de.jstacs.classifiers.differentiableSequenceScoreBased.gendismix

Class OneDataSetLogGenDisMixFunction

java.lang.Object

de.jstacs.algorithms.optimization.DifferentiableFunction

de.jstacs.classifiers.differentiableSequenceScoreBased.OptimizableFunction

de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractOptimizableFunction

de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractMultiThreadedOptimizableFunction

de.jstacs.classifiers.differentiableSequenceScoreBased.DiffSSBasedOptimizableFunction

de.jstacs.classifiers.differentiableSequenceScoreBased.gendismix.LogGenDisMixFunction

de.jstacs.classifiers.differentiableSequenceScoreBased.gendismix.OneDataSetLogGenDisMixFunction

All Implemented Interfaces:

Function, MultiThreadedFunction

public class OneDataSetLogGenDisMixFunction
extends LogGenDisMixFunction

This class implements the the following function

where is the weight for sequence and class . The weights have to sum to 1. For special weights the optimization turns out to be well known

• if the weights are (0,1,0), one obtains maximum likelihood,
• if the weights are (0,0.5,0.5), one obtains maximum a posteriori,
• if the weights are (1,0,0), one obtains maximum conditional likelihood,
• if the weights are (0.5,0,0.5), one obtains maximum supervised posterior,
• if the =0, one obtains the generative-discriminative trade-off,
• if the =0.5, one obtains the penalized generative-discriminative trade-off.

Of course, there are also some very interesting cases with other weights.

It can be used to maximize the parameters. This can also be done with LogGenDisMixFunction. However, we implemented this class to allow a faster function and gradient evaluation leading to a faster optimization. This becomes especially interesting if the number of classes increases.

This class enables the user to exploit all CPUs of the computer by using threads. The number of compute threads can be determined in the constructor.

It is very important for this class that the DifferentiableSequenceScore.clone() method works correctly, since each thread works on its own clones.

Author:

Jens Keilwagen

Nested Class Summary

Nested classes/interfaces inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.OptimizableFunction

OptimizableFunction.KindOfParameter

Field Summary

Fields inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.gendismix.LogGenDisMixFunction

beta, cllGrad, helpArray, llGrad, prGrad

Fields inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.DiffSSBasedOptimizableFunction

dList, iList, prior, score, shortcut

Fields inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractMultiThreadedOptimizableFunction

params, worker

Fields inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractOptimizableFunction

cl, clazz, data, freeParams, logClazz, norm, sum, weights

Constructor Summary

Constructors

Constructor and Description
OneDataSetLogGenDisMixFunction(int threads, DifferentiableSequenceScore[] score, DataSet data, double[][] weights, LogPrior prior, double[] beta, boolean norm, boolean freeParams) The constructor for creating an instance that can be used in an Optimizer.

Method Summary

Modifier and Type	Method and Description
protected void	evaluateFunction(int index, int startClass, int startSeq, int endClass, int endSeq) This method evaluates the function for a part of the data.
protected void	evaluateGradientOfFunction(int index, int startClass, int startSeq, int endClass, int endSeq) This method evaluates the gradient of the function for a part of the data.
DataSet[]	getData() Returns the data for each class used in this OptimizableFunction.
void	setDataAndWeights(DataSet[] data, double[][] weights) This method sets the data set and the sequence weights to be used.

Methods inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.gendismix.LogGenDisMixFunction

joinFunction, joinGradients, reset

Methods inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.DiffSSBasedOptimizableFunction

addTermToClassParameter, getClassParams, getDimensionOfScope, getParameters, reset, setParams, setThreadIndependentParameters

Methods inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractMultiThreadedOptimizableFunction

evaluateFunction, evaluateGradientOfFunction, getNumberOfAvailableProcessors, getNumberOfThreads, prepareThreads, setParams, stopThreads

Methods inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractOptimizableFunction

getParameters, getSequenceWeights

Methods inherited from class de.jstacs.algorithms.optimization.DifferentiableFunction

findOneDimensionalMin

Methods inherited from class java.lang.Object

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

Constructor Detail

OneDataSetLogGenDisMixFunction

public OneDataSetLogGenDisMixFunction(int threads,
                                      DifferentiableSequenceScore[] score,
                                      DataSet data,
                                      double[][] weights,
                                      LogPrior prior,
                                      double[] beta,
                                      boolean norm,
                                      boolean freeParams)
                               throws IllegalArgumentException

The constructor for creating an instance that can be used in an Optimizer.

Parameters:

threads - the number of threads used for evaluating the function and determining the gradient of the function

score - an array containing the DifferentiableSequenceScores that are used for determining the sequences scores; if the weight beta[LearningPrinciple.LIKELIHOOD_INDEX] is positive all elements of score have to be DifferentiableStatisticalModel

data - the array of DataSets containing the data that is needed to evaluate the function

weights - the weights for each Sequence in each DataSet of data

prior - the prior that is used for learning the parameters

beta - the beta-weights for the three terms of the learning principle

norm - the switch for using the normalization (division by the number of sequences)

freeParams - the switch for using only the free parameters

Throws:

IllegalArgumentException - if the number of threads is not positive, the number of classes or the dimension of the weights is not correct

Method Detail

setDataAndWeights

public void setDataAndWeights(DataSet[] data,
                              double[][] weights)
                       throws IllegalArgumentException

Description copied from class: OptimizableFunction

This method sets the data set and the sequence weights to be used. It also allows to do further preparation for the computation on this data.

Overrides:

setDataAndWeights in class AbstractMultiThreadedOptimizableFunction

Parameters:

data - the data sets

weights - the sequence weights for each sequence in each data set

Throws:

IllegalArgumentException - if the data or the weights can not be used

getData

public DataSet[] getData()

Description copied from class: OptimizableFunction

Returns the data for each class used in this OptimizableFunction.

Overrides:

getData in class AbstractOptimizableFunction

Returns:

the data for each class

See Also:

OptimizableFunction.getSequenceWeights()

evaluateGradientOfFunction

protected void evaluateGradientOfFunction(int index,
                                          int startClass,
                                          int startSeq,
                                          int endClass,
                                          int endSeq)

Description copied from class: AbstractMultiThreadedOptimizableFunction

This method evaluates the gradient of the function for a part of the data.

Overrides:

evaluateGradientOfFunction in class LogGenDisMixFunction

Parameters:

index - the index of the part

startClass - the index of the start class

startSeq - the index of the start sequence

endClass - the index of the end class (inclusive)

endSeq - the index of the end sequence (exclusive)

evaluateFunction

protected void evaluateFunction(int index,
                                int startClass,
                                int startSeq,
                                int endClass,
                                int endSeq)
                         throws EvaluationException

Description copied from class: AbstractMultiThreadedOptimizableFunction

This method evaluates the function for a part of the data.

Overrides:

evaluateFunction in class LogGenDisMixFunction

Parameters:

index - the index of the part

startClass - the index of the start class

startSeq - the index of the start sequence

endClass - the index of the end class (inclusive)

endSeq - the index of the end sequence (exclusive)

Throws:

EvaluationException - if the gradient could not be evaluated properly