollapse */
Flatten.className = 'Flatten';
registerClass(Flatten);
class Activation extends Layer {
constructor(args) {
super(args);
this.supportsMasking = true;
this.activation = getActivation(args.activation);
}
call(inputs, kwargs) {
return tidy(() => {
this.invokeCallHook(inputs, kwargs);
const input = getExactlyOneTensor(inputs);
return this.activation.apply(input);
});
}
getConfig() {
const config = { activation: serializeActivation(this.activation) };
const baseConfig = super.getConfig();
Object.assign(config, baseConfig);
return config;
}
}
/** @nocollapse */
Activation.className = 'Activation';
registerClass(Activation);
class RepeatVector extends Layer {
constructor(args) {
super(args);
this.n = args.n;
this.inputSpec = [{ ndim: 2 }];
}
computeOutputShape(inputShape) {
return [inputShape[0], this.n, inputShape[1]];
}
call(inputs, kwargs) {
return tidy(() => {
inputs = getExactlyOneTensor(inputs);
return repeat(inputs, this.n);
});
}
getConfig() {
const config = {
n: this.n,
};
const baseConfig = super.getConfig();
Object.assign(config, baseConfig);
return config;
}
}
/** @nocollapse */
RepeatVector.className = 'RepeatVector';
registerClass(RepeatVector);
class Reshape extends Layer {
constructor(args) {
super(args);
this.targetShape = args.targetShape;
// Make sure that all unknown dimensions are represented as `null`.
for (let i = 0; i < this.targetShape.length; ++i) {
if (this.isUnknown(this.targetShape[i])) {
this.targetShape[i] = null;
}
}
}
isUnknown(dim) {
return dim < 0 || dim == null;
}
/**
* Finds and replaces a missing dimension in output shape.
*
* This is a near direct port of the internal Numpy function
* `_fix_unknown_dimension` in `numpy/core/src/multiarray/shape.c`.
*
* @param inputShape: Original shape of array begin reshape.
* @param outputShape: Target shape of the array, with at most a single
* `null` or negative number, which indicates an underdetermined dimension
* that should be derived from `inputShape` and the known dimensions of
* `outputShape`.
* @returns: The output shape with `null` replaced with its computed value.
* @throws: ValueError: If `inputShape` and `outputShape` do not match.
*/
fixUnknownDimension(inputShape, outputShape) {
const errorMsg = 'Total size of new array must be unchanged.';
const finalShape = outputShape.slice();
let known = 1;
let unknown = null;
for (let i = 0; i < finalShape.length; ++i) {
const dim = finalShape[i];
if (this.isUnknown(dim)) {
if (unknown === null) {
unknown = i;
}
else {
throw new ValueError('Can only specifiy one unknown dimension.');
}
}
else {
known *= dim;
}
}
const originalSize = arrayProd(inputShape);
if (unknown !== null) {
if (known === 0 || originalSize % known !== 0) {
throw new ValueError(errorMsg);
}
finalShape[unknown] = originalSize / known;
}
else if (originalSize !== known) {
throw new ValueError(errorMsg);
}
return finalShape;
}
computeOutputShape(inputShape) {
let anyUnknownDims = false;
for (let i = 0; i < inputShape.length; ++i) {
if (this.isUnknown(inputShape[i])) {
anyUnknownDims = true;
break;
}
}
if (anyUnknownDims) {
return inputShape.slice(0, 1).concat(this.targetShape);
}
else {
return inputShape.slice(0, 1).concat(this.fixUnknownDimension(inputShape.slice(1), this.targetShape));
}
}
call(inputs, kwargs) {
return tidy(() => {
this.invokeCallHook(inputs, kwargs);
const input = getExactlyOneTensor(inputs);
const inputShape = input.shape;
const outputShape = inputShape.slice(0, 1).concat(this.fixUnknownDimension(inputShape.slice(1), this.targetShape));
return reshape$2(input, outputShape);
});
}
getConfig() {
const config = {
targetShape: this.targetShape,
};
const baseConfig = super.getConfig();
Object.assign(config, baseConfig);
return config;
}
}
/** @nocollapse */
Reshape.className = 'Reshape';
registerClass(Reshape);
class Permute extends Layer {
constructor(args) {
super(args);
if (args.dims == null) {
throw new Error('Required configuration field `dims` is missing during Permute ' +
'constructor call.');
}
if (!Array.isArray(args.dims)) {
throw new Error('Permute constructor requires `dims` to be an Array, but received ' +
`${args.dims} instead.`);
}
// Check the validity of the permutation indices.
const expectedSortedIndices = range(1, args.dims.length + 1);
if (!arraysEqual(args.dims.slice().sort(), expectedSortedIndices)) {
throw new Error('Invalid permutation `dims`: ' + JSON.stringify(args.dims) +
' `dims` must contain consecutive integers starting from 1.');
}
this.dims = args.dims;
this.dimsIncludingBatch = [0].concat(this.dims);
this.inputSpec = [new InputSpec({ ndim: this.dims.length + 1 })];
}
computeOutputShape(inputShape) {
inputShape = getExactlyOneShape(inputShape);
const outputShape = inputShape.slice();
this.dims.forEach((dim, i) => {
outputShape[i + 1] = inputShape[dim];
});
return outputShape;
}
call(inputs, kwargs) {
return transpose$2(getExactlyOneTensor(inputs), this.dimsIncludingBatch);
}
getConfig() {
const config = {
dims: this.dims,
};
const baseConfig = super.getConfig();
Object.assign(config, baseConfig);
return config;
}
}
/** @nocollapse */
Permute.className = 'Permute';
registerClass(Permute);
class Masking extends Layer {
constructor(args) {
super(args == null ? {} : args);
this.supportsMasking = true;
if (args != null) {
this.maskValue = args.maskValue == null ? 0 : args.maskValue;
}
else {
this.maskValue = 0;
}
}
computeOutputShape(inputShape) {
return inputShape;
}
getConfig() {
const baseConfig = super.getConfig();
const config = { maskValue: this.maskValue };
Object.assign(config, baseConfig);
return config;
}
computeMask(inputs, mask) {
const input = getExactlyOneTensor(inputs);
const axis = -1;
return any$2(notEqual$2(input, this.maskValue), axis);
}
call(inputs, kwargs) {
return tidy(() => {
this.invokeCallHook(inputs, kwargs);
const input = getExactlyOneTensor(inputs);
const axis = -1;
const keepDims = true;
const booleanMask = any$2(notEqual$2(input, this.maskValue), axis, keepDims);
const output = mul(input, cast$3(booleanMask, input.dtype));
return output;
});
}
}
/** @nocollapse */
Masking.className = 'Masking';
registerClass(Masking);
/**
* @license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT-style
* license that can be found in the LICENSE file or at
* https://opensource.org/licenses/MIT.
* =============================================================================
*/
/**
* Creates a dense (fully connected) layer.
*
* This layer implements the operation:
* `output = activation(dot(input, kernel) + bias)`
*
* `activation` is the element-wise activation function
* passed as the `activation` argument.
*
* `kernel` is a weights matrix created by the layer.
*
* `bias` is a bias vector created by the layer (only applicable if `useBias`
* is `true`).
*
* **Input shape:**
*
* nD `tf.Tensor` with shape: `(batchSize, ..., inputDim)`.
*
* The most common situation would be
* a 2D input with shape `(batchSize, inputDim)`.
*
* **Output shape:**
*
* nD tensor with shape: `(batchSize, ..., units)`.
*
* For instance, for a 2D input with shape `(batchSize, inputDim)`,
* the output would have shape `(batchSize, units)`.
*
* Note: if the input to the layer has a rank greater than 2, then it is
* flattened prior to the initial dot product with the kernel.
*
* @doc {heading: 'Layers', subheading: 'Basic', namespace: 'layers'}
*/
function dense(args) {
return new Dense(args);
}
/**
* Applies
* [dropout](http://www.cs.toronto.edu/~rsalakhu/papers/srivastava14a.pdf) to
* the input.
*
* Dropout consists in randomly setting a fraction `rate` of input units to 0 at
* each update during training time, which helps prevent overfitting.
*
* @doc {heading: 'Layers', subheading: 'Basic', namespace: 'layers'}
*/
function dropout(args) {
return new Dropout(args);
}
export { LayersModel, PlatformStub, dense, dropout, enableProdMode, env, fromMemory, glorotUniform, loadLayersModelFromIOHandler, sequential, stringToHashBucketFast$2 as stringToHashBucketFast, tensor1d, tensor2d, withSaveHandler };
Reference in a new issue View git blame Copy permalink