org.deeplearning4j.spark.api

Interface TrainingWorker<R extends TrainingResult>

All Superinterfaces:

java.io.Serializable

All Known Implementing Classes:

ParameterAveragingTrainingWorker

public interface TrainingWorker<R extends TrainingResult>
extends java.io.Serializable

TrainingWorker is a small serializable class that can be passed (in serialized form) to each Spark executor for actually conducting training. The results are then passed back to the TrainingMaster for processing.

TrainingWorker implementations provide a layer of abstraction for network learning tha should allow for more flexibility/ control over how learning is conducted (including for example asynchronous communication)

Method Summary

Modifier and Type	Method and Description
void	addHook(TrainingHook trainingHook) Add a training hook to be used during training of the worker
WorkerConfiguration	getDataConfiguration() Get the WorkerConfiguration that contains information such as minibatch sizes, etc
R	getFinalResult(ComputationGraph graph) Get the final result to be returned to the driver
R	getFinalResult(MultiLayerNetwork network) Get the final result to be returned to the driver
R	getFinalResultNoData() Get the final result to be returned to the driver, if no data was available for this executor
Pair<R,SparkTrainingStats>	getFinalResultNoDataWithStats() As per getFinalResultNoData() but used when SparkTrainingStats are being collected
Pair<R,SparkTrainingStats>	getFinalResultWithStats(ComputationGraph graph) As per getFinalResult(ComputationGraph) but used when SparkTrainingStats are being collected
Pair<R,SparkTrainingStats>	getFinalResultWithStats(MultiLayerNetwork network) As per getFinalResult(MultiLayerNetwork) but used when SparkTrainingStats are being collected
MultiLayerNetwork	getInitialModel() Get the initial model when training a MultiLayerNetwork/SparkDl4jMultiLayer
ComputationGraph	getInitialModelGraph() Get the initial model when training a ComputationGraph/SparkComputationGraph
R	processMinibatch(org.nd4j.linalg.dataset.api.DataSet dataSet, ComputationGraph graph, boolean isLast) Process (fit) a minibatch for a ComputationGraph
R	processMinibatch(org.nd4j.linalg.dataset.api.DataSet dataSet, MultiLayerNetwork network, boolean isLast) Process (fit) a minibatch for a MultiLayerNetwork
R	processMinibatch(org.nd4j.linalg.dataset.api.MultiDataSet dataSet, ComputationGraph graph, boolean isLast) Process (fit) a minibatch for a ComputationGraph using a MultiDataSet
Pair<R,SparkTrainingStats>	processMinibatchWithStats(org.nd4j.linalg.dataset.api.DataSet dataSet, ComputationGraph graph, boolean isLast) As per processMinibatch(DataSet, ComputationGraph, boolean) but used when SparkTrainingStats are being collected
Pair<R,SparkTrainingStats>	processMinibatchWithStats(org.nd4j.linalg.dataset.api.DataSet dataSet, MultiLayerNetwork network, boolean isLast) As per processMinibatch(DataSet, MultiLayerNetwork, boolean) but used when SparkTrainingStats are being collecte
Pair<R,SparkTrainingStats>	processMinibatchWithStats(org.nd4j.linalg.dataset.api.MultiDataSet dataSet, ComputationGraph graph, boolean isLast) As per processMinibatch(MultiDataSet, ComputationGraph, boolean) but used when SparkTrainingStats are being collected
void	removeHook(TrainingHook trainingHook) Remove a training hook from the worker

Method Detail

removeHook

void removeHook(TrainingHook trainingHook)

Remove a training hook from the worker

Parameters:

trainingHook - the training hook to remove

addHook

void addHook(TrainingHook trainingHook)

Add a training hook to be used during training of the worker

Parameters:

trainingHook - the training hook to add

getInitialModel

MultiLayerNetwork getInitialModel()

Get the initial model when training a MultiLayerNetwork/SparkDl4jMultiLayer

Returns:

Initial model for this worker

getInitialModelGraph

ComputationGraph getInitialModelGraph()

Get the initial model when training a ComputationGraph/SparkComputationGraph

Returns:

Initial model for this worker

processMinibatch

R processMinibatch(org.nd4j.linalg.dataset.api.DataSet dataSet,
                   MultiLayerNetwork network,
                   boolean isLast)

Process (fit) a minibatch for a MultiLayerNetwork

Parameters:

dataSet - Data set to train on

network - Network to train

isLast - If true: last data set currently available. If false: more data sets will be processed for this executor

Returns:

Null, or a training result if training should be terminated immediately.

processMinibatch

R processMinibatch(org.nd4j.linalg.dataset.api.DataSet dataSet,
                   ComputationGraph graph,
                   boolean isLast)

Process (fit) a minibatch for a ComputationGraph

Parameters:

dataSet - Data set to train on

graph - Network to train

isLast - If true: last data set currently available. If false: more data sets will be processed for this executor

Returns:

Null, or a training result if training should be terminated immediately.

processMinibatch

R processMinibatch(org.nd4j.linalg.dataset.api.MultiDataSet dataSet,
                   ComputationGraph graph,
                   boolean isLast)

Process (fit) a minibatch for a ComputationGraph using a MultiDataSet

Parameters:

dataSet - Data set to train on

graph - Network to train

isLast - If true: last data set currently available. If false: more data sets will be processed for this executor

Returns:

Null, or a training result if training should be terminated immediately.

processMinibatchWithStats

Pair<R,SparkTrainingStats> processMinibatchWithStats(org.nd4j.linalg.dataset.api.DataSet dataSet,
                                                     MultiLayerNetwork network,
                                                     boolean isLast)

As per processMinibatch(DataSet, MultiLayerNetwork, boolean) but used when SparkTrainingStats are being collecte

processMinibatchWithStats

Pair<R,SparkTrainingStats> processMinibatchWithStats(org.nd4j.linalg.dataset.api.DataSet dataSet,
                                                     ComputationGraph graph,
                                                     boolean isLast)

As per processMinibatch(DataSet, ComputationGraph, boolean) but used when SparkTrainingStats are being collected

processMinibatchWithStats

Pair<R,SparkTrainingStats> processMinibatchWithStats(org.nd4j.linalg.dataset.api.MultiDataSet dataSet,
                                                     ComputationGraph graph,
                                                     boolean isLast)

As per processMinibatch(MultiDataSet, ComputationGraph, boolean) but used when SparkTrainingStats are being collected

getFinalResult

R getFinalResult(MultiLayerNetwork network)

Get the final result to be returned to the driver

Parameters:

network - Current state of the network

Returns:

Result to return to the driver

getFinalResult

R getFinalResult(ComputationGraph graph)

Get the final result to be returned to the driver

Parameters:

graph - Current state of the network

Returns:

Result to return to the driver

getFinalResultNoData

R getFinalResultNoData()

Get the final result to be returned to the driver, if no data was available for this executor

Returns:

Result to return to the driver

getFinalResultNoDataWithStats

Pair<R,SparkTrainingStats> getFinalResultNoDataWithStats()

As per getFinalResultNoData() but used when SparkTrainingStats are being collected

getFinalResultWithStats

Pair<R,SparkTrainingStats> getFinalResultWithStats(MultiLayerNetwork network)

As per getFinalResult(MultiLayerNetwork) but used when SparkTrainingStats are being collected

getFinalResultWithStats

Pair<R,SparkTrainingStats> getFinalResultWithStats(ComputationGraph graph)

As per getFinalResult(ComputationGraph) but used when SparkTrainingStats are being collected

getDataConfiguration

WorkerConfiguration getDataConfiguration()

Get the WorkerConfiguration that contains information such as minibatch sizes, etc

Returns:

Worker configuration