org.nd4j.linalg.learning.AdaGrad Maven / Gradle / Ivy
/*
*
* * Copyright 2015 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 lombok.NoArgsConstructor;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.shape.Shape;
import org.nd4j.linalg.factory.Nd4j;
import java.io.Serializable;
import static org.nd4j.linalg.ops.transforms.Transforms.sqrt;
/**
* Vectorized Learning Rate used per Connection Weight
*
* Adapted from: http://xcorr.net/2014/01/23/adagrad-eliminating-learning-rates-in-stochastic-gradient-descent/
*
* @author Adam Gibson
*/
@Data
@NoArgsConstructor
public class AdaGrad implements Serializable,GradientUpdater {
//protected double squaredGradientSum = 0;
public INDArray historicalGradient;
public int[] shape;
protected double learningRate = 1e-1; // learning rate
protected int numIterations = 0;
private double epsilon = 1e-8;
/**
*
* @param rows
* @param cols
* @param learningRate
*/
public AdaGrad(int rows, int cols, double learningRate) {
this.shape = new int[]{rows, cols};
this.learningRate = learningRate;
}
public AdaGrad(int rows, int cols){
this(rows, cols, 0.1);
}
public AdaGrad(int[] shape, double learningRate) {
this.shape = shape;
this.learningRate = learningRate;
}
public AdaGrad(double learningRate) {
this.learningRate = learningRate;
}
@Override
public void update(Object... args) {
if(args.length > 0) {
learningRate = (Double) args[0];
}
}
/**
* Gets feature specific learning rates
* Adagrad keeps a history of gradients being passed in.
* Note that each gradient passed in becomes adapted over time, hence
* the name adagrad
*
* @param gradient the gradient to get learning rates for
* @param iteration
* @return the feature specific learning rates
*/
@Override
public INDArray getGradient(INDArray gradient, int iteration) {
if(historicalGradient == null) historicalGradient = gradient.mul(gradient);
else historicalGradient.addi(gradient.mul(gradient));
INDArray sqrtHistory = sqrt(historicalGradient.add(epsilon));
// lr * gradient / sqrt(sumSquaredGradients + 1e-8)
INDArray ret = sqrtHistory.rdivi(learningRate).muli(gradient);
numIterations++;
return ret;
}
public double getGradient(double gradient, int column, int[] shape) {
boolean historicalInitialized = false;
if (this.historicalGradient == null) {
this.historicalGradient = Nd4j.ones(shape);
historicalInitialized = true;
}
double sqrtHistory = !historicalInitialized ? Math.sqrt(historicalGradient.getDouble(column)) : historicalGradient.getDouble(column);
double learningRates = learningRate / (sqrtHistory + epsilon);
double adjustedGradient = gradient * (learningRates);
historicalGradient.putScalar(column, historicalGradient.getDouble(column) + gradient * gradient);
numIterations++;
//ensure no zeros
return adjustedGradient;
}
public INDArray getGradient(INDArray gradient, int slice, int[] shape) {
boolean historicalInitialized = false;
INDArray sqrtHistory;
if (this.historicalGradient == null) {
this.historicalGradient = Nd4j.ones(shape);
historicalInitialized = true;
} else if (!this.historicalGradient.isVector() && this.historicalGradient.slice(slice).length() != gradient.length())
throw new IllegalArgumentException("Illegal gradient");
if (historicalGradient.isVector())
sqrtHistory = sqrt(historicalGradient);
else
sqrtHistory = !historicalInitialized ? sqrt(historicalGradient.slice(slice)) : historicalGradient;
INDArray learningRates = sqrtHistory.add(epsilon).rdivi(learningRate);
if(gradient.length() != learningRates.length())
gradient.muli(learningRates.slice(slice));
else
gradient.muli(learningRates);
this.historicalGradient.slice(slice).addi(gradient.mul(gradient));
numIterations++;
//ensure no zeros
return gradient;
}
public AdaGrad createSubset(int index) {
if (historicalGradient == null)
this.historicalGradient = Nd4j.ones(shape);
if (Shape.isMatrix(shape)) {
AdaGrad a = new AdaGrad(1, historicalGradient.columns());
//grab only the needed elements
INDArray slice = historicalGradient.slice(index).dup();
a.historicalGradient = slice;
a.setLearningRate(learningRate);
return a;
} else {
AdaGrad a = new AdaGrad(1, 1);
//grab only the needed elements
INDArray slice = Nd4j.scalar(historicalGradient.getDouble(index));
a.historicalGradient = slice;
a.setLearningRate(learningRate);
return a;
}
}
@Override
public GradientUpdaterAggregator getAggregator(boolean addThis){
AdaGradAggregator ag = new AdaGradAggregator();
if(addThis) ag.aggregate(this);
return ag;
}
public static class AdaGradAggregator implements GradientUpdaterAggregator {
private INDArray historicalGradientSum;
private double lrSum;
private long numIterationsSum = 0;
private int count = 0;
@Override
public GradientUpdater getUpdater() {
AdaGrad adaGrad = new AdaGrad(lrSum/count);
adaGrad.setHistoricalGradient(historicalGradientSum.div(count));
adaGrad.setNumIterations((int)(numIterationsSum/count));
return adaGrad;
}
@Override
public void aggregate(GradientUpdater updater) {
if(!(updater instanceof AdaGrad)) throw new UnsupportedOperationException("Cannot aggregate AdaGrad with updater: " + updater);
AdaGrad adagrad = (AdaGrad)updater;
if(historicalGradientSum ==null){
historicalGradientSum = adagrad.historicalGradient.dup();
lrSum = adagrad.learningRate;
numIterationsSum = adagrad.numIterations;
} else {
historicalGradientSum.addi(adagrad.historicalGradient);
lrSum += adagrad.learningRate;
numIterationsSum += adagrad.numIterations;
}
count++;
}
@Override
public GradientUpdaterAggregator combine(GradientUpdaterAggregator other) {
if(!(other instanceof AdaGradAggregator))
throw new IllegalArgumentException("Cannot combine AdaGradAggregator with aggregator: " + other);
AdaGradAggregator aggregator = (AdaGradAggregator)other;
historicalGradientSum.addi(aggregator.historicalGradientSum);
lrSum += aggregator.lrSum;
numIterationsSum += aggregator.numIterationsSum;
count += aggregator.count;
return this;
}
}
}