Many resources are needed to download a project. Please understand that we have to compensate our server costs. Thank you in advance.
Project price only 1 $
You can buy this project and download/modify it how often you want.
org.nd4j.autodiff.loss.LossFunctions Maven / Gradle / Ivy
/*******************************************************************************
* Copyright (c) 2015-2018 Skymind, Inc.
*
* 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.
*
* 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.autodiff.loss;
import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.base.Preconditions;
/**
* SameDiff loss functions
*
* @author Alex Black
*/
public class LossFunctions {
private static final int[] SCALAR = new int[]{1,1};
public enum Reduction {
/**
* No reduction. Output is the same shape as the predictions/labels.
* Weights (if any) are applied. Dimension args are ignored.
* Example: 2d input, MSE.
* Output: sqDiff(predictions,labels) -> shape same as input/labels
*/
NONE,
/**
* Reduce as normal along the specified dimensions, but don't sum/mean etc the remaining
* dimensions.
* Example: 2d input, MSE loss along dimension 1.
* Output: mean(weights * sqDiff(predictions,labels),1) -> shape [dim0,1]
*/
SPECIFIED_DIMS,
/**
* Sum across the remaining dimensions, returning a scalar
* Example: 2d input, MSE loss along dimension 1.
* Output: mse_per_ex = mean(weights * sqDiff(predictions,labels),1) *Same as SPECIFIED_DIMS*
* output = sum(mse_per_ex)
*/
SUM,
/**
* Weighted mean: sum(weights * loss) / sum(weights)
* Example: 2d input, MSE loss along dimension 1.
* Output: mse_per_ex = mean(weights * sqDiff(predictions,labels),1) *Same as SPECIFIED_DIMS*
* output = sum(mse_per_ex) / sum(weights)
*
* NOTE: if weights array is not provided, then weights default to (effectively) 1.0 for all entries - and hence
* MEAN_BY_WEIGHT is equivalent to SUM (as sum(1.0) = 1.0)
*/
MEAN_BY_WEIGHT,
/**
* Weighted mean: sum(weights * loss) / count(weights != 0)
* Example: 2d input, MSE loss along dimension 1.
* Output: mse_per_ex = mean(weights * sqDiff(predictions,labels),1) *Same as SPECIFIED_DIMS*
* output = sum(mse_per_ex) / count(weights != 0)
*
* NOTE: if weights array is not provided, then weights default to scalar 1.0 and hence MEAN_BY_COUNT
* is equivalent to
*/
MEAN_BY_COUNT
}
private LossFunctions(){ }
private static LossInfo.Builder validate(String lossName, SDVariable predictions, SDVariable label, Reduction reduction){
Preconditions.checkNotNull(predictions, "Predictions variable cannot be null for loss function - %s", lossName);
Preconditions.checkNotNull(label, "Label variable cannot be null for loss function - %s", lossName);
Preconditions.checkNotNull(reduction, "Reduction enumeration cannot be null for loss function - %s", lossName);
return LossInfo.builder()
.lossName(lossName)
.reduction(reduction)
.label(label)
.predictions(predictions);
}
/**
* Mean squared error: L = mean( (predicted - label)^2)
*
* @param outputName Name of the output SDVariable
* @param predictions Predictions variable
* @param label Label variable
* @param weights Weights array. May be null, or any broadcastable shape (with predictions/label arrays).
* Note that this is also used for masking (weight of 0 = 'masked out')
* @param reduction Type of reduction to perform for the loss function
* @param dimensions Dimension(s) to apply the loss function on
* @return LossInfo - bean with variables etc for the loss function
*/
public static LossInfo mse(String outputName, SDVariable predictions, SDVariable label, SDVariable weights,
Reduction reduction, int... dimensions){
LossInfo.Builder b = validate("mse", predictions, label, reduction);
SameDiff sd = predictions.getSameDiff();
if(weights == null){
weights = sd.one("mse_loss_weights", SCALAR);
}
SDVariable diff = predictions.sub(label);
String name = (reduction == Reduction.NONE ? outputName : null);
SDVariable preReduceLoss = sd.square(diff).mul(name, weights);
return doReduce(sd, outputName, true, b, reduction, preReduceLoss, label, weights, dimensions);
}
/**
* L1 loss - sum of absolute errors. L = sum_i abs(predicted_i - actual_i)
*
* @param outputName
* @param predictions
* @param label
* @param weights
* @param reduction
* @param dimensions
* @return
*/
public static LossInfo l1(String outputName, SDVariable predictions, SDVariable label, SDVariable weights,
Reduction reduction, int... dimensions){
LossInfo.Builder b = validate("l1", predictions, label, reduction);
SameDiff sd = predictions.getSameDiff();
if(weights == null){
weights = sd.one("l1_loss_weights", SCALAR);
}
String name = (reduction == Reduction.NONE ? outputName : null);
SDVariable preReduceLoss = sd.abs(predictions.sub(label)).mul(name, weights);
return doReduce(sd, outputName, false, b, reduction, preReduceLoss, label, weights, dimensions);
}
/**
* L2 loss function: i.e., sum of squared errors, L = sum_i (actual_i - predicted)^2
*
* @param outputName
* @param predictions
* @param label
* @param weights
* @param reduction
* @param dimensions
* @return
*/
public static LossInfo l2(String outputName, SDVariable predictions, SDVariable label, SDVariable weights,
Reduction reduction, int... dimensions){
LossInfo.Builder b = validate("l2", predictions, label, reduction);
SameDiff sd = predictions.getSameDiff();
if(weights == null){
weights = sd.one("l2_loss_weights", SCALAR);
}
SDVariable diff = predictions.sub(label);
String name = (reduction == Reduction.NONE ? outputName : null);
SDVariable preReduceLoss = sd.square(diff).mul(name, weights);
return doReduce(sd, outputName, false, b, reduction, preReduceLoss, label, weights, dimensions);
}
public static LossInfo negativeLogLikelihood(String outputName, SDVariable predictions, SDVariable label, SDVariable weights,
Reduction reduction, int... dimensions){
return mcxent(outputName, predictions, label, weights, reduction, dimensions);
}
/**
* Multi-Class Cross Entropy loss function:
