LogGenDisMixFunction

Overview

Package

Class

Use

Tree

Deprecated

Index

Help

PREV CLASS NEXT CLASS

FRAMES NO FRAMES All Classes

SUMMARY: NESTED | FIELD | CONSTR | METHOD

DETAIL: FIELD | CONSTR | METHOD

de.jstacs.classifiers.differentiableSequenceScoreBased.gendismix
Class LogGenDisMixFunction

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

All Implemented Interfaces:: Function

Direct Known Subclasses:: OneDataSetLogGenDisMixFunction

public class LogGenDisMixFunction
extends DiffSSBasedOptimizableFunction
extends DiffSSBasedOptimizableFunction

This class implements the the following function

$\[f(\underline{\lambda}|C,D,\underline{\alpha},\underline{\beta})$

The weights $\beta_i$ 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 $\beta_2$ =0, one obtains the generative-discriminative trade-off,
if the $\beta_2$ =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 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
`protected double[]`	`beta` The mixture parameters of the GenDisMix
`protected double[][]`	`cllGrad` Array for the gradient of the conditional log-likelihood
`protected double[][]`	`helpArray` General temporary array
`protected double[][]`	`llGrad` Array for the gradient of the log-likelihood
`protected double[]`	`prGrad` Array for the gradient of the prior

Fields inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.DiffSSBasedOptimizableFunction
`dList, iList, prior, score, shortcut`

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

Fields inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractOptimizableFunction
`cl, clazz, data, freeParams, logClazz, norm, sum, weights`

Constructor Summary
`LogGenDisMixFunction(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
`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.
`protected double`	`joinFunction()` This method joins the partial results that have been computed using `AbstractMultiThreadedOptimizableFunction.evaluateFunction(int, int, int, int, int)`.
`protected double[]`	`joinGradients()` This method joins the gradients of each part that have been computed using `AbstractMultiThreadedOptimizableFunction.evaluateGradientOfFunction(int, int, int, int, int)`.
`void`	`reset(DifferentiableSequenceScore[] funs)` This method allows to reset the internally used functions and the corresponding objects.

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, setDataAndWeights, setParams, stopThreads`

Methods inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.AbstractOptimizableFunction
`getData, 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`

Field Detail

helpArray

protected double[][] helpArray

General temporary array

llGrad

protected double[][] llGrad

Array for the gradient of the log-likelihood

cllGrad

protected double[][] cllGrad

Array for the gradient of the conditional log-likelihood

beta

protected double[] beta

The mixture parameters of the GenDisMix

prGrad

protected double[] prGrad

Array for the gradient of the prior

Constructor Detail