org.deeplearning4j.optimize.api

Interface TrainingListener

All Known Implementing Classes:

BaseTrainingListener, CheckpointListener, CollectScoresIterationListener, CollectScoresListener, ComposableIterationListener, EvaluativeListener, IterationListener, ParamAndGradientIterationListener, PerformanceListener, ScoreIterationListener, SleepyTrainingListener, TimeIterationListener

public interface TrainingListener

A listener interface for training DL4J models.
The methods here will be called at various points during training, and only during training.
Note that users can extend BaseTrainingListener and selectively override the required methods, instead of implementing TrainingListener directly and having a number of no-op methods.

Author:

Alex Black

Method Summary

Modifier and Type	Method and Description
void	iterationDone(Model model, int iteration, int epoch) Event listener for each iteration.
void	onBackwardPass(Model model) Called once per iteration (backward pass) after gradients have been calculated, and updated Gradients are available via Model.gradient().
void	onEpochEnd(Model model) Called once at the end of each epoch, when using methods such as MultiLayerNetwork.fit(DataSetIterator), ComputationGraph.fit(DataSetIterator) or ComputationGraph.fit(MultiDataSetIterator)
void	onEpochStart(Model model) Called once at the start of each epoch, when using methods such as MultiLayerNetwork.fit(DataSetIterator), ComputationGraph.fit(DataSetIterator) or ComputationGraph.fit(MultiDataSetIterator)
void	onForwardPass(Model model, List<org.nd4j.linalg.api.ndarray.INDArray> activations) Called once per iteration (forward pass) for activations (usually for a MultiLayerNetwork), only at training time
void	onForwardPass(Model model, Map<String,org.nd4j.linalg.api.ndarray.INDArray> activations) Called once per iteration (forward pass) for activations (usually for a ComputationGraph), only at training time
void	onGradientCalculation(Model model) Called once per iteration (backward pass) before the gradients are updated Gradients are available via Model.gradient().

Method Detail

iterationDone

void iterationDone(Model model,
                   int iteration,
                   int epoch)

Event listener for each iteration. Called once, after each parameter update has ocurred while training the network

Parameters:

iteration - the iteration

model - the model iterating

onEpochStart

void onEpochStart(Model model)

Called once at the start of each epoch, when using methods such as MultiLayerNetwork.fit(DataSetIterator), ComputationGraph.fit(DataSetIterator) or ComputationGraph.fit(MultiDataSetIterator)

onEpochEnd

void onEpochEnd(Model model)

Called once at the end of each epoch, when using methods such as MultiLayerNetwork.fit(DataSetIterator), ComputationGraph.fit(DataSetIterator) or ComputationGraph.fit(MultiDataSetIterator)

onForwardPass

void onForwardPass(Model model,
                   List<org.nd4j.linalg.api.ndarray.INDArray> activations)

Called once per iteration (forward pass) for activations (usually for a MultiLayerNetwork), only at training time

Parameters:

model - Model

activations - Layer activations (including input)

onForwardPass

void onForwardPass(Model model,
                   Map<String,org.nd4j.linalg.api.ndarray.INDArray> activations)

Called once per iteration (forward pass) for activations (usually for a ComputationGraph), only at training time

Parameters:

model - Model

activations - Layer activations (including input)

onGradientCalculation

void onGradientCalculation(Model model)

Called once per iteration (backward pass) before the gradients are updated Gradients are available via Model.gradient(). Note that gradients will likely be updated in-place - thus they should be copied or processed synchronously in this method.

For updates (gradients post learning rate/momentum/rmsprop etc) see onBackwardPass(Model)

Parameters:

model - Model

onBackwardPass

void onBackwardPass(Model model)

Called once per iteration (backward pass) after gradients have been calculated, and updated Gradients are available via Model.gradient().

Unlike onGradientCalculation(Model) the gradients at this point will be post-update, rather than raw (pre-update) gradients at that method call.

Parameters:

model - Model