UNPKG

@tensorflow/tfjs-core

Version:

Hardware-accelerated JavaScript library for machine intelligence

108 lines 4.19 kB
/** * @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. * ============================================================================= */ import { ENGINE } from '../engine'; import { dispose, tidy } from '../globals'; import { add } from '../ops/add'; import { mul } from '../ops/mul'; import { scalar } from '../ops/scalar'; import { zerosLike } from '../ops/zeros_like'; import { registerClass } from '../serialization'; import { SGDOptimizer } from './sgd_optimizer'; /** @doclink Optimizer */ export class MomentumOptimizer extends SGDOptimizer { 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(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(mul(this.m, accumulation), gradient); if (this.useNesterov) { newValue = add(mul(this.c, add(gradient, mul(newAccumulation, this.m))), value); } else { newValue = add(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']); } } /** @nocollapse */ MomentumOptimizer.className = 'Momentum'; // Name matters for Python compatibility. registerClass(MomentumOptimizer); //# sourceMappingURL=momentum_optimizer.js.map