Single-Class Hidden Markov Model. More...

#include <xmmHmmSingleClass.hpp>

Inheritance diagram for xmm::SingleClassHMM:

Collaboration diagram for xmm::SingleClassHMM:

Public Member Functions
	SingleClassHMM (std::shared_ptr< SharedParameters > p=NULL)
	Constructor. More...

	SingleClassHMM (SingleClassHMM const &src)
	Copy constructor. More...

	SingleClassHMM (std::shared_ptr< SharedParameters > p, Json::Value const &root)
	Copy constructor. More...

SingleClassHMM &	operator= (SingleClassHMM const &src)
	Assignment. More...

Accessors
void	addExitPoint (int stateIndex, float proba)
	Set the exit probability of a specific state. More...

Performance
void	reset ()
	Resets the fitering process (recognition or regression) More...

double	filter (std::vector< float > const &observation)
	filters a incoming observation (performs recognition or regression) More...

Json I/O
virtual Json::Value	toJson () const
	Write the object to a JSON Structure. More...

virtual void	fromJson (Json::Value const &root)
	Read the object from a JSON Structure. More...

Training
bool	isTraining () const
	Checks if the model is training. More...

void	train (TrainingSet *trainingSet)
	Main training method based on the EM algorithm. More...

void	cancelTraining ()
	Cancels the training process : sets a flag so that the training stops at the next possible exit in the training process. More...

Python File I/O
void	writeFile (char *fileName) const
	write method for python wrapping ('write' keyword forbidden, name has to be different) More...

void	readFile (char *fileName)
	read method for python wrapping ('read' keyword forbidden, name has to be different) More...

std::string	__str__ () const
	"print" method for python => returns the results of write method More...

Static Public Member Functions
static const float	DEFAULT_EXITPROBABILITY_LAST_STATE ()

static const float	TRANSITION_REGULARIZATION ()

Public Attributes
ClassParameters< HMM >	parameters
	Convert to bimodal HMM in place. More...

ClassResults< HMM >	results
	Results of the filtering process (recognition & regression) More...

std::vector< double >	alpha
	State probabilities estimated by the forward algorithm. More...

std::vector< double >	alpha_h [3]
	State probabilities estimated by the hierarchical forward algorithm. More...

std::vector< SingleClassGMM >	states
	States of the model (Gaussian Mixture Models) More...

std::vector< float >	prior
	Prior probabilities. More...

std::vector< float >	transition
	Transition Matrix. More...

std::string	label
	label associated with the given model More...

std::shared_ptr< SharedParameters >	shared_parameters
	Pointer to the shared parameters owned by a multi-class model. More...

EventGenerator< TrainingEvent >	training_events
	Generator for events monitoring the training process. More...

TrainingEvent	training_status
	Event containing information on the current status of the training process. More...

Protected Member Functions
void	allocate ()
	Allocate model parameters. More...

void	initParametersToDefault (std::vector< float > const &dataStddev)
	initialize model parameters to their default values More...

void	initMeansWithAllPhrases (TrainingSet *trainingSet)
	initialize the means of each state with all training phrases (single gaussian) More...

void	initCovariances_fullyObserved (TrainingSet *trainingSet)
	initialize the covariances of each state with all training phrases (single gaussian) More...

void	initMeansCovariancesWithGMMEM (TrainingSet *trainingSet)
	initialize the means and covariances of each state using GMM-EM on segments. More...

void	setErgodic ()
	set the prior and transition matrix to ergodic More...

void	setLeftRight ()
	set the prior and transition matrix to left-right (no state skip) More...

void	normalizeTransitions ()
	Normalize transition probabilities. More...

double	forward_init (const float observation, const float observation_output=NULL)
	Initialization of the forward algorithm. More...

double	forward_update (const float observation, const float observation_output=NULL)
	Update of the forward algorithm. More...

void	backward_init (double ct)
	Initialization Backward algorithm. More...

void	backward_update (double ct, const float observation, const float observation_output=NULL)
	Update of the Backward algorithm. More...

void	emAlgorithmInit (TrainingSet *trainingSet)
	Initialization of the parameters before training. More...

void	emAlgorithmTerminate ()
	Termination of the training algorithm. More...

virtual double	emAlgorithmUpdate (TrainingSet *trainingSet)
	update method of the EM algorithm (calls Baum-Welch Algorithm) More...

double	baumWelch_forwardBackward (std::shared_ptr< Phrase > currentPhrase, int phraseIndex)
	Compute the forward-backward algorithm on a phrase of the training set. More...

double	baumWelch_forward_update (std::vector< double >::iterator observation_likelihoods)
	Update of the forward algorithm for Training (observation probabilities are pre-computed) More...

void	baumWelch_backward_update (double ct, std::vector< double >::iterator observation_likelihoods)
	Update of the Backward algorithm for Training (observation probabilities are pre-computed) More...

void	baumWelch_gammaSum (TrainingSet *trainingSet)
	Compute the sum of the gamma variable (for use in EM) More...

void	baumWelch_estimateMixtureCoefficients (TrainingSet *trainingSet)
	Estimate the Coefficients of the Gaussian Mixture for each state. More...

void	baumWelch_estimateMeans (TrainingSet *trainingSet)
	Estimate the Means of the Gaussian Distribution for each state. More...

void	baumWelch_estimateCovariances (TrainingSet *trainingSet)
	Estimate the Covariances of the Gaussian Distribution for each state. More...

void	baumWelch_estimatePrior (TrainingSet *trainingSet)
	Estimate the Prior Probabilities. More...

void	baumWelch_estimateTransitions (TrainingSet *trainingSet)
	Estimate the Transition Probabilities. More...

void	addCyclicTransition (double proba)
	Adds a cyclic Transition probability (from last state to first state) More...

void	updateAlphaWindow ()
	Estimates the likeliest state and compute the bounds of the windows over the states. More...

void	regression (std::vector< float > const &observation_input)
	Compute the regression for the case of a bimodal model, given the estimated state probabilities estimated by forward algorithm. More...

void	updateResults ()
	update the content of the likelihood buffer and return average likelihood. More...

void	updateExitProbabilities (float *_exitProbabilities=NULL)
	Update the exit probability vector given the probabilities. More...

bool	emAlgorithmHasConverged (int step, double log_prob, double old_log_prob) const
	checks if the training has converged according to the object's EM stop criterion More...

bool	cancelTrainingIfRequested ()
	checks if a cancel request has been sent and accordingly cancels the training process More...

void	check_training () const
	Checks if the model is still training. More...

Protected Attributes
bool	forward_initialized_
	Defines if the forward algorithm has been initialized. More...

std::vector< double >	previous_alpha_
	used to store the alpha estimated at the previous time step More...

std::vector< double >	beta_
	backward state probabilities More...

std::vector< double >	previous_beta_
	used to store the beta estimated at the previous time step More...

std::vector< std::vector< double > >	gamma_sequence_
	Sequence of Gamma probabilities. More...

std::vector< std::vector< double > >	epsilon_sequence_
	Sequence of Epsilon probabilities. More...

std::vector< std::vector< std::vector< double > > >	gamma_sequence_per_mixture_
	Sequence of Gamma probabilities for each mixture component. More...

std::vector< double >	alpha_seq_
	Sequence of alpha (forward) probabilities. More...

std::vector< double >	beta_seq_
	Sequence of beta (backward) probabilities. More...

std::vector< double >	gamma_sum_
	Used to store the sums of the gamma variable. More...

std::vector< double >	gamma_sum_per_mixture_
	Used to store the sums of the gamma variable for each mixture component. More...

bool	is_hierarchical_
	Defines if the model is a submodel of a hierarchical HMM. More...

std::vector< float >	exit_probabilities_
	Exit probabilities for a hierarchical model. More...

int	window_minindex_
	minimum index of the alpha window (used for regression & time progression) More...

int	window_maxindex_
	minimum index of the alpha window (used for regression & time progression) More...

double	window_normalization_constant_
	normalization constant of the alpha window (used for regression & time progression) More...

CircularBuffer< double >	likelihood_buffer_
	Likelihood buffer used for smoothing. More...

std::mutex	training_mutex_
	Mutex used in Concurrent Mode. More...

bool	is_training_
	defines if the model is being trained. More...

bool	cancel_training_
	defines if the model received a request to cancel training More...

Friends
template<typename SingleClassModel , typename ModelType >
class	Model

class	HierarchicalHMM

Detailed Description

Single-Class Hidden Markov Model.

Hidden Markov Model with Multivariate Gaussian Mixture Model Observation Distributions. Supports Bimodal data and Hidden Markov Regression. Built for Hierarchical HMMs

Constructor & Destructor Documentation

xmm::SingleClassHMM::SingleClassHMM ( std::shared_ptr< SharedParameters > p = NULL )

Constructor.

Parameters

p	Shared Parameters (owned by a multiclass object)

xmm::SingleClassHMM::SingleClassHMM ( SingleClassHMM const & src )

Copy constructor.

Parameters

src	Source Model

xmm::SingleClassHMM::SingleClassHMM	(	std::shared_ptr< SharedParameters >	p,
		Json::Value const &	root
	)

explicit

Copy constructor.

Parameters

p	pointer to a shared parameters object (owned by a Model)
root	Json structure

Member Function Documentation

std::string xmm::Writable::__str__ ( ) const

inlineinherited

"print" method for python => returns the results of write method

Warning: only defined if SWIGPYTHON is defined

void xmm::SingleClassHMM::addCyclicTransition ( double proba )

protected

Adds a cyclic Transition probability (from last state to first state)

avoids getting stuck at the end of the model. this method is idle for a hierarchical model.

Parameters

proba probability of the transition form last to first state

void xmm::SingleClassHMM::addExitPoint	(	int	stateIndex,
		float	proba
	)

Set the exit probability of a specific state.

this method is only active in 'HIERARCHICAL' mode. The probability

Parameters

stateIndex	index of the state to add the exit point
proba	probability to exit the gesture from this state

Exceptions

runtime_error	if the model is not hierarchical
out_of_range	if the state index is out of bounds

void xmm::SingleClassHMM::allocate ( )

protectedvirtual

Allocate model parameters.

Implements xmm::SingleClassProbabilisticModel.

void xmm::SingleClassHMM::backward_init ( double ct )

protected

Initialization Backward algorithm.

Parameters

ct	inverse of the likelihood at time step t computed with the forward algorithm (see Rabiner 1989)

void xmm::SingleClassHMM::backward_update	(	double	ct,
		const float *	observation,
		const float *	observation_output = `NULL`
	)

protected

Update of the Backward algorithm.

Parameters

ct	inverse of the likelihood at time step t computed with the forward algorithm (see Rabiner 1989)
observation	observation vector at time t. If the model is bimodal, this vector should be only the observation on the input modality.
observation_output	observation on the output modality (only used if the model is bimodal). If unspecified, the update is performed on the input modality only.

void xmm::SingleClassHMM::baumWelch_backward_update	(	double	ct,
		std::vector< double >::iterator	observation_likelihoods
	)

protected

Update of the Backward algorithm for Training (observation probabilities are pre-computed)

Parameters

ct	inverse of the likelihood at time step t computed with the forward algorithm (see Rabiner 1989)
observation_likelihoods	likelihoods of the observations for each state

void xmm::SingleClassHMM::baumWelch_estimateCovariances ( TrainingSet * trainingSet )

protected

Estimate the Covariances of the Gaussian Distribution for each state.

void xmm::SingleClassHMM::baumWelch_estimateMeans ( TrainingSet * trainingSet )

protected

Estimate the Means of the Gaussian Distribution for each state.

void xmm::SingleClassHMM::baumWelch_estimateMixtureCoefficients ( TrainingSet * trainingSet )

protected

Estimate the Coefficients of the Gaussian Mixture for each state.

void xmm::SingleClassHMM::baumWelch_estimatePrior ( TrainingSet * trainingSet )

protected

Estimate the Prior Probabilities.

void xmm::SingleClassHMM::baumWelch_estimateTransitions ( TrainingSet * trainingSet )

protected

Estimate the Transition Probabilities.

double xmm::SingleClassHMM::baumWelch_forward_update ( std::vector< double >::iterator observation_likelihoods )

protected

Update of the forward algorithm for Training (observation probabilities are pre-computed)

Parameters

observation_likelihoods likelihoods of the observations for each state

Returns: instantaneous likelihood

double xmm::SingleClassHMM::baumWelch_forwardBackward	(	std::shared_ptr< Phrase >	currentPhrase,
		int	phraseIndex
	)

protected

Compute the forward-backward algorithm on a phrase of the training set.

Parameters

currentPhrase	pointer to the phrase of the training set
phraseIndex	index of the phrase

Returns: lieklihood of the phrase given the model's current parameters

void xmm::SingleClassHMM::baumWelch_gammaSum ( TrainingSet * trainingSet )

protected

Compute the sum of the gamma variable (for use in EM)

void xmm::SingleClassProbabilisticModel::cancelTraining ( )

inherited

Cancels the training process : sets a flag so that the training stops at the next possible exit in the training process.

Warning: the model is still training when this function returns. This function only requests the training process to cancel.

bool xmm::SingleClassProbabilisticModel::cancelTrainingIfRequested ( )

protectedinherited

checks if a cancel request has been sent and accordingly cancels the training process

Returns: true if the training has been canceled.

void xmm::SingleClassProbabilisticModel::check_training ( ) const

inlineprotectedinherited

Checks if the model is still training.

Exceptions

runtime_error if the model is training.

static const float xmm::SingleClassHMM::DEFAULT_EXITPROBABILITY_LAST_STATE ( )

inlinestatic

bool xmm::SingleClassProbabilisticModel::emAlgorithmHasConverged	(	int	step,
		double	log_prob,
		double	old_log_prob
	)		const

protectedinherited

checks if the training has converged according to the object's EM stop criterion

Parameters

step	index of the current step of the EM algorithm
log_prob	log-likelihood returned by the EM update
old_log_prob	log-likelihood returned by the EM update at the previous step

void xmm::SingleClassHMM::emAlgorithmInit ( TrainingSet * trainingSet )

protectedvirtual

Initialization of the parameters before training.

Implements xmm::SingleClassProbabilisticModel.

void xmm::SingleClassHMM::emAlgorithmTerminate ( )

protectedvirtual

Termination of the training algorithm.

Reimplemented from xmm::SingleClassProbabilisticModel.

double xmm::SingleClassHMM::emAlgorithmUpdate ( TrainingSet * trainingSet )

protectedvirtual

update method of the EM algorithm (calls Baum-Welch Algorithm)

Implements xmm::SingleClassProbabilisticModel.

double xmm::SingleClassHMM::filter ( std::vector< float > const & observation )

virtual

filters a incoming observation (performs recognition or regression)

the results of the inference process are stored in the results attribute

Parameters

observation observation vector

Returns: likelihood of the observation

Implements xmm::SingleClassProbabilisticModel.

double xmm::SingleClassHMM::forward_init	(	const float *	observation,
		const float *	observation_output = `NULL`
	)

protected

Initialization of the forward algorithm.

Parameters

observation	observation vector at time t. If the model is bimodal, this vector should be only the observation on the input modality.
observation_output	observation on the output modality (only used if the model is bimodal). If unspecified, the update is performed on the input modality only.

Returns: instantaneous likelihood

double xmm::SingleClassHMM::forward_update	(	const float *	observation,
		const float *	observation_output = `NULL`
	)

protected

Update of the forward algorithm.

Parameters

observation	observation vector at time t. If the model is bimodal, this vector should be only the observation on the input modality.
observation_output	observation on the output modality (only used if the model is bimodal). If unspecified, the update is performed on the input modality only.

Returns: instantaneous likelihood

void xmm::SingleClassHMM::fromJson ( Json::Value const & root )

virtual

Read the object from a JSON Structure.

Parameters

root	JSON value containing the object's information

Exceptions

JsonException if the JSON value has a wrong format

Implements xmm::SingleClassProbabilisticModel.

void xmm::SingleClassHMM::initCovariances_fullyObserved ( TrainingSet * trainingSet )

protected

initialize the covariances of each state with all training phrases (single gaussian)

void xmm::SingleClassHMM::initMeansCovariancesWithGMMEM ( TrainingSet * trainingSet )

protected

initialize the means and covariances of each state using GMM-EM on segments.

void xmm::SingleClassHMM::initMeansWithAllPhrases ( TrainingSet * trainingSet )

protected

initialize the means of each state with all training phrases (single gaussian)

void xmm::SingleClassHMM::initParametersToDefault ( std::vector< float > const & dataStddev )

protected

initialize model parameters to their default values

bool xmm::SingleClassProbabilisticModel::isTraining ( ) const

inherited

Checks if the model is training.

Returns: true if the model is training

void xmm::SingleClassHMM::normalizeTransitions ( )

protected

Normalize transition probabilities.

xmm::SingleClassHMM & xmm::SingleClassHMM::operator= ( SingleClassHMM const & src )

Assignment.

Parameters

src	Source Model

void xmm::Writable::readFile ( char * fileName )

inlineinherited

read method for python wrapping ('read' keyword forbidden, name has to be different)

Warning: only defined if SWIGPYTHON is defined

void xmm::SingleClassHMM::regression ( std::vector< float > const & observation_input )

protected

Compute the regression for the case of a bimodal model, given the estimated state probabilities estimated by forward algorithm.

predicted output parameters are stored in the result structure.

Parameters

observation_input observation on the input modality

void xmm::SingleClassHMM::reset ( )

virtual

Resets the fitering process (recognition or regression)

Reimplemented from xmm::SingleClassProbabilisticModel.

void xmm::SingleClassHMM::setErgodic ( )

protected

set the prior and transition matrix to ergodic

void xmm::SingleClassHMM::setLeftRight ( )

protected

set the prior and transition matrix to left-right (no state skip)

Json::Value xmm::SingleClassHMM::toJson ( ) const

virtual

Write the object to a JSON Structure.

Returns: Json value containing the object's information

Reimplemented from xmm::SingleClassProbabilisticModel.

void xmm::SingleClassProbabilisticModel::train ( TrainingSet * trainingSet )

inherited

Main training method based on the EM algorithm.

the method performs a loop over the pure virtual method emAlgorithmUpdate() until convergence. The emAlgorithmUpdate method computes both E and M steps of the EM algorithm.

Parameters

trainingSet Training Set to train the model.

See also: emAlgorithmUpdate

static const float xmm::SingleClassHMM::TRANSITION_REGULARIZATION ( )

inlinestatic

void xmm::SingleClassHMM::updateAlphaWindow ( )

protected

Estimates the likeliest state and compute the bounds of the windows over the states.

The window is centered around the likeliest state, and its size is the number of states. The window is clipped to the first and last states.

void xmm::SingleClassHMM::updateExitProbabilities ( float * _exitProbabilities = NULL )

protected

Update the exit probability vector given the probabilities.

this method is only active in 'HIERARCHICAL' mode. The probability vector defines the probability of exiting the gesture from each state. If unspecified, only the last state of the gesture has a non-zero probability.

Parameters

_exitProbabilities vector of exit probabilities (size must be nbStates)

Exceptions

runtime_error if the model is not hierarchical

void xmm::SingleClassHMM::updateResults ( )

protected

update the content of the likelihood buffer and return average likelihood.

The method also updates the cumulative log-likelihood computed over a window (cumulativeloglikelihood)

void xmm::Writable::writeFile ( char * fileName ) const

inlineinherited

write method for python wrapping ('write' keyword forbidden, name has to be different)

Warning: only defined if SWIGPYTHON is defined

Friends And Related Function Documentation

friend class HierarchicalHMM

friend

template<typename SingleClassModel , typename ModelType >

friend class Model

friend

Member Data Documentation

std::vector<double> xmm::SingleClassHMM::alpha

State probabilities estimated by the forward algorithm.

std::vector<double> xmm::SingleClassHMM::alpha_h[3]

State probabilities estimated by the hierarchical forward algorithm.

the variable is only allocated/used in hierarchical mode (see 'HIERARCHICAL' flag)

std::vector<double> xmm::SingleClassHMM::alpha_seq_

protected

Sequence of alpha (forward) probabilities.

std::vector<double> xmm::SingleClassHMM::beta_

protected

backward state probabilities

std::vector<double> xmm::SingleClassHMM::beta_seq_

protected

Sequence of beta (backward) probabilities.

bool xmm::SingleClassProbabilisticModel::cancel_training_

protectedinherited

defines if the model received a request to cancel training

std::vector<std::vector<double> > xmm::SingleClassHMM::epsilon_sequence_

protected

Sequence of Epsilon probabilities.

std::vector<float> xmm::SingleClassHMM::exit_probabilities_

protected

Exit probabilities for a hierarchical model.

bool xmm::SingleClassHMM::forward_initialized_

protected

Defines if the forward algorithm has been initialized.

std::vector<std::vector<double> > xmm::SingleClassHMM::gamma_sequence_

protected

Sequence of Gamma probabilities.

std::vector<std::vector<std::vector<double> > > xmm::SingleClassHMM::gamma_sequence_per_mixture_

protected

Sequence of Gamma probabilities for each mixture component.

std::vector<double> xmm::SingleClassHMM::gamma_sum_

protected

Used to store the sums of the gamma variable.

std::vector<double> xmm::SingleClassHMM::gamma_sum_per_mixture_

protected

Used to store the sums of the gamma variable for each mixture component.

bool xmm::SingleClassHMM::is_hierarchical_

protected

Defines if the model is a submodel of a hierarchical HMM.

in practice this adds exit probabilities to each state. These probabilities are set to the last state by default.

bool xmm::SingleClassProbabilisticModel::is_training_

protectedinherited

defines if the model is being trained.

std::string xmm::SingleClassProbabilisticModel::label

inherited

label associated with the given model

CircularBuffer<double> xmm::SingleClassProbabilisticModel::likelihood_buffer_

protectedinherited

Likelihood buffer used for smoothing.

ClassParameters<HMM> xmm::SingleClassHMM::parameters

Convert to bimodal HMM in place.

Parameters

dimension_input dimension of the input modality

Exceptions

runtime_error	if the model is already bimodal
out_of_range	if the requested input dimension is too large Convert to unimodal HMM in place
runtime_error	if the model is already unimodal extract a submodel with the given columns

Parameters

columns columns indices in the target order

Exceptions

runtime_error	if the model is training
out_of_range	if the number or indices of the requested columns exceeds the current dimension

Returns: a HMM from the current model considering only the target columns extract the submodel of the input modality

Exceptions

runtime_error if the model is training or if it is not bimodal

Returns: a unimodal HMM of the input modality from the current bimodal model extract the submodel of the output modality

Exceptions

runtime_error if the model is training or if it is not bimodal

Returns: a unimodal HMM of the output modality from the current bimodal model extract the model with reversed input and output modalities

Exceptions

runtime_error if the model is training or if it is not bimodal

Returns: a bimodal HMM that swaps the input and output modalities Model Parameters

std::vector<double> xmm::SingleClassHMM::previous_alpha_

protected

used to store the alpha estimated at the previous time step

std::vector<double> xmm::SingleClassHMM::previous_beta_

protected

used to store the beta estimated at the previous time step

std::vector<float> xmm::SingleClassHMM::prior

Prior probabilities.

ClassResults<HMM> xmm::SingleClassHMM::results

Results of the filtering process (recognition & regression)

std::shared_ptr<SharedParameters> xmm::SingleClassProbabilisticModel::shared_parameters

inherited

Pointer to the shared parameters owned by a multi-class model.

std::vector<SingleClassGMM> xmm::SingleClassHMM::states

States of the model (Gaussian Mixture Models)

EventGenerator<TrainingEvent> xmm::SingleClassProbabilisticModel::training_events

inherited

Generator for events monitoring the training process.

std::mutex xmm::SingleClassProbabilisticModel::training_mutex_

protectedinherited

Mutex used in Concurrent Mode.

TrainingEvent xmm::SingleClassProbabilisticModel::training_status

inherited

Event containing information on the current status of the training process.

std::vector<float> xmm::SingleClassHMM::transition

Transition Matrix.

int xmm::SingleClassHMM::window_maxindex_

protected

minimum index of the alpha window (used for regression & time progression)

int xmm::SingleClassHMM::window_minindex_

protected

minimum index of the alpha window (used for regression & time progression)

double xmm::SingleClassHMM::window_normalization_constant_

protected

normalization constant of the alpha window (used for regression & time progression)

The documentation for this class was generated from the following files:

XMM - Probabilistic Models for Motion Recognition and Mapping

Public Member Functions

Static Public Member Functions

Public Attributes

Protected Member Functions

Protected Attributes

Friends

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

Friends And Related Function Documentation

Member Data Documentation