Package ai.djl.training.loss

Class AbstractCompositeLoss

java.lang.Object

ai.djl.training.evaluator.Evaluator

ai.djl.training.loss.Loss

ai.djl.training.loss.AbstractCompositeLoss

Direct Known Subclasses:

BertPretrainingLoss, SimpleCompositeLoss, SingleShotDetectionLoss

public abstract class AbstractCompositeLoss
extends Loss

AbstractCompositeLoss is a Loss class that can combine other Losses together to make a larger loss.

The AbstractCompositeLoss is designed to be extended for more complicated composite losses. For simpler use cases, consider using the SimpleCompositeLoss.

Field Summary

Fields

Modifier and Type	Field	Description
protected java.util.List<Loss>	components

Fields inherited from class ai.djl.training.evaluator.Evaluator

totalInstances

Constructor Summary

Constructors

Constructor	Description
AbstractCompositeLoss(java.lang.String name)	Constructs a composite loss with the given name.

Method Summary

Modifier and Type	Method	Description
void	addAccumulator(java.lang.String key)	Adds an accumulator for the results of the evaluation with the given key.
NDArray	evaluate(NDList labels, NDList predictions)	Calculates the evaluation between the labels and the predictions.
float	getAccumulator(java.lang.String key)	Returns the accumulated evaluator value.
java.util.List<Loss>	getComponents()	Returns the component losses that make up the composite loss.
protected abstract ai.djl.util.Pair<NDList,NDList>	inputForComponent(int componentIndex, NDList labels, NDList predictions)	Returns the inputs to computing the loss for a component loss.
void	resetAccumulator(java.lang.String key)	Resets the evaluator value with the given key.
void	updateAccumulator(java.lang.String key, NDList labels, NDList predictions)	Updates the evaluator with the given key based on a NDList of labels and predictions.

Methods inherited from class ai.djl.training.loss.Loss

elasticNetWeightedDecay, elasticNetWeightedDecay, elasticNetWeightedDecay, elasticNetWeightedDecay, hingeLoss, hingeLoss, hingeLoss, l1Loss, l1Loss, l1Loss, l1WeightedDecay, l1WeightedDecay, l1WeightedDecay, l2Loss, l2Loss, l2Loss, l2WeightedDecay, l2WeightedDecay, l2WeightedDecay, maskedSoftmaxCrossEntropyLoss, maskedSoftmaxCrossEntropyLoss, maskedSoftmaxCrossEntropyLoss, quantileL1Loss, quantileL1Loss, sigmoidBinaryCrossEntropyLoss, sigmoidBinaryCrossEntropyLoss, sigmoidBinaryCrossEntropyLoss, softmaxCrossEntropyLoss, softmaxCrossEntropyLoss, softmaxCrossEntropyLoss

Methods inherited from class ai.djl.training.evaluator.Evaluator

checkLabelShapes, checkLabelShapes, getName

Methods inherited from class java.lang.Object

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

Field Detail

components

protected java.util.List<Loss> components

Constructor Detail

AbstractCompositeLoss

public AbstractCompositeLoss(java.lang.String name)

Constructs a composite loss with the given name.

Parameters:

name - the display name of the loss

Method Detail

inputForComponent

protected abstract ai.djl.util.Pair<NDList,NDList> inputForComponent(int componentIndex,
                                                                           NDList labels,
                                                                           NDList predictions)

Returns the inputs to computing the loss for a component loss.

Parameters:

componentIndex - the index of the component loss

labels - the label input to the composite loss

predictions - the predictions input to the composite loss

Returns:

a pair of the (labels, predictions) inputs to the component loss

getComponents

public java.util.List<Loss> getComponents()

Returns the component losses that make up the composite loss.

Returns:

the component losses that make up the composite loss

evaluate

public NDArray evaluate(NDList labels,
                        NDList predictions)

Calculates the evaluation between the labels and the predictions.

Specified by:

evaluate in class Evaluator

Parameters:

labels - the correct values

predictions - the predicted values

Returns:

the evaluation result

addAccumulator

public void addAccumulator(java.lang.String key)

Adds an accumulator for the results of the evaluation with the given key.

Overrides:

addAccumulator in class Loss

Parameters:

key - the key for the new accumulator

updateAccumulator

public void updateAccumulator(java.lang.String key,
                              NDList labels,
                              NDList predictions)

Updates the evaluator with the given key 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.

Overrides:

updateAccumulator in class Loss

Parameters:

key - the key of the accumulator to update

labels - a NDList of labels

predictions - a NDList of predictions

resetAccumulator

public void resetAccumulator(java.lang.String key)

Resets the evaluator value with the given key.

Overrides:

resetAccumulator in class Loss

Parameters:

key - the key of the accumulator to reset

getAccumulator

public float getAccumulator(java.lang.String key)

Returns the accumulated evaluator value.

Overrides:

getAccumulator in class Loss

Parameters:

key - the key of the accumulator to get

Returns:

the accumulated value