A class to hold the several feature vectors needed during training (especially for the perceptron family of algorithms), as well as for performing the updates to those feature vectors (again, with the perceptron family of algorithms in mind). More...

#include <training-vector-set.H>

Public Member Functions
	TrainingVectorSet ()
	Constructs a new set of feature vectors (models) for use during training. More...

virtual	~TrainingVectorSet ()
	Destroys this instance. More...

const FeatureVector< int, double > &	weights () const
	Returns the "raw" feature weights computed during training. More...

const FeatureVector< int, double > &	average_weights () const
	Returns the feature vector corresponding to the averaged perceptron. More...

const FeatureVector< int, double > &	GetModel (bool raw) const
	Returns either the raw or averaged feature vector, depending on the argument. More...

template<typename Collection >
void	UpdateWeights (const Time &time, const Collection &feature_uids, const FeatureVector< int, double > &feature_vector, double scalar)
	Increments the weights for the specified collection of features. More...

template<typename Collection >
void	UpdateGoldAndCandidateFeatureAverages (const Time &time, const Collection &gold_feature_uids, const Collection &candidate_feature_uids)
	Updates the feature averages the specified pair of feature uid collections, one for a gold reference training instance and the other for a one-best hypothesis. More...

void	UpdateAllFeatureAverages (const Time &time)

void	RemapFeatureUids (const unordered_map< int, int > &old_to_new_uids)

Friends
class	PerceptronModelProtoReader

ostream &	operator<< (ostream &os, const TrainingVectorSet &tvs)

Detailed Description

A class to hold the several feature vectors needed during training (especially for the perceptron family of algorithms), as well as for performing the updates to those feature vectors (again, with the perceptron family of algorithms in mind).

Definition at line 59 of file training-vector-set.H.

Constructor & Destructor Documentation

reranker::TrainingVectorSet::TrainingVectorSet ( )

inline

Constructs a new set of feature vectors (models) for use during training.

Definition at line 63 of file training-vector-set.H.

virtual reranker::TrainingVectorSet::~TrainingVectorSet ( )

inlinevirtual

Destroys this instance.

Definition at line 65 of file training-vector-set.H.

Member Function Documentation

const FeatureVector<int,double>& reranker::TrainingVectorSet::average_weights ( ) const

inline

Returns the feature vector corresponding to the averaged perceptron.

Definition at line 74 of file training-vector-set.H.

const FeatureVector<int,double>& reranker::TrainingVectorSet::GetModel ( bool raw ) const

inline

Returns either the raw or averaged feature vector, depending on the argument.

Parameters

raw	if true, return the raw model; otherwise, return the averaged perceptron model

Definition at line 83 of file training-vector-set.H.

void reranker::TrainingVectorSet::RemapFeatureUids ( const unordered_map< int, int > & old_to_new_uids )

inline

Definition at line 136 of file training-vector-set.H.

void reranker::TrainingVectorSet::UpdateAllFeatureAverages ( const Time & time )

inline

Definition at line 128 of file training-vector-set.H.

template<typename Collection >

void reranker::TrainingVectorSet::UpdateGoldAndCandidateFeatureAverages	(	const Time &	time,
		const Collection &	gold_feature_uids,
		const Collection &	candidate_feature_uids
	)

inline

Updates the feature averages the specified pair of feature uid collections, one for a gold reference training instance and the other for a one-best hypothesis.

Definition at line 115 of file training-vector-set.H.

template<typename Collection >

void reranker::TrainingVectorSet::UpdateWeights	(	const Time &	time,
		const Collection &	feature_uids,
		const FeatureVector< int, double > &	feature_vector,
		double	scalar
	)

inline

Increments the weights for the specified collection of features.

Technically, this method adds a scaled version of the specified vector projected into the subspace specified by the collection of feature uid's.

Parameters

time	the current training time
feature_uids	the set of features to be updated (essentially a subspace into which to project `feature_vector`)
feature_vector	the feature vector to be projected into the specified subspace, scaled and then added to the current set of model weights
scalar	the amount by which to scale the specified subvector