org.nd4j.parameterserver.distributed.training.impl

Class SkipGramTrainer

java.lang.Object

org.nd4j.parameterserver.distributed.training.BaseTrainer<SkipGramRequestMessage>

org.nd4j.parameterserver.distributed.training.impl.SkipGramTrainer

All Implemented Interfaces:

TrainingDriver<SkipGramRequestMessage>

public class SkipGramTrainer
extends BaseTrainer<SkipGramRequestMessage>

Distributed SkipGram trainer TrainingDriver idea is simple: 1) We get request from Client 2) We initiate training by issuing DotRequest 3) Each Shard does Dot accumulation 4) As soon as Dot aggregated, we calculate gradients independently 5) As soon as they are ready - we just apply them to appropriate

Author:

[email protected]

Field Summary

Fields

Modifier and Type	Field and Description
protected java.util.Map<RequestDescriptor,SkipGramChain>	chains
protected java.util.concurrent.atomic.AtomicLong	cntRounds

Fields inherited from class org.nd4j.parameterserver.distributed.training.BaseTrainer

clipboard, completionHandler, storage, transport, voidConfiguration

Constructor Summary

Constructors

Constructor and Description
SkipGramTrainer()

Method Summary

Modifier and Type	Method and Description
void	aggregationFinished(VoidAggregation aggregation) This method is invoked after particular aggregation finished
void	finishTraining(long originatorId, long taskId)
void	pickTraining(SkipGramRequestMessage message) This method will be called from non-initialized Shard context
void	startTraining(SkipGramRequestMessage message)
java.lang.String	targetMessageClass()

Methods inherited from class org.nd4j.parameterserver.distributed.training.BaseTrainer

addCompletionHook, init, replicate

Methods inherited from class java.lang.Object

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

Field Detail

chains

protected java.util.Map<RequestDescriptor,SkipGramChain> chains

cntRounds

protected java.util.concurrent.atomic.AtomicLong cntRounds

Constructor Detail

SkipGramTrainer

public SkipGramTrainer()

Method Detail

startTraining

public void startTraining(SkipGramRequestMessage message)

pickTraining

public void pickTraining(@NonNull
                         SkipGramRequestMessage message)

This method will be called from non-initialized Shard context

Parameters:

message -

targetMessageClass

public java.lang.String targetMessageClass()

aggregationFinished

public void aggregationFinished(@NonNull
                                VoidAggregation aggregation)

This method is invoked after particular aggregation finished

Parameters:

aggregation -

finishTraining

public void finishTraining(long originatorId,
                           long taskId)