@tensorflow/tfjs-core/dist/optimizers/momentum_optimizer_test.js

Hardware-accelerated JavaScript library for machine intelligence

/**
 * @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 * as tf from '../index';
import { ALL_ENVS, describeWithFlags } from '../jasmine_util';
import { expectArraysClose } from '../test_util';
describeWithFlags('MomentumOptimizer', ALL_ENVS, () => {
    it('basic', async () => {
        const learningRate = .1;
        const momentum = .5;
        const optimizer = tf.train.momentum(learningRate, momentum);
        const x = tf.tensor1d([1, 2]).variable();
        const f = () => x.square().sum();
        let numTensors = tf.memory().numTensors;
        let cost = optimizer.minimize(f, /* returnCost */ true);
        // Cost & velocity should be the only additional arrays.
        expect(tf.memory().numTensors).toBe(numTensors + 2);
        // newAccumulation = momentum * accumulation + gradient
        // newVariable += -learningRate * newAccumulation + variable
        //
        // de/dx = [2, 4]
        // newAccumulation = [2, 4]
        // x = [.8, 1.6]
        expectArraysClose(await x.data(), [.8, 1.6]);
        cost.dispose();
        numTensors = tf.memory().numTensors;
        cost = optimizer.minimize(f, /* returnCost */ false);
        // de/dx = [1.6, 3.2]
        // accumulation = [2, 4]
        // newAccumulation = [2.6, 5.2]
        // x = [0.54, 1.08]
        expectArraysClose(await x.data(), [0.54, 1.08]);
        // There should be no new additional Tensors.
        expect(tf.memory().numTensors).toBe(numTensors);
        expect(cost).toBe(null);
        x.dispose();
        numTensors = tf.memory().numTensors;
        optimizer.dispose();
        // The optimizer.dispose() call should have disposed the m variable and the
        // momentum variable for x.
        expect(tf.memory().numTensors).toBe(numTensors - 2);
    });
    it('basic - with Nesterov', async () => {
        const learningRate = .1;
        const momentum = .5;
        const useNesterov = true;
        const optimizer = tf.train.momentum(learningRate, momentum, useNesterov);
        const x = tf.tensor1d([1, 2]).variable();
        const f = () => x.square().sum();
        let numTensors = tf.memory().numTensors;
        let cost = optimizer.minimize(f, /* returnCost */ true);
        // Cost and velocity should be the only additional arrays.
        expect(tf.memory().numTensors).toBe(numTensors + 2);
        // newAccumulation = momentum * accumulation + gradient
        // newVariable = -learningRate * (newAccumulation * momentum + gradient) +
        //                variable
        //
        // de/dx = [2, 4]
        // newAccumulation = [2, 4]
        // newVariable = -0.1 * ([2, 4] * 0.5 + [2, 4]) + [1, 2]
        // x = [.7, 1.4]
        expectArraysClose(await x.data(), [.7, 1.4]);
        cost.dispose();
        numTensors = tf.memory().numTensors;
        cost = optimizer.minimize(f, /* returnCost */ false);
        // de/dx = [1.4, 2.8]
        // accumulation = [2, 4]
        // newAccumulation = [0.5 * 2 + 1.4, 0.5 * 4 + 2.8] = [2.4, 4.8]
        // newVariable = -0.1 * ([2.4, 4.8] * 0.5 + [1.4, 2.8]) + [0.7, 1.4]
        // x = [0.44, 0.88]
        expectArraysClose(await x.data(), [0.44, 0.88]);
        // There should be no new additional Tensors.
        expect(tf.memory().numTensors).toBe(numTensors);
        expect(cost).toBe(null);
        x.dispose();
        numTensors = tf.memory().numTensors;
        optimizer.dispose();
        // The optimizer.dispose() call should have disposed the m variable and the
        // momentum variable for x.
        expect(tf.memory().numTensors).toBe(numTensors - 2);
    });
    it('Save, load weights and conntinue training', async () => {
        const learningRate = .1;
        const momentum = .5;
        const useNesterov = true;
        const optimizer1 = tf.train.momentum(learningRate, momentum, useNesterov);
        const x = tf.tensor1d([1, 2]).variable();
        const f = () => x.square().sum();
        let cost = optimizer1.minimize(f, /* returnCost */ true);
        // The iterations counter and the accumulation for the variable x.
        const optimizer2 = tf.train.momentum(learningRate, momentum, useNesterov);
        await optimizer2.setWeights(await optimizer1.getWeights());
        cost = optimizer2.minimize(f, /* returnCost */ true);
        expectArraysClose(await cost.data(), 2.45);
        expectArraysClose(await x.data(), [0.44, 0.88]);
        expect(optimizer2.iterations).toEqual(2);
    });
    it('serialization round-trip', () => {
        const originalOpt = tf.train.momentum(0.1, 0.2, true);
        const reserialized = tf.MomentumOptimizer.fromConfig(tf.MomentumOptimizer, originalOpt.getConfig());
        expect(reserialized.getConfig()).toEqual(originalOpt.getConfig());
    });
});
//# sourceMappingURL=momentum_optimizer_test.js.map