de.jstacs.sequenceScores.statisticalModels.trainable.hmm.transitions

Interface DifferentiableTransition

All Superinterfaces:

Cloneable, Storable, Transition

All Known Subinterfaces:

TrainableAndDifferentiableTransition

All Known Implementing Classes:

HigherOrderTransition

public interface DifferentiableTransition
extends Transition

This class declares methods that allow for optimizing the parameters numerically using the Optimizer.

Author:

Jens Keilwagen

Method Summary

Modifier and Type	Method and Description
void	addGradientForLogPriorTerm(double[] gradient, int start) This method computes the gradient of Transition.getLogPriorTerm() for each parameter of this transition.
void	fillParameters(double[] params) This method allows to fill the parameters of the transition in a given array.
void	fillSamplingGroups(int parameterOffset, LinkedList<int[]> list) Adds the groups of indexes of those parameters of this transition that should be sampled together in one step of a grouped sampling procedure, each as an int[], into list.
double	getLogScoreAndPartialDerivation(int layer, int index, int childIdx, IntList indices, DoubleList partDer, Sequence sequence, int sequencePosition) This method allows to compute the logarithm of the score and the gradient for a specific transition.
int	getSizeOfEventSpace(int index) Returns the size of the event space, i.e., the number of possible outcomes, for the random variable of parameter index
int	setParameterOffset(int offset) This method sets the internal offset of the parameter index.
void	setParameters(double[] params, int start) This method allows to set the parameters of the transition.

Methods inherited from interface de.jstacs.sequenceScores.statisticalModels.trainable.hmm.transitions.Transition

clone, fillTransitionInformation, getChildIdx, getGraphizNetworkRepresentation, getLastContextState, getLogPriorTerm, getLogScoreFor, getMaximalInDegree, getMaximalMarkovOrder, getMaximalNumberOfChildren, getNumberOfChildren, getNumberOfIndexes, getNumberOfStates, hasAnySelfTransitions, initializeRandomly, isAbsoring, setParameters, toString

Methods inherited from interface de.jstacs.Storable

toXML

Method Detail

setParameterOffset

int setParameterOffset(int offset)

This method sets the internal offset of the parameter index.

Parameters:

offset - the offset

Returns:

the offset for the next parameters

setParameters

void setParameters(double[] params,
                   int start)

This method allows to set the parameters of the transition.

Parameters:

params - the parameters

start - the (global) start position

fillParameters

void fillParameters(double[] params)

This method allows to fill the parameters of the transition in a given array.

Parameters:

params - the parameters

getLogScoreAndPartialDerivation

double getLogScoreAndPartialDerivation(int layer,
                                       int index,
                                       int childIdx,
                                       IntList indices,
                                       DoubleList partDer,
                                       Sequence sequence,
                                       int sequencePosition)

This method allows to compute the logarithm of the score and the gradient for a specific transition.

Parameters:

layer - the layer of the matrix

index - the index encoding the context

childIdx - the index of the child

indices - a list for the parameter indices

partDer - a list for the partial derivations

sequencePosition - the position within the sequence

sequence - the sequence

Returns:

the logarithm of the score

See Also:

Transition.getLogScoreFor(int, int, int, Sequence, int)

addGradientForLogPriorTerm

void addGradientForLogPriorTerm(double[] gradient,
                                int start)

This method computes the gradient of Transition.getLogPriorTerm() for each parameter of this transition. The results are added to the array gradient beginning at index start.

Parameters:

gradient - the array of gradients

start - the start index in the gradient array, where the partial derivations for the parameters of this Transition shall be entered

getSizeOfEventSpace

int getSizeOfEventSpace(int index)

Returns the size of the event space, i.e., the number of possible outcomes, for the random variable of parameter index

Parameters:

index - the index of the parameter

Returns:

the size of the event space

fillSamplingGroups

void fillSamplingGroups(int parameterOffset,
                        LinkedList<int[]> list)

Adds the groups of indexes of those parameters of this transition that should be sampled together in one step of a grouped sampling procedure, each as an int[], into list. In most cases, one group should contain the parameters that are living on a common simplex, e.g. the parameters of one TransitionElement of this transition. The internal indexes of the parameters are incremeneted by an external parameterOffset

Parameters:

parameterOffset - the external parameter offset

list - the list of sampling groups