org.nd4j.linalg.lossfunctions.impl.LossKLD Maven / Gradle / Ivy

package org.nd4j.linalg.lossfunctions.impl;


import lombok.EqualsAndHashCode;
import org.nd4j.linalg.primitives.Pair;
import org.nd4j.linalg.activations.IActivation;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.lossfunctions.ILossFunction;
import org.nd4j.linalg.lossfunctions.LossUtil;
import org.nd4j.linalg.ops.transforms.Transforms;

/**
 * Kullback Leibler Divergence loss function
 *
 * @author Susan Eraly
 */
@EqualsAndHashCode
public class LossKLD implements ILossFunction {

    private INDArray scoreArray(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask) {
        if (labels.size(1) != preOutput.size(1)) {
            throw new IllegalArgumentException("Labels array numColumns (size(1) = " + labels.size(1)
                            + ") does not match output layer" + " number of outputs (nOut = " + preOutput.size(1)
                            + ") ");
            
        }
        INDArray output = activationFn.getActivation(preOutput.dup(), true);

        // Clip output and labels to be between Nd4j.EPS_THREsHOLD and 1, i.e. a valid non-zero probability
        output = Transforms.min(Transforms.max(output, Nd4j.EPS_THRESHOLD, false), 1, false);
        labels = Transforms.min(Transforms.max(labels, Nd4j.EPS_THRESHOLD, true), 1, false);

        INDArray logRatio = Transforms.log(output.rdivi(labels), false);

        INDArray scoreArr = logRatio.muli(labels);
        if (mask != null) {
            LossUtil.applyMask(scoreArr, mask);
        }
        return scoreArr;
    }

    @Override
    public double computeScore(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask,
                    boolean average) {
        INDArray scoreArr = scoreArray(labels, preOutput, activationFn, mask);

        double score = scoreArr.sumNumber().doubleValue();

        if (average) {
            score /= scoreArr.size(0);
        }

        return score;
    }

    @Override
    public INDArray computeScoreArray(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask) {
        INDArray scoreArr = scoreArray(labels, preOutput, activationFn, mask);
        return scoreArr.sum(1);
    }

    @Override
    public INDArray computeGradient(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask) {
        if (labels.size(1) != preOutput.size(1)) {
            throw new IllegalArgumentException("Labels array numColumns (size(1) = " + labels.size(1)
                            + ") does not match output layer" + " number of outputs (nOut = " + preOutput.size(1)
                            + ") ");
            
        }
        //INDArray output = Nd4j.getExecutioner().execAndReturn(Nd4j.getOpFactory().createTransform(activationFn, preOutput.dup()));
        INDArray output = activationFn.getActivation(preOutput.dup(), true);

        INDArray dLda = labels.div(output).negi();

        if(mask != null && LossUtil.isPerOutputMasking(dLda, mask)){
            //For *most* activation functions: we don't actually need to mask dL/da in addition to masking dL/dz later
            //but: some, like softmax, require both (due to dL/dz_i being a function of dL/da_j, for i != j)
            //We could add a special case for softmax (activationFn instanceof ActivationSoftmax) but that would be
            // error prone - though buy us a tiny bit of performance
            LossUtil.applyMask(dLda, mask);
        }

        INDArray grad = activationFn.backprop(preOutput, dLda).getFirst(); //TODO activation functions with params

        if (mask != null) {
            LossUtil.applyMask(grad, mask);
        }

        return grad;
    }

    @Override
    public Pair computeGradientAndScore(INDArray labels, INDArray preOutput, IActivation activationFn,
                    INDArray mask, boolean average) {
        //TODO: probably a more efficient way to do this...

        return new Pair<>(computeScore(labels, preOutput, activationFn, mask, average),
                        computeGradient(labels, preOutput, activationFn, mask));
    }


    /**
     * The name of this function
     *
     * @return
     */
    @Override
    public String name() {
        return toString();
    }



    @Override
    public String toString() {
        return "LossKLD()";
    }
}