nuroko.module

Class Dropout

java.lang.Object

nuroko.module.AComponent

nuroko.module.AOperationComponent

nuroko.module.Dropout

All Implemented Interfaces:

Cloneable, Iterable<IComponent>, IComponent, IFunction<mikera.vectorz.AVector,mikera.vectorz.AVector>, IInput, IInputOutput, IInputState, IModule, IOutput, IOutputState, IParameterised, ISynthesiser, IThinker, ITrainable

public class Dropout
extends AOperationComponent

Component that performs dropout on activation values. Useful for avoiding over-fitting.

Author:

Mike

Field Summary

Fields

Modifier and Type	Field and Description
protected mikera.vectorz.Vector	input
protected mikera.vectorz.Vector	inputGradient
protected mikera.vectorz.Vector	output
protected mikera.vectorz.Vector	outputGradient

Fields inherited from class nuroko.module.AOperationComponent

length

Constructor Summary

Constructors

Constructor and Description
Dropout(int length)
Dropout(int length, double dropoutRate)

Method Summary

Methods

Modifier and Type	Method and Description
Dropout	clone() Creates a clone of a module, including a deep copy of any mutable state.
mikera.vectorz.Vector	getInput()
mikera.vectorz.Vector	getInputGradient()
IInputState	getInputState()
mikera.vectorz.Vector	getOutput()
mikera.vectorz.Vector	getOutputGradient()
boolean	hasDifferentTrainingThinking()
void	setInput(mikera.vectorz.AVector input)
void	thinkInternal() Thinks within the scope of the component.
void	thinkInternalTraining() Thinks within the scope of the component.
void	trainGradientInternal(double factor) Abstract method for training gradients.

Methods inherited from class nuroko.module.AOperationComponent

getComponents, getGradient, getInputLength, getOutputLength, getParameters

Methods inherited from class nuroko.module.AComponent

applyConstraints, applyConstraintsInternal, generate, generate, getComponent, getConstraints, getDefaultLossFunction, getDownStack, getLearnFactor, getModules, getParameterLength, getUpStack, initRandom, isStochastic, isSynthesiser, iterator, setConstraint, setLearnFactor, setOutput, think, think, topComponent, train, train, trainSynth, trainSynth

Methods inherited from class java.lang.Object

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

Field Detail

input

protected final mikera.vectorz.Vector input

inputGradient

protected final mikera.vectorz.Vector inputGradient

output

protected final mikera.vectorz.Vector output

outputGradient

protected final mikera.vectorz.Vector outputGradient

Constructor Detail

Dropout

public Dropout(int length)

Dropout

public Dropout(int length,
       double dropoutRate)

Method Detail

thinkInternal

public void thinkInternal()

Description copied from interface: IComponent

Thinks within the scope of the component. Updates component's output.

thinkInternalTraining

public void thinkInternalTraining()

Description copied from interface: IComponent

Thinks within the scope of the component. Update's component's output. May be different from regular thinking.

Specified by:

thinkInternalTraining in interface IComponent

Specified by:

thinkInternalTraining in interface ITrainable

Overrides:

thinkInternalTraining in class AComponent

trainGradientInternal

public void trainGradientInternal(double factor)

Description copied from class: AComponent

Abstract method for training gradients. Assumes use of the current activation states for all inputs/outputs. Should overwrite inputGradient, and adjust gradient for all parameters. Should apply own internal learn rate factor

Specified by:

trainGradientInternal in interface IComponent

Specified by:

trainGradientInternal in class AComponent

hasDifferentTrainingThinking

public boolean hasDifferentTrainingThinking()

clone

public Dropout clone()

Description copied from interface: IModule

Creates a clone of a module, including a deep copy of any mutable state.

Specified by:

clone in interface IComponent

Specified by:

clone in interface IModule

Specified by:

clone in interface IParameterised

Specified by:

clone in interface ISynthesiser

Specified by:

clone in interface IThinker

Specified by:

clone in interface ITrainable

Specified by:

clone in class AOperationComponent

Returns:

getInputState

public IInputState getInputState()

Specified by:

getInputState in interface IComponent

Overrides:

getInputState in class AComponent

getInput

public mikera.vectorz.Vector getInput()

getOutput

public mikera.vectorz.Vector getOutput()

getOutputGradient

public mikera.vectorz.Vector getOutputGradient()

setInput

public void setInput(mikera.vectorz.AVector input)

Specified by:

setInput in interface IInputState

Overrides:

setInput in class AComponent

getInputGradient

public mikera.vectorz.Vector getInputGradient()