ai.djl.training.loss

Class Loss

java.lang.Object

ai.djl.training.metrics.TrainingMetric

ai.djl.training.loss.Loss

All Implemented Interfaces:

java.lang.Cloneable

Direct Known Subclasses:

CompositeLoss, HingeLoss, L1Loss, L2Loss, SigmoidBinaryCrossEntropyLoss, SingleShotDetectionLoss, SoftmaxCrossEntropyLoss

public abstract class Loss
extends TrainingMetric

Loss functions or Cost Functions to evaluate model predictions against true labels.

Constructor Summary

Constructors

Constructor and Description
Loss(java.lang.String name) Base class for metric with abstract update methods.

Method Summary

Modifier and Type	Method and Description
Loss	duplicate() Creates and returns a copy of this object.
protected int[]	excludeBatchAxis(NDArray loss, int batchAxis) Gets all axes except the batch axis because loss functions require reduction on all axes except the batch axis.
abstract NDArray	getLoss(NDList label, NDList prediction) Calculates loss between the label and prediction.
float	getValue() Calculates metric values.
static HingeLoss	hingeLoss() Returns a new instance of HingeLoss with default arguments.
static HingeLoss	hingeLoss(int margin, float weight, int batchAxis) Returns a new instance of HingeLoss with the given arguments.
static L1Loss	l1Loss() Returns a new instance of L1Loss with default weight and batch axis.
static L1Loss	l1Loss(float weight, int batchAxis) Returns a new instance of L1Loss with given weight and batch axis.
static L2Loss	l2Loss() Returns a new instance of L2Loss with default weight and batch axis.
static L2Loss	l2Loss(float weight, int batchAxis) Returns a new instance of L2Loss with given weight and batch axis.
void	reset() Resets metric values.
static SigmoidBinaryCrossEntropyLoss	sigmoidBinaryCrossEntropyLoss() Returns a new instance of SigmoidBinaryCrossEntropyLoss with default arguments.
static SigmoidBinaryCrossEntropyLoss	sigmoidBinaryCrossEntropyLoss(float weight, int batchAxis, boolean fromSigmoid) Returns a new instance of SigmoidBinaryCrossEntropyLoss with the given arguments.
static SoftmaxCrossEntropyLoss	softmaxCrossEntropyLoss() Returns a new instance of SoftmaxCrossEntropyLoss with default arguments.
static SoftmaxCrossEntropyLoss	softmaxCrossEntropyLoss(float weight, int batchAxis, int classAxis, boolean sparseLabel, boolean fromLogit) Returns a new instance of SoftmaxCrossEntropyLoss with the given arguments.
void	update(NDList labels, NDList predictions) Updates the training metrics based on a NDList of labels and predictions.

Methods inherited from class ai.djl.training.metrics.TrainingMetric

checkLabelShapes, checkLabelShapes, getName

Methods inherited from class java.lang.Object

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

Constructor Detail

Loss

public Loss(java.lang.String name)

Base class for metric with abstract update methods.

Parameters:

name - The display name of the Loss

Method Detail

getLoss

public abstract NDArray getLoss(NDList label,
                                NDList prediction)

Calculates loss between the label and prediction.

Parameters:

label - the true label

prediction - the predicted label

Returns:

the loss value

l1Loss

public static L1Loss l1Loss()

Returns a new instance of L1Loss with default weight and batch axis.

Returns:

a new instance of L1Loss

l1Loss

public static L1Loss l1Loss(float weight,
                            int batchAxis)

Returns a new instance of L1Loss with given weight and batch axis.

Parameters:

weight - the weight to apply on loss value, default 1

batchAxis - the axis that represents mini-batch, default 0

Returns:

a new instance of L1Loss

l2Loss

public static L2Loss l2Loss()

Returns a new instance of L2Loss with default weight and batch axis.

Returns:

a new instance of L2Loss

l2Loss

public static L2Loss l2Loss(float weight,
                            int batchAxis)

Returns a new instance of L2Loss with given weight and batch axis.

Parameters:

weight - the weight to apply on loss value, default 1

batchAxis - the axis that represents mini-batch, default 0

Returns:

a new instance of L2Loss

sigmoidBinaryCrossEntropyLoss

public static SigmoidBinaryCrossEntropyLoss sigmoidBinaryCrossEntropyLoss()

Returns a new instance of SigmoidBinaryCrossEntropyLoss with default arguments.

Returns:

a new instance of SigmoidBinaryCrossEntropyLoss

sigmoidBinaryCrossEntropyLoss

public static SigmoidBinaryCrossEntropyLoss sigmoidBinaryCrossEntropyLoss(float weight,
                                                                          int batchAxis,
                                                                          boolean fromSigmoid)

Returns a new instance of SigmoidBinaryCrossEntropyLoss with the given arguments.

Parameters:

weight - the weight to apply on the loss value, default 1

batchAxis - the axis that represents the mini-batch, default 0

fromSigmoid - whether the input is from the output of sigmoid, default false

Returns:

a new instance of SigmoidBinaryCrossEntropyLoss

softmaxCrossEntropyLoss

public static SoftmaxCrossEntropyLoss softmaxCrossEntropyLoss()

Returns a new instance of SoftmaxCrossEntropyLoss with default arguments.

Returns:

a new instance of SoftmaxCrossEntropyLoss

softmaxCrossEntropyLoss

public static SoftmaxCrossEntropyLoss softmaxCrossEntropyLoss(float weight,
                                                              int batchAxis,
                                                              int classAxis,
                                                              boolean sparseLabel,
                                                              boolean fromLogit)

Returns a new instance of SoftmaxCrossEntropyLoss with the given arguments.

Parameters:

weight - the weight to apply on the loss value, default 1

batchAxis - the axis that represents the mini-batch, default 0

classAxis - the axis that represents the class probabilities, default -1

sparseLabel - whether labels are integer array or probabilities, default true

fromLogit - whether labels are log probabilities or un-normalized numbers

Returns:

a new instance of SoftmaxCrossEntropyLoss

hingeLoss

public static HingeLoss hingeLoss()

Returns a new instance of HingeLoss with default arguments.

Returns:

a new instance of HingeLoss

hingeLoss

public static HingeLoss hingeLoss(int margin,
                                  float weight,
                                  int batchAxis)

Returns a new instance of HingeLoss with the given arguments.

Parameters:

margin - the margin in hinge loss. Defaults to 1.0

weight - the weight to apply on loss value, default 1

batchAxis - the axis that represents mini-batch, default 0

Returns:

a new instance of HingeLoss

duplicate

public Loss duplicate()

Creates and returns a copy of this object.

Overrides:

duplicate in class TrainingMetric

Returns:

a copy of this TrainingMetric

update

public void update(NDList labels,
                   NDList predictions)

Updates the training metrics based on a NDList of labels and predictions.

This is a synchronized operation. You should only call it at the end of a batch or epoch.

Specified by:

update in class TrainingMetric

Parameters:

labels - a NDList of labels

predictions - a NDList of predictions

reset

public void reset()

Resets metric values.

Specified by:

reset in class TrainingMetric

getValue

public float getValue()

Calculates metric values.

Specified by:

getValue in class TrainingMetric

Returns:

a Pair of metric name and value

excludeBatchAxis

protected int[] excludeBatchAxis(NDArray loss,
                                 int batchAxis)

Gets all axes except the batch axis because loss functions require reduction on all axes except the batch axis.

Parameters:

loss - the loss NDArray

batchAxis - the axis that represents the mini-batch

Returns:

all axes except the batch axis