org.nd4j.autodiff.loss

Class LossFunctions

java.lang.Object

org.nd4j.autodiff.loss.LossFunctions

public class LossFunctions
extends Object

SameDiff loss functions

Author:

Alex Black

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	LossFunctions.Reduction

Method Summary

Modifier and Type	Method and Description
static LossInfo	l1(String outputName, SDVariable predictions, SDVariable label, SDVariable weights, LossFunctions.Reduction reduction, int... dimensions) L1 loss - sum of absolute errors.
static LossInfo	l2(String outputName, SDVariable predictions, SDVariable label, SDVariable weights, LossFunctions.Reduction reduction, int... dimensions) L2 loss function: i.e., sum of squared errors, L = sum_i (actual_i - predicted)^2
static LossInfo	mcxent(String outputName, SDVariable predictions, SDVariable label, SDVariable weights, LossFunctions.Reduction reduction, int... dimensions) Multi-Class Cross Entropy loss function: L = sum_i actual_i * log( predicted_i )
static LossInfo	mse(String outputName, SDVariable predictions, SDVariable label, SDVariable weights, LossFunctions.Reduction reduction, int... dimensions) Mean squared error: L = mean( (predicted - label)^2)
static LossInfo	negativeLogLikelihood(String outputName, SDVariable predictions, SDVariable label, SDVariable weights, LossFunctions.Reduction reduction, int... dimensions)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Detail

mse

public static LossInfo mse(String outputName,
                           SDVariable predictions,
                           SDVariable label,
                           SDVariable weights,
                           LossFunctions.Reduction reduction,
                           int... dimensions)

Mean squared error: L = mean( (predicted - label)^2)

Parameters:

outputName - Name of the output SDVariable

predictions - Predictions variable

label - Label variable

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')

reduction - Type of reduction to perform for the loss function

dimensions - Dimension(s) to apply the loss function on

Returns:

LossInfo - bean with variables etc for the loss function

l1

public static LossInfo l1(String outputName,
                          SDVariable predictions,
                          SDVariable label,
                          SDVariable weights,
                          LossFunctions.Reduction reduction,
                          int... dimensions)

L1 loss - sum of absolute errors. L = sum_i abs(predicted_i - actual_i)

Parameters:

outputName -

predictions -

label -

weights -

reduction -

dimensions -

Returns:

l2

public static LossInfo l2(String outputName,
                          SDVariable predictions,
                          SDVariable label,
                          SDVariable weights,
                          LossFunctions.Reduction reduction,
                          int... dimensions)

L2 loss function: i.e., sum of squared errors, L = sum_i (actual_i - predicted)^2

Parameters:

outputName -

predictions -

label -

weights -

reduction -

dimensions -

Returns:

negativeLogLikelihood

public static LossInfo negativeLogLikelihood(String outputName,
                                             SDVariable predictions,
                                             SDVariable label,
                                             SDVariable weights,
                                             LossFunctions.Reduction reduction,
                                             int... dimensions)

mcxent

public static LossInfo mcxent(String outputName,
                              SDVariable predictions,
                              SDVariable label,
                              SDVariable weights,
                              LossFunctions.Reduction reduction,
                              int... dimensions)

Multi-Class Cross Entropy loss function:
L = sum_i actual_i * log( predicted_i )

Parameters:

outputName -

predictions -

label -

weights -

reduction -

dimensions -

Returns: