org.deeplearning4j.parallelism

Class ParallelWrapper.Builder<T extends Model>

java.lang.Object

org.deeplearning4j.parallelism.ParallelWrapper.Builder<T>

Enclosing class:

ParallelWrapper

public static class ParallelWrapper.Builder<T extends Model>
extends java.lang.Object

Field Summary

Fields

Modifier and Type	Field and Description
protected boolean	averageUpdaters
protected int	averagingFrequency
protected boolean	isMQ
protected boolean	legacyAveraging
protected T	model
protected int	prefetchSize
protected boolean	reportScore
protected TrainerContext	trainerContext
protected java.lang.Object[]	trainerContextArgs
protected int	workers

Constructor Summary

Constructors

Constructor and Description
Builder(T model) Build ParallelWrapper for MultiLayerNetwork

Method Summary

Modifier and Type	Method and Description
ParallelWrapper.Builder	averageUpdaters(boolean reallyAverage) This method enables/disables updaters averaging.
ParallelWrapper.Builder	averagingFrequency(int freq) Model averaging frequency.
ParallelWrapper	build() This method returns ParallelWrapper instance
ParallelWrapper.Builder	prefetchBuffer(int size) Size of prefetch buffer that will be used for background data prefetching.
ParallelWrapper.Builder	reportScoreAfterAveraging(boolean reallyReport) This method enables/disables averaged model score reporting
ParallelWrapper.Builder	trainerContextArgs(java.lang.Object... trainerContextArgs) Transer context args are for calling a TrainerContext init method when ParallelWrapper starts training
ParallelWrapper.Builder	trainerFactory(TrainerContext trainerContext) Specify a TrainerContext for the given ParallelWrapper instance.
ParallelWrapper.Builder	useLegacyAveraging(boolean reallyUse) If set to true, legacy averaging method is used.
ParallelWrapper.Builder	useMQ(boolean reallyUse) This method enables/disable MagicQueue use If set to true, all datasets will be spread among all available devices at prefetch phase using AsyncDataSetIterator PLEASE NOTE: This is experimental feature.
ParallelWrapper.Builder	workers(int num) This method allows to configure number of workers that'll be used for parallel training

Methods inherited from class java.lang.Object

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

Field Detail

model

protected T extends Model model

workers

protected int workers

prefetchSize

protected int prefetchSize

averagingFrequency

protected int averagingFrequency

reportScore

protected boolean reportScore

averageUpdaters

protected boolean averageUpdaters

legacyAveraging

protected boolean legacyAveraging

isMQ

protected boolean isMQ

trainerContext

protected TrainerContext trainerContext

trainerContextArgs

protected java.lang.Object[] trainerContextArgs

Constructor Detail

Builder

public Builder(@NonNull
               T model)

Build ParallelWrapper for MultiLayerNetwork

Parameters:

model -

Method Detail

trainerContextArgs

public ParallelWrapper.Builder trainerContextArgs(java.lang.Object... trainerContextArgs)

Transer context args are for calling a TrainerContext init method when ParallelWrapper starts training

Parameters:

trainerContextArgs - the args to use (maybe null)

Returns:

trainerFactory

public ParallelWrapper.Builder trainerFactory(TrainerContext trainerContext)

Specify a TrainerContext for the given ParallelWrapper instance. Defaults to DefaultTrainerContext otherwise

Parameters:

trainerContext - the trainer factory to use

Returns:

builder pattern

workers

public ParallelWrapper.Builder workers(int num)

This method allows to configure number of workers that'll be used for parallel training

Parameters:

num -

Returns:

averagingFrequency

public ParallelWrapper.Builder averagingFrequency(int freq)

Model averaging frequency.

Parameters:

freq - number of iterations between averaging

Returns:

averageUpdaters

public ParallelWrapper.Builder averageUpdaters(boolean reallyAverage)

This method enables/disables updaters averaging. Default value: TRUE PLEASE NOTE: This method is suitable for debugging purposes mostly. So don't change default value, unless you're sure why you need it.

Parameters:

reallyAverage -

Returns:

useMQ

public ParallelWrapper.Builder useMQ(boolean reallyUse)

This method enables/disable MagicQueue use If set to true, all datasets will be spread among all available devices at prefetch phase using AsyncDataSetIterator PLEASE NOTE: This is experimental feature. Default: false

Parameters:

reallyUse -

Returns:

prefetchBuffer

public ParallelWrapper.Builder prefetchBuffer(int size)

Size of prefetch buffer that will be used for background data prefetching. Usually it's better to keep this value equal to the number of workers. Default value: 2

Parameters:

size - 0 to disable prefetching, any positive number

Returns:

useLegacyAveraging

public ParallelWrapper.Builder useLegacyAveraging(boolean reallyUse)

If set to true, legacy averaging method is used. This might be used as fallback on multi-gpu systems without P2P access available. Default value: false

Parameters:

reallyUse -

Returns:

reportScoreAfterAveraging

public ParallelWrapper.Builder reportScoreAfterAveraging(boolean reallyReport)

This method enables/disables averaged model score reporting

Parameters:

reallyReport -

Returns:

build

public ParallelWrapper build()

This method returns ParallelWrapper instance

Returns: