cc.mallet.fst.semi_supervised.pr.constraints
Class OneLabelL2PRConstraints

java.lang.Object
  cc.mallet.fst.semi_supervised.pr.constraints.OneLabelL2PRConstraints

All Implemented Interfaces:

PRConstraint

public class OneLabelL2PRConstraints
extends java.lang.Object
implements PRConstraint

A set of constraints on distributions over single labels conditioned on the presence of input features. This is to be used with PR, and penalizes L_2^2 difference from target expectations. Multiple constraints are grouped together here to make things more efficient.

Author:

Gregory Druck

Nested Class Summary

protected class

OneLabelL2PRConstraints.OneLabelPRConstraint

Field Summary

protected gnu.trove.TIntArrayList	cache
protected gnu.trove.TIntIntHashMap	constraintIndices
protected gnu.trove.TIntObjectHashMap<OneLabelL2PRConstraints.OneLabelPRConstraint>	constraints
protected StateLabelMap	map
protected boolean	normalized

Constructor Summary

	OneLabelL2PRConstraints(boolean normalized)
protected	OneLabelL2PRConstraints(gnu.trove.TIntObjectHashMap<OneLabelL2PRConstraints.OneLabelPRConstraint> constraints, gnu.trove.TIntIntHashMap constraintIndices, StateLabelMap map, boolean normalized)

Method Summary

void	addConstraint(int fi, double[] target, double weight)
void	addExpectations(double[] expectations)
PRConstraint	copy() This is used in multi-threading.
double	getAuxiliaryValueContribution(double[] parameters)
double	getCompleteValueContribution(double[] parameters)
void	getExpectations(double[] expectations)
void	getGradient(double[] parameters, double[] gradient)
double	getScore(FeatureVector input, int inputPosition, int srcIndex, int destIndex, double[] parameters)
void	incrementExpectations(FeatureVector input, int inputPosition, int srcIndex, int destIndex, double prob)
boolean	isOneStateConstraint()
int	numDimensions()
void	preProcess(FeatureVector fv) Gives the constraint the option to do some caching using only the FeatureVector.
java.util.BitSet	preProcess(InstanceList data)
void	setStateLabelMap(StateLabelMap map) Sets that map between the state indices and label indices.
void	zeroExpectations() Zero expectation values.

Methods inherited from class java.lang.Object

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

Field Detail

constraints

protected gnu.trove.TIntObjectHashMap<OneLabelL2PRConstraints.OneLabelPRConstraint> constraints

constraintIndices

protected gnu.trove.TIntIntHashMap constraintIndices

map

protected StateLabelMap map

normalized

protected boolean normalized

cache

protected gnu.trove.TIntArrayList cache

Constructor Detail

OneLabelL2PRConstraints

public OneLabelL2PRConstraints(boolean normalized)

OneLabelL2PRConstraints

protected OneLabelL2PRConstraints(gnu.trove.TIntObjectHashMap<OneLabelL2PRConstraints.OneLabelPRConstraint> constraints,
                                  gnu.trove.TIntIntHashMap constraintIndices,
                                  StateLabelMap map,
                                  boolean normalized)

Method Detail

copy

public PRConstraint copy()

Description copied from interface: PRConstraint

This is used in multi-threading.

Specified by:

copy in interface PRConstraint

Returns:

A copy of the GEConstraint.

addConstraint

public void addConstraint(int fi,
                          double[] target,
                          double weight)

numDimensions

public int numDimensions()

Specified by:

numDimensions in interface PRConstraint

isOneStateConstraint

public boolean isOneStateConstraint()

setStateLabelMap

public void setStateLabelMap(StateLabelMap map)

Description copied from interface: PRConstraint

Sets that map between the state indices and label indices.

Specified by:

setStateLabelMap in interface PRConstraint

Parameters:

map - StateLabelMap

preProcess

public void preProcess(FeatureVector fv)

Description copied from interface: PRConstraint

Gives the constraint the option to do some caching using only the FeatureVector. For example, the constrained input features could be cached.

Specified by:

preProcess in interface PRConstraint

Parameters:

fv - FeatureVector input

preProcess

public java.util.BitSet preProcess(InstanceList data)

Specified by:

preProcess in interface PRConstraint

Parameters:

data - Unlabeled data

Returns:

Returns a bitset of the size of the data, with the bit set if a constraint feature fires in that instance.

getScore

public double getScore(FeatureVector input,
                       int inputPosition,
                       int srcIndex,
                       int destIndex,
                       double[] parameters)

Specified by:

getScore in interface PRConstraint

incrementExpectations

public void incrementExpectations(FeatureVector input,
                                  int inputPosition,
                                  int srcIndex,
                                  int destIndex,
                                  double prob)

Specified by:

incrementExpectations in interface PRConstraint

getExpectations

public void getExpectations(double[] expectations)

Specified by:

getExpectations in interface PRConstraint

addExpectations

public void addExpectations(double[] expectations)

Specified by:

addExpectations in interface PRConstraint

zeroExpectations

public void zeroExpectations()

Description copied from interface: PRConstraint

Zero expectation values. Called before re-computing gradient.

Specified by:

zeroExpectations in interface PRConstraint

getAuxiliaryValueContribution

public double getAuxiliaryValueContribution(double[] parameters)

Specified by:

getAuxiliaryValueContribution in interface PRConstraint

getCompleteValueContribution

public double getCompleteValueContribution(double[] parameters)

Specified by:

getCompleteValueContribution in interface PRConstraint

getGradient

public void getGradient(double[] parameters,
                        double[] gradient)

Specified by:

getGradient in interface PRConstraint