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

Show more of this group Show more artifacts with this name
Show all versions of nd4j-api Show documentation
There is a newer version: 1.0.0-M2.1
package org.nd4j.linalg.lossfunctions.impl;

import lombok.EqualsAndHashCode;
import lombok.Getter;
import org.apache.commons.math3.util.Pair;
import org.nd4j.linalg.activations.IActivation;
import org.nd4j.linalg.activations.impl.ActivationSoftmax;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.ops.impl.transforms.LogSoftMax;
import org.nd4j.linalg.api.ops.impl.transforms.TimesOneMinus;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.lossfunctions.ILossFunction;
import org.nd4j.linalg.lossfunctions.LossUtil;
import org.nd4j.linalg.lossfunctions.serde.RowVectorDeserializer;
import org.nd4j.linalg.lossfunctions.serde.RowVectorSerializer;
import org.nd4j.linalg.ops.transforms.Transforms;
import org.nd4j.shade.jackson.annotation.JsonInclude;
import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
import org.nd4j.shade.jackson.databind.annotation.JsonSerialize;

/**
 * Binary cross entropy loss function
 * 
 * https://en.wikipedia.org/wiki/Cross_entropy#Cross-entropy_error_function_and_logistic_regression
 * Labels are assumed to take values 0 or 1
 *
 * @author Susan Eraly
 */
@EqualsAndHashCode
@JsonInclude(JsonInclude.Include.NON_NULL)
@Getter
public class LossBinaryXENT implements ILossFunction {

    @JsonSerialize(using = RowVectorSerializer.class)
    @JsonDeserialize(using = RowVectorDeserializer.class)
    private final INDArray weights;

    public LossBinaryXENT() {
        this(null);
    }

    /**
     * Binary cross entropy where each the output is
     * (optionally) weighted/scaled by a fixed scalar value.
     * Note that the weights array must be a row vector, of length equal to
     * the labels/output dimension 1 size.
     * A weight vector of 1s should give identical results to no weight vector.
     *
     * @param weights Weights array (row vector). May be null.
     */
    public LossBinaryXENT(INDArray weights) {
        if (weights != null && !weights.isRowVector()) {
            throw new IllegalArgumentException("Weights array must be a row vector");
        }

        this.weights = weights;
    }

    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 scoreArr;
        if (activationFn instanceof ActivationSoftmax) {
            //Use LogSoftMax op to avoid numerical issues when calculating score
            INDArray logsoftmax = Nd4j.getExecutioner().execAndReturn(new LogSoftMax(preOutput.dup()));
            scoreArr = logsoftmax.muli(labels);

        } else {
            //INDArray output = Nd4j.getExecutioner().execAndReturn(Nd4j.getOpFactory().createTransform(activationFn, preOutput.dup()));
            INDArray output = activationFn.getActivation(preOutput.dup(), true);
            scoreArr = Transforms.log(output, true).muli(labels);
            INDArray secondTerm = output.rsubi(1);
            Transforms.log(secondTerm, false);
            secondTerm.muli(labels.rsub(1));
            scoreArr.addi(secondTerm);
        }

        //Weighted loss function
        if (weights != null) {
            if (weights.length() != preOutput.size(1)) {
                throw new IllegalStateException("Weights vector (length " + weights.length()
                        + ") does not match output.size(1)=" + preOutput.size(1));
            }

            scoreArr.muliRowVector(weights);
        }

        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).muli(-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 = activationFn.getActivation(preOutput.dup(), true);

        INDArray numerator = output.sub(labels);
        INDArray denominator = Nd4j.getExecutioner().execAndReturn(new TimesOneMinus(output));  // output * (1-output)
        INDArray dLda = numerator.divi(denominator);

        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 - but buy us a tiny bit of performance
            LossUtil.applyMask(dLda, mask);
        }

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

        //Weighted loss function
        if (weights != null) {
            if (weights.length() != output.size(1)) {
                throw new IllegalStateException("Weights vector (length " + weights.length()
                        + ") does not match output.size(1)=" + output.size(1));
            }

            grad.muliRowVector(weights);
        }

        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() {
        if (weights == null)
            return "LossBinaryXENT()";
        return "LossBinaryXENT(weights=" + weights + ")";
    }
}