org.nd4j.linalg.lossfunctions.impl.LossMSLE Maven / Gradle / Ivy
The newest version!
/*
* ******************************************************************************
* *
* *
* * This program and the accompanying materials are made available under the
* * terms of the Apache License, Version 2.0 which is available at
* * https://www.apache.org/licenses/LICENSE-2.0.
* *
* * See the NOTICE file distributed with this work for additional
* * information regarding copyright ownership.
* * 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.
* *
* * SPDX-License-Identifier: Apache-2.0
* *****************************************************************************
*/
package org.nd4j.linalg.lossfunctions.impl;
import lombok.EqualsAndHashCode;
import lombok.Getter;
import org.nd4j.common.base.Preconditions;
import org.nd4j.linalg.activations.IActivation;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.lossfunctions.ILossFunction;
import org.nd4j.linalg.lossfunctions.LossUtil;
import org.nd4j.linalg.ops.transforms.Transforms;
import org.nd4j.common.primitives.Pair;
import org.nd4j.serde.jackson.shaded.NDArrayTextDeSerializer;
import org.nd4j.serde.jackson.shaded.NDArrayTextSerializer;
import org.nd4j.shade.jackson.annotation.JsonInclude;
import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
import org.nd4j.shade.jackson.databind.annotation.JsonSerialize;
@EqualsAndHashCode
@JsonInclude(JsonInclude.Include.NON_NULL)
@Getter
public class LossMSLE implements ILossFunction {
@JsonSerialize(using = NDArrayTextSerializer.class)
@JsonDeserialize(using = NDArrayTextDeSerializer.class)
private final INDArray weights;
public LossMSLE() {
this(null);
}
/**
* Mean Squared Logarithmic Error loss function where each the output is (optionally) weighted/scaled by a flags 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 LossMSLE(INDArray weights) {
if (weights != null && !weights.isRowVectorOrScalar()) {
throw new IllegalArgumentException("Weights array must be a row vector");
}
this.weights = weights;
}
public INDArray scoreArray(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask) {
if(!labels.equalShapes(preOutput)){
Preconditions.throwEx("Labels and preOutput must have equal shapes: got shapes %s vs %s", labels.shape(), preOutput.shape());
}
labels = labels.castTo(preOutput.dataType()); //No-op if already correct dtype
INDArray scoreArr;
INDArray output = activationFn.getActivation(preOutput.dup(), true);
scoreArr = Transforms.log(output.addi(1.0).divi(labels.add(1.0)), false);
scoreArr = scoreArr.muli(scoreArr).divi(labels.size(1));
//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));
}
scoreArr.muliRowVector(weights.castTo(scoreArr.dataType()));
}
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(true,1);
}
@Override
public INDArray computeGradient(INDArray labels, INDArray preOutput, IActivation activationFn, INDArray mask) {
if(!labels.equalShapes(preOutput)){
Preconditions.throwEx("Labels and preOutput must have equal shapes: got shapes %s vs %s", labels.shape(), preOutput.shape());
}
labels = labels.castTo(preOutput.dataType()); //No-op if already correct dtype
INDArray output = activationFn.getActivation(preOutput.dup(), true);
INDArray p1 = output.add(1.0);
INDArray dlda = p1.rdiv(2.0 / labels.size(1));
INDArray logRatio = Transforms.log(p1.divi(labels.add(1.0)), false);
dlda.muli(logRatio);
if (weights != null) {
dlda.muliRowVector(weights.castTo(dlda.dataType()));
}
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);
}
//dL/dz
INDArray gradients = activationFn.backprop(preOutput, dlda).getFirst(); //TODO activation functions with weights
if (mask != null) {
LossUtil.applyMask(gradients, mask);
}
return gradients;
}
@Override
public Pair computeGradientAndScore(INDArray labels,
INDArray preOutput, IActivation activationFn, INDArray mask, boolean average) {
//TODO: probably a more efficient way to do this...
//Yes - will implement in round two. Just want to get done now.
return new Pair<>(computeScore(labels, preOutput, activationFn, mask, average),
computeGradient(labels, preOutput, activationFn, mask));
}
/**
* The opName of this function
*
* @return
*/
@Override
public String name() {
return toString();
}
@Override
public String toString() {
if (weights == null)
return "LossMSLE()";
return "LossMSLE(weights=" + weights + ")";
}
}