Package io.github.kirstenali.deepj.loss

Class CrossEntropyLoss

java.lang.Object

io.github.kirstenali.deepj.loss.CrossEntropyLoss

All Implemented Interfaces:

LossFunction

public final class CrossEntropyLoss
extends Object
implements LossFunction

Cross-entropy loss with integer class targets.

Expected shapes:

• predicted (logits): [nTokens x vocab]
• actual (class indices): [nTokens x 1] where each entry is an integer in [0, vocab)

This class also provides helpers for common language-modeling usage where targets are provided as int[].

Constructor Summary

Constructors

Constructor	Description
CrossEntropyLoss()

Method Summary

Modifier and Type	Method	Description
static Tensor	fromIntTargets(int[] targets)	Builds a [n x 1] Tensor from int[] targets.
static Tensor	gradient(Tensor logits, int[] targets)	Convenience helper: gradient w.r.t.
Tensor	gradient(Tensor predicted, Tensor actual)
static double	loss(Tensor logits, int[] targets)	Convenience helper: compute loss from logits and int targets.
double	loss(Tensor predicted, Tensor actual)
static int[]	toIntTargets(Tensor actual)	Converts a [n x 1] Tensor of class indices into an int[].

Methods inherited from class java.lang.Object

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

Constructor Details

CrossEntropyLoss

public CrossEntropyLoss()

Method Details

loss

public double loss(Tensor predicted,
 Tensor actual)

Specified by:

loss in interface LossFunction

gradient

public Tensor gradient(Tensor predicted,
 Tensor actual)

Specified by:

gradient in interface LossFunction

loss

public static double loss(Tensor logits,
 int[] targets)

Convenience helper: compute loss from logits and int targets.

gradient

public static Tensor gradient(Tensor logits,
 int[] targets)

Convenience helper: gradient w.r.t. logits, averaged over rows.

toIntTargets

public static int[] toIntTargets(Tensor actual)

Converts a [n x 1] Tensor of class indices into an int[].

fromIntTargets

public static Tensor fromIntTargets(int[] targets)

Builds a [n x 1] Tensor from int[] targets.