nuroko.module.layers

Class SparseWeightLayer

java.lang.Object

nuroko.module.AComponent

nuroko.module.AWeightLayer

nuroko.module.layers.SparseWeightLayer

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 final class SparseWeightLayer
extends AWeightLayer

Fully connected weight layer

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.AWeightLayer

BIAS_INITIAL_SCALE, INITIAL_WEIGHT_SCALE, inputLength, outputLength

Constructor Summary

Constructors

Constructor and Description
SparseWeightLayer(int inputLength, int outputLength, int maxLinks)
SparseWeightLayer(SparseWeightLayer wl)

Method Summary

Methods

Modifier and Type	Method and Description
mikera.matrixx.AMatrix	asMatrix()
SparseWeightLayer	clone() Creates a clone of a module, including a deep copy of any mutable state.
mikera.vectorz.Vector	getBias()
List<IComponent>	getComponents()
mikera.vectorz.AVector	getGradient() Return an AVector referencing the accumulated gradient in this model
mikera.vectorz.Vector	getInput()
mikera.vectorz.Vector	getInputGradient()
IInputState	getInputState()
SparseWeightLayer	getInverse()
int	getLinkCount(int outputIndex)
int	getLinkSource(int outputIndex, int number)
double	getLinkWeight(int outputIndex, int number)
mikera.vectorz.Vector	getOutput()
mikera.vectorz.Vector	getOutputGradient()
int	getParameterLength() Returns the length of the parameter vector for this model
mikera.vectorz.AVector	getParameters() Return an AVector referring to the parameters in the model.
mikera.indexz.Index	getSourceIndex(int outputIndex)
mikera.vectorz.AVector	getSourceWeights(int outputIndex)
boolean	hasDifferentTrainingThinking()
void	initRandom() Randomly initialises a component's parameters
void	setInput(mikera.vectorz.AVector input)
void	think(mikera.vectorz.AVector input, mikera.vectorz.AVector output)
void	thinkInternal() Thinks within the scope of the component.
void	trainGradientInternal(double factor) Abstract method for training gradients.

Methods inherited from class nuroko.module.AWeightLayer

getInputLength, getModules, getOutputLength, trainGradient

Methods inherited from class nuroko.module.AComponent

applyConstraints, applyConstraintsInternal, generate, generate, getComponent, getConstraints, getDefaultLossFunction, getDownStack, getLearnFactor, getUpStack, isStochastic, isSynthesiser, iterator, setConstraint, setLearnFactor, setOutput, think, thinkInternalTraining, 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

SparseWeightLayer

public SparseWeightLayer(int inputLength,
                 int outputLength,
                 int maxLinks)

SparseWeightLayer

public SparseWeightLayer(SparseWeightLayer wl)

Method Detail

asMatrix

public mikera.matrixx.AMatrix asMatrix()

Specified by:

asMatrix in class AWeightLayer

getParameters

public mikera.vectorz.AVector getParameters()

Description copied from interface: IParameterised

Return an AVector referring to the parameters in the model. Expected to be mutable.

Returns:

getParameterLength

public int getParameterLength()

Description copied from interface: IParameterised

Returns the length of the parameter vector for this model

Specified by:

getParameterLength in interface IParameterised

Overrides:

getParameterLength in class AComponent

Returns:

think

public void think(mikera.vectorz.AVector input,
         mikera.vectorz.AVector output)

Specified by:

think in interface IThinker

Overrides:

think in class AComponent

thinkInternal

public void thinkInternal()

Description copied from interface: IComponent

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

getGradient

public mikera.vectorz.AVector getGradient()

Description copied from interface: IParameterised

Return an AVector referencing the accumulated gradient in this model

Returns:

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

clone

public SparseWeightLayer 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 AWeightLayer

Returns:

getLinkCount

public int getLinkCount(int outputIndex)

Specified by:

getLinkCount in class AWeightLayer

getLinkWeight

public double getLinkWeight(int outputIndex,
                   int number)

Specified by:

getLinkWeight in class AWeightLayer

getLinkSource

public int getLinkSource(int outputIndex,
                int number)

Specified by:

getLinkSource in class AWeightLayer

initRandom

public void initRandom()

Description copied from interface: IComponent

Randomly initialises a component's parameters

Specified by:

initRandom in interface IComponent

Specified by:

initRandom in class AWeightLayer

getSourceIndex

public mikera.indexz.Index getSourceIndex(int outputIndex)

Specified by:

getSourceIndex in class AWeightLayer

getSourceWeights

public mikera.vectorz.AVector getSourceWeights(int outputIndex)

Specified by:

getSourceWeights in class AWeightLayer

getInverse

public SparseWeightLayer getInverse()

Specified by:

getInverse in class AWeightLayer

hasDifferentTrainingThinking

public boolean hasDifferentTrainingThinking()

getBias

public mikera.vectorz.Vector getBias()

Specified by:

getBias in class AWeightLayer

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()

getComponents

public List<IComponent> getComponents()