cmu.arktweetnlp.impl

Class Model

java.lang.Object

cmu.arktweetnlp.impl.Model

public class Model
extends java.lang.Object

This contains (1) Feature and label vocabularies (therefore knowledge of numberization) (2) Model coefficients (and knowledge how to flattenize them for LBFGS's sake) (3) Decoding/posterior and gradient computation

Field Summary

Fields

Modifier and Type	Field and Description
double[]	biasCoefs dim: N_labels
double[][]	edgeCoefs dim: (N_labels+1 x N_labels)
Vocabulary	featureVocab
Vocabulary	labelVocab
int	numLabels
double[][]	observationFeatureCoefs dim: (N_base_features x N_labels)

Constructor Summary

Constructors

Constructor and Description
Model()

Method Summary

Methods

Modifier and Type	Method and Description
void	allocateCoefs(int numLabels, int numObsFeats)
void	computeBiasScores(double[] labelScores) Adds into labelScores
void	computeEdgeScores(int t, ModelSentence sentence, double[] labelScores) Adds into labelScores
void	computeGradient(ModelSentence sentence, double[] grad) Training-only add-in loglik gradient (direction of higher likelihood)
void	computeLabelScores(int t, ModelSentence sentence, double[] labelScores) Computes unnormalized log-potentials.
double	computeLogLik(ModelSentence s)
void	computeObservedFeatureScores(int t, ModelSentence sentence, double[] labelScores) Adds into labelScores
void	computeVitLabelScores(int t, int prior, ModelSentence sentence, double[] labelScores)
double[]	convertCoefsToFlat()
static void	copyCoefsForIntersectingFeatures(Model sourceModel, Model destModel) Copies coefs from sourceModel into destModel.
int	flatIDsize()
void	greedyDecode(ModelSentence sentence, boolean storeConfidences) THIS CLOBBERS THE LABELS, stores its decoding into them.
double[][]	inferPosteriorForUnknownLabels(ModelSentence sentence) This needs forward-backward I think
double[][]	inferPosteriorGivenLabels(ModelSentence sentence) "given labels" i.e.
static Model	loadModelFromText(java.lang.String filename)
void	lockdownAfterFeatureExtraction()
void	mbrDecode(ModelSentence sentence)
void	saveModelAsText(java.lang.String outputFilename)
void	setCoefsFromFlat(double[] flatCoefs)
int	startMarker()
double[]	ThreewiseMultiply(double[] a, double[] b, double[] c)
void	viterbiDecode(ModelSentence sentence) vit[t][k] is the max probability such that the sequence from 0 to t has token t labeled with tag k.
void	viterbiEdgeScores(int prior, ModelSentence sentence, double[] EdgeScores)

Methods inherited from class java.lang.Object

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

Field Detail

labelVocab

public Vocabulary labelVocab

featureVocab

public Vocabulary featureVocab

biasCoefs

public double[] biasCoefs

dim: N_labels

edgeCoefs

public double[][] edgeCoefs

dim: (N_labels+1 x N_labels)

observationFeatureCoefs

public double[][] observationFeatureCoefs

dim: (N_base_features x N_labels)

numLabels

public int numLabels

Constructor Detail

Model

public Model()

Method Detail

startMarker

public int startMarker()

lockdownAfterFeatureExtraction

public void lockdownAfterFeatureExtraction()

allocateCoefs

public void allocateCoefs(int numLabels,
                 int numObsFeats)

inferPosteriorGivenLabels

public double[][] inferPosteriorGivenLabels(ModelSentence sentence)

"given labels" i.e. at trainingtime labels are observed. You hide the current one and predict it given you know the previous. So you get funny incremental posteriors per position that an MEMM uses at trainingtime. (They don't have a proper full-model posterior marginal interpretation like a CRF forward-backward-computed posterior does. no?)

Parameters:

sentence - - must its have .labels set

greedyDecode

public void greedyDecode(ModelSentence sentence,
                boolean storeConfidences)

THIS CLOBBERS THE LABELS, stores its decoding into them. Does progressive rolling edge feature extraction

inferPosteriorForUnknownLabels

public double[][] inferPosteriorForUnknownLabels(ModelSentence sentence)

This needs forward-backward I think

Returns:

dim: (T x K) posterior marginals at each position

viterbiDecode

public void viterbiDecode(ModelSentence sentence)

vit[t][k] is the max probability such that the sequence from 0 to t has token t labeled with tag k. (0<=tstartMarker)

mbrDecode

public void mbrDecode(ModelSentence sentence)

computeLabelScores

public void computeLabelScores(int t,
                      ModelSentence sentence,
                      double[] labelScores)

Computes unnormalized log-potentials. CLOBBERS labelScores

computeVitLabelScores

public void computeVitLabelScores(int t,
                         int prior,
                         ModelSentence sentence,
                         double[] labelScores)

computeBiasScores

public void computeBiasScores(double[] labelScores)

Adds into labelScores

computeEdgeScores

public void computeEdgeScores(int t,
                     ModelSentence sentence,
                     double[] labelScores)

Adds into labelScores

viterbiEdgeScores

public void viterbiEdgeScores(int prior,
                     ModelSentence sentence,
                     double[] EdgeScores)

computeObservedFeatureScores

public void computeObservedFeatureScores(int t,
                                ModelSentence sentence,
                                double[] labelScores)

Adds into labelScores

ThreewiseMultiply

public double[] ThreewiseMultiply(double[] a,
                         double[] b,
                         double[] c)

computeGradient

public void computeGradient(ModelSentence sentence,
                   double[] grad)

Training-only add-in loglik gradient (direction of higher likelihood)

computeLogLik

public double computeLogLik(ModelSentence s)

setCoefsFromFlat

public void setCoefsFromFlat(double[] flatCoefs)

convertCoefsToFlat

public double[] convertCoefsToFlat()

flatIDsize

public int flatIDsize()

saveModelAsText

public void saveModelAsText(java.lang.String outputFilename)
                     throws java.io.IOException

Throws:

java.io.IOException

loadModelFromText

public static Model loadModelFromText(java.lang.String filename)
                               throws java.io.IOException

Throws:

java.io.IOException

copyCoefsForIntersectingFeatures

public static void copyCoefsForIntersectingFeatures(Model sourceModel,
                                    Model destModel)

Copies coefs from sourceModel into destModel. For observation features, only copies features that exist in both. (Therefore if a feature exists in destModel but not sourceModel, it's not touched.)