lti::hmmTrainer::parameters Class Reference

the parameters for the class hmmTrainer More...

#include <ltiHmmTrainer.h>

Inheritance diagram for lti::hmmTrainer::parameters:

[legend]

Collaboration diagram for lti::hmmTrainer::parameters:

[legend]

List of all members.

Public Types
enum	poolingType { noPooling, statePooling, modelPooling }
enum	estimatorType { estimMaximumLikelihood, estimStandardDeviation, estimMeanDeviation, estimMeanDeviationRoot }
enum	{ minimumSequenceLength = 0, averageSequenceLength = -1 }
Public Member Functions
	parameters ()
	parameters (const parameters &other)
	~parameters ()
const char *	getTypeName () const
parameters &	copy (const parameters &other)
parameters &	operator= (const parameters &other)
virtual functor::parameters *	clone () const
virtual bool	write (ioHandler &handler, const bool complete=true) const
virtual bool	read (ioHandler &handler, const bool complete=true)
Public Attributes
Training termination
These members are the training iteration control parameters. For experienced users!
int	maxIterations
double	convergenceThreshold
HMM attributes
These members are the model attributes.
int	numberOfStates
int	minimumJump
int	maximumJump
dvector	initialScore
Emission density
Parameters defining the kind of emission density and the estimator to use.
hiddenMarkovModel::scoreType	scoreFunction
estimatorType	estimatorFunction
double	minScalingFactor
Splitting (mixture densities)
These members can be adjusted to change the splitting process, where the density functions are successively divided into smaller ones to describe the actual distribution better. For intermediate users!
int	maxDensities
int	reassignmentIterations
int	firstSplit
poolingType	pooling
int	minFramesForSplit

Detailed Description

the parameters for the class hmmTrainer

Member Enumeration Documentation

anonymous enum

special definitions for numberOfStates

Enumerator:

minimumSequenceLength	uses the minimum length of all sequences as number of states
averageSequenceLength	uses the average length of all sequences as number of states

enum lti::hmmTrainer::parameters::estimatorType

Estimator types:.

Enumerator:

estimMaximumLikelihood	select estimator for given score function (default)
estimStandardDeviation	standard deviation (gaussian maximum-likelihood)
estimMeanDeviation	mean absolute deviation (laplacian maximum-likelihood)
estimMeanDeviationRoot	root of mean absolute deviation (laplacian maximum-likelihood)

enum lti::hmmTrainer::parameters::poolingType

Pooling modes.

Enumerator:

noPooling	don't pool variances
statePooling	pool variances within one state
modelPooling	pool variances of a dimension for all states

Constructor & Destructor Documentation

lti::hmmTrainer::parameters::parameters ( )

default constructor

Reimplemented from lti::functor::parameters.

lti::hmmTrainer::parameters::parameters ( const parameters & other )

copy constructor

Parameters:

other

the parameters object to be copied

lti::hmmTrainer::parameters::~parameters ( ) [virtual]

destructor

Reimplemented from lti::functor::parameters.

Member Function Documentation

virtual functor::parameters* lti::hmmTrainer::parameters::clone ( ) const [virtual]

returns a pointer to a clone of the parameters

Implements lti::functor::parameters.

parameters& lti::hmmTrainer::parameters::copy ( const parameters & other )

copy the contents of a parameters object

Parameters:

other

the parameters object to be copied

Returns:: a reference to this parameters object

const char* lti::hmmTrainer::parameters::getTypeName ( ) const [virtual]

returns name of this type

Reimplemented from lti::functor::parameters.

parameters& lti::hmmTrainer::parameters::operator= ( const parameters & other )

copy the contents of a parameters object

Parameters:

other

the parameters object to be copied

Returns:: a reference to this parameters object

virtual bool lti::hmmTrainer::parameters::read	(	ioHandler &	handler,
		const bool	complete = `true`
	)			`[virtual]`

read the parameters from the given ioHandler

Parameters:

	handler	the ioHandler to be used
	complete	if true (the default) the enclosing begin/end will be also written, otherwise only the data block will be written.

Returns:: true if write was successful

Reimplemented from lti::functor::parameters.

virtual bool lti::hmmTrainer::parameters::write	(	ioHandler &	handler,
		const bool	complete = `true`
	)			const `[virtual]`

write the parameters in the given ioHandler

Parameters:

	handler	the ioHandler to be used
	complete	if true (the default) the enclosing begin/end will be also written, otherwise only the data block will be written.

Returns:: true if write was successful

Reimplemented from lti::functor::parameters.

Member Data Documentation

double lti::hmmTrainer::parameters::convergenceThreshold

Estimation of the HMM parameters is finished, when score enhancement within one iteration is below or equal to convergence threshold for given training data.

Default is 0.0 (iterate, while model score is improving)

estimatorType lti::hmmTrainer::parameters::estimatorFunction

Method for estimating the scaling factor (ie spread) of the density functions.

Default: estimMaximumLikelihood

int lti::hmmTrainer::parameters::firstSplit

Number of iteration, where splitting is taken into consideration for the first time.

By default, this is set to 0.

dvector lti::hmmTrainer::parameters::initialScore

Vector of initial scores.

If the size of this vector differs from numberOfStates, the default initialScore is used:





state 0 
other states  


score
0
oo   


(prob) 
(1) 
(0)

Remember to specify the initialScore as scores (i.e. ~ -ln(p)). Note that the initialScore vector doesn't have to be normalized.

int lti::hmmTrainer::parameters::maxDensities

Maximum allowed number of densities per state.

Default = 5

int lti::hmmTrainer::parameters::maximumJump

Maximum relative forward jump within HMM.

Default: 2

int lti::hmmTrainer::parameters::maxIterations

Maximum number of training iterations.

Default is 0 (iterate, until convergenceThreshold is reached).

int lti::hmmTrainer::parameters::minFramesForSplit

Minimum number of observations assigned to a density to be splittable.

Default: 3

int lti::hmmTrainer::parameters::minimumJump

Minimum relative forward jump within HMM.

Default: 0 , i.e. loop to current state

double lti::hmmTrainer::parameters::minScalingFactor

Minimum scaling factor.

This is required to avoid very narrow and high density functions, which are obtained, when variation in data is too little.

Default is 0.1

int lti::hmmTrainer::parameters::numberOfStates

Number of states of the model to be generated.

A value of minimumSequenceLength (0) indicates to pick the size of the shortest observation sequence, averageSequenceLength (-1) picks the average size of all considered observation sequences.

Default: minimumSequenceLength (0)

poolingType lti::hmmTrainer::parameters::pooling

Pooling means to compute a common scaling factor (spread, variance) for several densities.

You can select to have no pooling, state pooling or model pooling.

Default: noPooling

int lti::hmmTrainer::parameters::reassignmentIterations

Number of iterations between two split processes.

A value of 0 will reassign observations until no further changes occur (nope, not yet!).

The default value is 5 (split densities at every 5th iteration).

hiddenMarkovModel::scoreType lti::hmmTrainer::parameters::scoreFunction

Density function type to be used for the emission scores.

Default: laplaceScore