org.deeplearning4j.parallelism.trainer

Class DefaultTrainer

java.lang.Object

java.lang.Thread

org.deeplearning4j.parallelism.trainer.DefaultTrainer

All Implemented Interfaces:

Runnable, Trainer

Direct Known Subclasses:

SymmetricTrainer

public class DefaultTrainer
extends Thread
implements Trainer

Trains datasets using a standard in memory parameter averaging technique. Think of this worker as the simplest form of doing parameter averaging

Author:

Adam Gibson

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	DefaultTrainer.DefaultTrainerBuilder

Nested classes/interfaces inherited from class java.lang.Thread

Thread.State, Thread.UncaughtExceptionHandler

Field Summary

Fields

Modifier and Type	Field and Description
protected int	averagingFrequency
protected AtomicBoolean	isStopped
protected AtomicLong	lastEtlTime
protected org.nd4j.linalg.dataset.api.DataSet	nullDataSet
protected AtomicBoolean	nullMode
protected boolean	onRootModel
protected org.deeplearning4j.nn.api.Model	originalModel
protected ParallelWrapper	parallelWrapper
protected LinkedBlockingQueue<org.nd4j.linalg.dataset.api.DataSet>	queue
protected LinkedBlockingQueue<org.nd4j.linalg.dataset.api.MultiDataSet>	queueMDS
protected org.deeplearning4j.nn.api.Model	replicatedModel
protected AtomicInteger	running
protected AtomicBoolean	shouldStop
protected AtomicBoolean	shouldUpdate
protected int	threadId
protected Exception	thrownException
protected boolean	useMDS
protected String	uuid
protected org.deeplearning4j.nn.conf.WorkspaceMode	workspaceMode

Fields inherited from class java.lang.Thread

MAX_PRIORITY, MIN_PRIORITY, NORM_PRIORITY

Constructor Summary

Constructors

Constructor and Description
DefaultTrainer()

Method Summary

Modifier and Type	Method and Description
boolean	averagingRequired() This method returns TRUE if this Trainer implementation assumes periodic aver
protected static org.deeplearning4j.optimize.api.IterationListener	cloneListener(org.deeplearning4j.optimize.api.IterationListener original)
protected void	configureListeners(String workerUUID, Collection<org.deeplearning4j.optimize.api.IterationListener> oldListeners, Collection<org.deeplearning4j.optimize.api.IterationListener> replicatedListeners)
void	feedDataSet(org.nd4j.linalg.dataset.api.DataSet dataSet, long etlTime) Train on a DataSet
void	feedMultiDataSet(org.nd4j.linalg.dataset.api.MultiDataSet dataSet, long etlTime) Train on a MultiDataSet
protected void	fit(org.nd4j.linalg.dataset.api.DataSet dataSet)
protected void	fit(org.nd4j.linalg.dataset.api.MultiDataSet dataSet)
org.deeplearning4j.nn.api.Model	getModel() THe current model for the trainer
boolean	isRunning()
protected void	postInit() This method does post-initialization configuration of Model.
void	run()
protected void	setupIfNeccessary()
void	shutdown() Shutdown this worker
void	updateModel(org.deeplearning4j.nn.api.Model model) Update the current Model for the worker
void	waitTillRunning() Block the main thread till the trainer is up and running.

Methods inherited from class java.lang.Thread

activeCount, checkAccess, clone, countStackFrames, currentThread, destroy, dumpStack, enumerate, getAllStackTraces, getContextClassLoader, getDefaultUncaughtExceptionHandler, getId, getName, getPriority, getStackTrace, getState, getThreadGroup, getUncaughtExceptionHandler, holdsLock, interrupt, interrupted, isAlive, isDaemon, isInterrupted, join, join, join, resume, setContextClassLoader, setDaemon, setDefaultUncaughtExceptionHandler, setName, setPriority, setUncaughtExceptionHandler, sleep, sleep, start, stop, stop, suspend, toString, yield

Methods inherited from class java.lang.Object

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

Methods inherited from interface org.deeplearning4j.parallelism.trainer.Trainer

setUncaughtExceptionHandler, start

Field Detail

replicatedModel

protected org.deeplearning4j.nn.api.Model replicatedModel

queue

protected LinkedBlockingQueue<org.nd4j.linalg.dataset.api.DataSet> queue

queueMDS

protected LinkedBlockingQueue<org.nd4j.linalg.dataset.api.MultiDataSet> queueMDS

running

protected AtomicInteger running

shouldUpdate

protected AtomicBoolean shouldUpdate

shouldStop

protected AtomicBoolean shouldStop

thrownException

protected Exception thrownException

useMDS

protected volatile boolean useMDS

uuid

protected final String uuid

onRootModel

protected boolean onRootModel

lastEtlTime

protected volatile AtomicLong lastEtlTime

nullMode

protected AtomicBoolean nullMode

nullDataSet

protected org.nd4j.linalg.dataset.api.DataSet nullDataSet

isStopped

protected AtomicBoolean isStopped

parallelWrapper

protected ParallelWrapper parallelWrapper

workspaceMode

protected org.deeplearning4j.nn.conf.WorkspaceMode workspaceMode

averagingFrequency

protected int averagingFrequency

threadId

protected int threadId

originalModel

protected org.deeplearning4j.nn.api.Model originalModel

Constructor Detail

DefaultTrainer

public DefaultTrainer()

Method Detail

feedMultiDataSet

public void feedMultiDataSet(@NonNull
                             org.nd4j.linalg.dataset.api.MultiDataSet dataSet,
                             long etlTime)

Description copied from interface: Trainer

Train on a MultiDataSet

Specified by:

feedMultiDataSet in interface Trainer

Parameters:

dataSet - the data set to train on

feedDataSet

public void feedDataSet(org.nd4j.linalg.dataset.api.DataSet dataSet,
                        long etlTime)

Description copied from interface: Trainer

Train on a DataSet

Specified by:

feedDataSet in interface Trainer

Parameters:

dataSet - the data set to train on

getModel

public org.deeplearning4j.nn.api.Model getModel()

Description copied from interface: Trainer

THe current model for the trainer

Specified by:

getModel in interface Trainer

Returns:

the current Model for the worker

updateModel

public void updateModel(@NonNull
                        org.deeplearning4j.nn.api.Model model)

Description copied from interface: Trainer

Update the current Model for the worker

Specified by:

updateModel in interface Trainer

Parameters:

model - the new model for this worker

setupIfNeccessary

protected void setupIfNeccessary()

isRunning

public boolean isRunning()

Specified by:

isRunning in interface Trainer

shutdown

public void shutdown()

Description copied from interface: Trainer

Shutdown this worker

Specified by:

shutdown in interface Trainer

fit

protected void fit(org.nd4j.linalg.dataset.api.DataSet dataSet)

fit

protected void fit(org.nd4j.linalg.dataset.api.MultiDataSet dataSet)

postInit

protected void postInit()

This method does post-initialization configuration of Model. Good place to configure listeners and all such a things

run

public void run()

Specified by:

run in interface Runnable

Overrides:

run in class Thread

waitTillRunning

public void waitTillRunning()

Description copied from interface: Trainer

Block the main thread till the trainer is up and running.

Specified by:

waitTillRunning in interface Trainer

averagingRequired

public boolean averagingRequired()

Description copied from interface: Trainer

This method returns TRUE if this Trainer implementation assumes periodic aver

Specified by:

averagingRequired in interface Trainer

Returns:

cloneListener

protected static org.deeplearning4j.optimize.api.IterationListener cloneListener(org.deeplearning4j.optimize.api.IterationListener original)

configureListeners

protected void configureListeners(String workerUUID,
                                  Collection<org.deeplearning4j.optimize.api.IterationListener> oldListeners,
                                  Collection<org.deeplearning4j.optimize.api.IterationListener> replicatedListeners)