org.encog.neural.networks.training.propagation

Class Propagation

java.lang.Object

org.encog.ml.train.BasicTraining

org.encog.neural.networks.training.propagation.Propagation

All Implemented Interfaces:

MLTrain, BatchSize, GradientWorkerOwner, Train, MultiThreadable

Direct Known Subclasses:

Backpropagation, ManhattanPropagation, QuickPropagation, ResilientPropagation, ScaledConjugateGradient

public abstract class Propagation
extends BasicTraining
implements Train, MultiThreadable, BatchSize, GradientWorkerOwner

Implements basic functionality that is needed by each of the propagation methods. The specifics of each of the propagation methods is implemented inside of the PropagationMethod interface implementors.

Constructor Summary

Constructors

Constructor and Description
Propagation(ContainsFlat network, MLDataSet training) Construct a propagation object.

Method Summary

Modifier and Type	Method and Description
void	calculateGradients() Calculate the gradients.
void	finishTraining() Should be called after training has completed and the iteration method will not be called any further.
void	finishTraining(double dropoutRate)
void	fixFlatSpot(boolean b) Default is true.
int	getBatchSize() The batch size.
FlatNetwork	getCurrentFlatNetwork()
double	getDropoutRate()
double	getL1()
double	getL2()
double[]	getLastGradient()
MLMethod	getMethod() Get the current best machine learning method from the training.
int	getThreadCount()
abstract void	initOthers()
void	iteration() Perform one training iteration.
void	iteration(int count) Perform the specified number of training iterations.
void	report(double[] gradients, double error, java.lang.Throwable ex) Called by the worker threads to report the progress at each step.
void	rollIteration() Increase the iteration by one.
void	setBatchSize(int theBatchSize) Set the batch size.
void	setDroupoutRate(double rate) Change the dropout rate
void	setErrorFunction(ErrorFunction ef)
void	setL1(double l1)
void	setL2(double l2)
void	setThreadCount(int numThreads) Set the number of threads.
abstract double	updateWeight(double[] gradients, double[] lastGradient, int index) Update a weight, the means by which weights are updated vary depending on the training.
abstract double	updateWeight(double[] gradients, double[] lastGradient, int index, double dropoutRate) Update a weight using dropout, the means by which weights are updated vary depending on the training.

Methods inherited from class org.encog.ml.train.BasicTraining

addStrategy, getError, getImplementationType, getIteration, getStrategies, getTraining, isTrainingDone, postIteration, preIteration, setError, setIteration, setTraining

Methods inherited from class java.lang.Object

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

Methods inherited from interface org.encog.ml.train.MLTrain

addStrategy, canContinue, getError, getImplementationType, getIteration, getStrategies, getTraining, isTrainingDone, pause, resume, setError, setIteration

Constructor Detail

Propagation

public Propagation(ContainsFlat network,
                   MLDataSet training)

Construct a propagation object.

Parameters:

network - The network.

training - The training set.

Method Detail

setDroupoutRate

public void setDroupoutRate(double rate)

Change the dropout rate

Parameters:

rate - The dropout rate.

getDropoutRate

public double getDropoutRate()

Returns:

the current dropout rate

finishTraining

public void finishTraining()

Should be called after training has completed and the iteration method will not be called any further.

Specified by:

finishTraining in interface MLTrain

Overrides:

finishTraining in class BasicTraining

finishTraining

public void finishTraining(double dropoutRate)

getCurrentFlatNetwork

public FlatNetwork getCurrentFlatNetwork()

Returns:

the currentFlatNetwork

getMethod

public MLMethod getMethod()

Get the current best machine learning method from the training.

Specified by:

getMethod in interface MLTrain

Returns:

The best machine learning method.

iteration

public void iteration()

Perform one training iteration.

Specified by:

iteration in interface MLTrain

rollIteration

public void rollIteration()

Increase the iteration by one.

iteration

public void iteration(int count)

Perform the specified number of training iterations. This can be more efficient than single training iterations. This is particularly true if you are training with a GPU.

Specified by:

iteration in interface MLTrain

Overrides:

iteration in class BasicTraining

Parameters:

count - The number of training iterations.

setThreadCount

public void setThreadCount(int numThreads)

Set the number of threads. Specify zero to tell Encog to automatically determine the best number of threads for the processor. If OpenCL is used as the target device, then this value is not used.

Specified by:

setThreadCount in interface MultiThreadable

Parameters:

numThreads - The number of threads.

getThreadCount

public int getThreadCount()

Specified by:

getThreadCount in interface MultiThreadable

Returns:

The number of threads to use, 0 to automatically determine based on core count.

fixFlatSpot

public void fixFlatSpot(boolean b)

Default is true. Call this with false to disable flat spot fix. For more info on flat spot: http://www.heatonresearch.com/wiki/Flat_Spot

Parameters:

b - True to fix flat spots, false otherwise.

setErrorFunction

public void setErrorFunction(ErrorFunction ef)

calculateGradients

public void calculateGradients()

Calculate the gradients.

report

public void report(double[] gradients,
                   double error,
                   java.lang.Throwable ex)

Called by the worker threads to report the progress at each step.

Specified by:

report in interface GradientWorkerOwner

Parameters:

gradients - The gradients from that worker.

error - The error for that worker.

ex - The exception.

initOthers

public abstract void initOthers()

updateWeight

public abstract double updateWeight(double[] gradients,
                                    double[] lastGradient,
                                    int index)

Update a weight, the means by which weights are updated vary depending on the training.

Parameters:

gradients - The gradients.

lastGradient - The last gradients.

index - The index.

Returns:

The update value.

updateWeight

public abstract double updateWeight(double[] gradients,
                                    double[] lastGradient,
                                    int index,
                                    double dropoutRate)

Update a weight using dropout, the means by which weights are updated vary depending on the training.

Parameters:

gradients - The gradients.

lastGradient - The last gradients.

index - The index.

dropoutRate - The dropout rate

Returns:

The update value.

getLastGradient

public double[] getLastGradient()

Returns:

the lastGradient

getBatchSize

public int getBatchSize()

The batch size. Specify 1 for pure online training. Specify 0 for pure batch training (complete training set in one batch). Otherwise specify the batch size for batch training.

Specified by:

getBatchSize in interface BatchSize

Returns:

The batch size.

setBatchSize

public void setBatchSize(int theBatchSize)

Set the batch size. Specify 1 for pure online training. Specify 0 for pure batch training (complete training set in one batch). Otherwise specify the batch size for batch training.

Specified by:

setBatchSize in interface BatchSize

Parameters:

theBatchSize - The batch size.

getL1

public double getL1()

Specified by:

getL1 in interface GradientWorkerOwner

Returns:

How much to apply l1 regularization penalty, 0 (default) for none.

setL1

public void setL1(double l1)

Parameters:

l1 - How much to apply l1 regularization penalty, 0 (default) for none.

getL2

public double getL2()

Specified by:

getL2 in interface GradientWorkerOwner

Returns:

How much to apply l2 regularization penalty, 0 (default) for none.

setL2

public void setL2(double l2)

Parameters:

l2 - How much to apply l2 regularization penalty, 0 (default) for none.