org.nd4j.linalg.lossfunctions.impl.LossMAE Maven / Gradle / Ivy
package org.nd4j.linalg.lossfunctions.impl;
import lombok.EqualsAndHashCode;
import org.nd4j.linalg.activations.IActivation;
import org.nd4j.linalg.api.ndarray.INDArray;
/**
* Mean absolute error loss function: L = 1/N sum_i abs(predicted_i - actual_i)
* See also {@link LossL1} for a mathematically similar loss function (LossL1 does not have division by N, where N is output size)
*
* @author Susan Eraly
*/
@EqualsAndHashCode(callSuper = true)
public class LossMAE extends LossL1 {
public LossMAE(){
}
/**
* Mean Absolute Error loss function 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 LossMAE(INDArray weights){
super(weights);
}
@Override
public double computeScore(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask, boolean average) {
double score = super.computeScore(labels,preOutput,activationFn,mask,average);
score /= (labels.size(1));
return score;
}
@Override
public INDArray computeScoreArray(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask) {
INDArray scoreArr = super.computeScoreArray(labels,preOutput,activationFn,mask);
scoreArr.divi(scoreArr.size(1));
return scoreArr;
}
@Override
public INDArray computeGradient(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask) {
INDArray gradients = super.computeGradient(labels,preOutput,activationFn,mask);
gradients.divi(labels.size(1));
return gradients;
}
@Override
public String toString() {
if(weights == null) return "LossMAE()";
return "LossMAE(weights=" + weights + ")";
}
}