de.jstacs.classifiers.differentiableSequenceScoreBased.logPrior
Class CompositeLogPrior

java.lang.Object
  de.jstacs.algorithms.optimization.DifferentiableFunction
      de.jstacs.classifiers.differentiableSequenceScoreBased.logPrior.LogPrior
          de.jstacs.classifiers.differentiableSequenceScoreBased.logPrior.CompositeLogPrior

All Implemented Interfaces:

Function, Storable

public class CompositeLogPrior
extends LogPrior

This class implements a composite prior that can be used for DifferentiableStatisticalModel. The prior for each DifferentiableStatisticalModel should be the (transformed) prior of the corresponding generative model. So e.g. for a PWM one should use an product of Dirichlets. The prior is more or less implemented in the DifferentiableStatisticalModel. For the class variables the prior uses a (transformed) Dirichlet with hyperparameters equal to the ESS of the classes.

If this class uses only the free parameters the class implements a real prior the is normalized to 1. If it used all parameters the function does not have (and is in general) not normalized to 1. Fortunately this is no problem, since it can be shown the it makes no difference in the optimization.

Author:

Jan Grau, Jens Keilwagen

See Also:

DifferentiableStatisticalModel, DifferentiableStatisticalModel.addGradientOfLogPriorTerm(double[], int), DifferentiableStatisticalModel.getLogPriorTerm()

Field Summary

Fields inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.logPrior.LogPrior

UNKNOWN

Constructor Summary

CompositeLogPrior()
The main constructor.

CompositeLogPrior(StringBuffer xml)
The constructor for the Storable interface.

Method Summary

void	addGradientFor(double[] params, double[] grad) Adds the gradient of the log-prior using the current parameters to a given vector.
double	evaluateFunction(double[] x) Evaluates the function at a certain vector (in mathematical sense) x.
int	getDimensionOfScope() Returns the dimension of the scope of the Function.
String	getInstanceName() Returns a short instance name.
CompositeLogPrior	getNewInstance() This method returns an empty new instance of the current prior.
void	set(boolean freeParameters, DifferentiableSequenceScore... funs) Resets all pre-computed values to their initial values using the DifferentiableSequenceScores funs.
StringBuffer	toXML() Encodes the prior as an XML representation.

Methods inherited from class de.jstacs.classifiers.differentiableSequenceScoreBased.logPrior.LogPrior

evaluateGradientOfFunction

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

CompositeLogPrior

public CompositeLogPrior()

The main constructor.

CompositeLogPrior

public CompositeLogPrior(StringBuffer xml)

The constructor for the Storable interface.

Parameters:

xml - the StringBuffer

Method Detail

set

public void set(boolean freeParameters,
                DifferentiableSequenceScore... funs)
         throws Exception

Description copied from class: LogPrior

Resets all pre-computed values to their initial values using the DifferentiableSequenceScores funs.

Overrides:

set in class LogPrior

Parameters:

freeParameters - the switch for using only the free parameters or all parameters in a DifferentiableSequenceScore

funs - the DifferentiableSequenceScores for the prior

Throws:

Exception - if the DifferentiableSequenceScores could not be set

addGradientFor

public void addGradientFor(double[] params,
                           double[] grad)
                    throws EvaluationException

Description copied from class: LogPrior

Adds the gradient of the log-prior using the current parameters to a given vector.

Specified by:

addGradientFor in class LogPrior

Parameters:

params - the parameters

grad - the vector

Throws:

EvaluationException - if the gradient could not be evaluated

evaluateFunction

public double evaluateFunction(double[] x)
                        throws DimensionException,
                               EvaluationException

Description copied from interface: Function

Evaluates the function at a certain vector (in mathematical sense) x.

Parameters:

x - the current vector

Returns:

the evaluation of the function

Throws:

DimensionException - if dim(x) != n, with

EvaluationException - if there was something wrong during the evaluation of the function

getDimensionOfScope

public int getDimensionOfScope()

Description copied from interface: Function

Returns the dimension of the scope of the Function.

Returns:

the dimension of the scope: n with

getNewInstance

public CompositeLogPrior getNewInstance()
                                 throws CloneNotSupportedException

Description copied from class: LogPrior

This method returns an empty new instance of the current prior. The method works similar to clone, but does not clone DifferentiableSequenceScore s that may be inside the instance. The DifferentiableSequenceScores must be set by an invocation of the method LogPrior.set(boolean, DifferentiableSequenceScore...).

Specified by:

getNewInstance in class LogPrior

Returns:

a new empty instance of the prior

Throws:

CloneNotSupportedException - if something went wrong while cloning

See Also:

LogPrior.set(boolean, DifferentiableSequenceScore...)

toXML

public StringBuffer toXML()

Description copied from class: LogPrior

Encodes the prior as an XML representation. It does not encode all information, since the method LogPrior.set(boolean, DifferentiableSequenceScore...) has to be invoked after decoding.

Specified by:

toXML in interface Storable

Specified by:

toXML in class LogPrior

Returns:

the XML representation;

getInstanceName

public String getInstanceName()

Description copied from class: LogPrior

Returns a short instance name.

Specified by:

getInstanceName in class LogPrior

Returns:

a short instance name