org.nd4j.linalg.learning.NadamUpdater Maven / Gradle / Ivy
/*-
*
* * Copyright 2017 Skymind,Inc.
* *
* * Licensed under the Apache License, Version 2.0 (the "License");
* * you may not use this file except in compliance with the License.
* * You may obtain a copy of the License at
* *
* * http://www.apache.org/licenses/LICENSE-2.0
* *
* * Unless required by applicable law or agreed to in writing, software
* * distributed under the License is distributed on an "AS IS" BASIS,
* * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* * See the License for the specific language governing permissions and
* * limitations under the License.
*
*
*/
package org.nd4j.linalg.learning;
import lombok.Data;
import org.apache.commons.math3.util.FastMath;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.shape.Shape;
import org.nd4j.linalg.indexing.NDArrayIndex;
import org.nd4j.linalg.learning.config.Nadam;
import org.nd4j.linalg.ops.transforms.Transforms;
/**
* The Nadam updater.
* https://arxiv.org/pdf/1609.04747.pdf
*
* @author Andrey Spiridonov
*/
@Data
public class NadamUpdater implements GradientUpdater {
private Nadam config;
private INDArray m, v; // moving avg & sqrd gradients
private char gradientReshapeOrder;
public NadamUpdater(Nadam config) {
this.config = config;
}
@Override
public void setStateViewArray(INDArray viewArray, int[] gradientShape, char gradientOrder, boolean initialize) {
if (!viewArray.isRowVector())
throw new IllegalArgumentException("Invalid input: expect row vector input");
if (initialize)
viewArray.assign(0);
int length = viewArray.length();
this.m = viewArray.get(NDArrayIndex.point(0), NDArrayIndex.interval(0, length / 2));
this.v = viewArray.get(NDArrayIndex.point(0), NDArrayIndex.interval(length / 2, length));
//Reshape to match the expected shape of the input gradient arrays
this.m = Shape.newShapeNoCopy(this.m, gradientShape, gradientOrder == 'f');
this.v = Shape.newShapeNoCopy(this.v, gradientShape, gradientOrder == 'f');
if (m == null || v == null)
throw new IllegalStateException("Could not correctly reshape gradient view arrays");
this.gradientReshapeOrder = gradientOrder;
}
/**
* Calculate the update based on the given gradient
*
* @param gradient the gradient to get the update for
* @param iteration
* @return the gradient
*/
@Override
public void applyUpdater(INDArray gradient, int iteration, int epoch) {
if (m == null || v == null)
throw new IllegalStateException("Updater has not been initialized with view state");
double beta1 = config.getBeta1();
double beta2 = config.getBeta2();
double learningRate = config.getLearningRate(iteration, epoch);
double epsilon = config.getEpsilon();
INDArray oneMinusBeta1Grad = gradient.mul(1.0 - beta1);
m.muli(beta1).addi(oneMinusBeta1Grad);
INDArray oneMinusBeta2GradSquared = gradient.mul(gradient).muli(1.0 - beta2);
v.muli(beta2).addi(oneMinusBeta2GradSquared);
double beta1t = FastMath.pow(beta1, iteration + 1);
INDArray biasCorrectedEstimateOfMomentum = m.mul(beta1).divi(1.0 - beta1t);
INDArray secondTerm = oneMinusBeta1Grad.divi(1 - beta1t);
INDArray alphat = biasCorrectedEstimateOfMomentum.add(secondTerm).muli(learningRate);
INDArray sqrtV = Transforms.sqrt(v.dup(gradientReshapeOrder), false).addi(epsilon);
gradient.assign(alphat).divi(sqrtV);
}
}