spr_pylib.train.master.MasterTrainer Class Reference

Defines the trainer object that will perform the various training steps. More...

Inheritance diagram for spr_pylib.train.master.MasterTrainer:

Collaboration diagram for spr_pylib.train.master.MasterTrainer:

Public Member Functions
def	__init__
	Constructor. More...
def	mvg_init
	Create a new mvg file based on collected data (not yet used. More...
def	tied
	Create + train models with tied gaussians (reduced semi-continuous) More...
def	phtied
	Create + train models with gaussians tied within the monophones only (reduced semi-continuous) More...
def	untied_new
	Create + train models with untied gaussians. More...
def	untied
	Create + train models with untied gaussians. More...
def	fvg
	Add a single full covariance matrix to the tied gaussian models. More...
def	cdtree
	Create context dependent models by means of a descision tree. More...
def	vitalign
	Create a new segmentation based on the model created in the previous training step. More...
def	install
	Copy the most current model files to the specified location. More...

Detailed Description

Defines the trainer object that will perform the various training steps.

Constructor & Destructor Documentation

def spr_pylib.train.master.MasterTrainer.__init__	(		self,
			config
	)

Constructor.

Parameters

self	The callee object.
config	The MasterConfig object holding all configuration data.

Member Function Documentation

def spr_pylib.train.master.MasterTrainer.mvg_init	(		self,
			mvg_file,
			Ngauss,
			trk_file,
			siz_file
	)

Create a new mvg file based on collected data (not yet used.

probably wrong location)

Parameters

self	The callee object.
mvg_file	The destiantion file
Ngauss	Number of Gaussians per mixture (int) or file name containing state dependent values
trk_file	The collected data
siz_file	The data partitioning

def spr_pylib.train.master.MasterTrainer.tied	(		self,
			niter = 3,
			tng = "auto",
			mfpg = "auto",
			cfpg = "auto",
			ngps = "auto",
			mxszi = "auto",
			mxszs = "auto",
			ocs = "auto",
			ocv = "auto",
			_type = "tied"
	)

Create + train models with tied gaussians (reduced semi-continuous)

Parameters

self	The callee object.
niter	The number of training iterations (excluding the initial segpass).
tng	Initial gaussians : total number of gaussians (maximum).
mfpg	Initial gaussians : create one gaussian per mfpg points.
cfpg	Initial gaussians : collect at max cfpg points per gaussians to estimate.
ngps	Initial gaussians : create at max ngps gaussians for a state.
mxszi	Initial HMM : make gaussians mixtures with mxszi gaussians per state.
mxszs	Before vitpass1 : limit the number of gaussians per state to mxszs .
ocs	Before vitpass1 : occupancy based culling of the gaussians weights.
ocv	Before vitpass2,3,... : occupancy based culling of the gaussians weights.
_type	Internal use only. Do not use.

def spr_pylib.train.master.MasterTrainer.phtied	(		self,
			niter = 3,
			tng = "auto",
			mfpg = "auto",
			cfpg = "auto",
			ngps = "auto",
			mxszi = "auto",
			mxszs = "auto",
			ocs = "auto",
			ocv = "auto"
	)

Create + train models with gaussians tied within the monophones only (reduced semi-continuous)

Parameters

self	The callee object.
niter	The number of training iterations (excluding the initial segpass).
tng	Initial gaussians : total number of gaussians (maximum).
mfpg	Initial gaussians : create one gaussian per mfpg points.
cfpg	Initial gaussians : collect at max cfpg points per gaussians to estimate.
ngps	Initial gaussians : create at max ngps gaussians for a state.
mxszi	Initial HMM : make gaussians mixtures with mxszi gaussians per state.
mxszs	Before vitpass1 : limit the number of gaussians per state to mxszs .
ocs	Before vitpass1 : occupancy based culling of the gaussians weights.
ocv	Before vitpass2,3,... : occupancy based culling of the gaussians weights.

def spr_pylib.train.master.MasterTrainer.untied_new	(		self,
			niter = 3,
			tng = "auto",
			mfpg = "auto",
			cfpg = "auto",
			ngps = "auto",
			mxszi = "auto",
			mxszs = "auto",
			ocs = "auto",
			ocv = "auto"
	)

Create + train models with untied gaussians.

(Limited number of gaussians)

Parameters

self	The callee object.
niter	The number of training iterations (excluding the initial segpass).
tng	Initial gaussians : total number of gaussians (maximum).
mfpg	Initial gaussians : create one gaussian per mfpg points.
cfpg	Initial gaussians : collect at max cfpg points per gaussians to estimate.
ngps	Initial gaussians : create at max ngps gaussians for a state.
mxszi	Initial HMM : make gaussians mixtures with mxszi gaussians per state.
mxszs	Before vitpass1 : limit the number of gaussians per state to mxszs .
ocs	Before vitpass1 : occupancy based culling of the gaussians weights.
ocv	Before vitpass2,3,... : occupancy based culling of the gaussians weights.

def spr_pylib.train.master.MasterTrainer.untied	(		self,
			niter = 3,
			tng = "auto",
			mfpg = "auto",
			cfpg = "auto",
			ngps = "auto"
	)

Create + train models with untied gaussians.

(Fixed number of gaussians)

If ngps and mfpg conflict, a balancing act is done.

Parameters

self	The callee object.
niter	The number of training iterations.
tng	Total number of gaussians (maximum).
mfpg	Limit in the number of frames needed to make a gaussian.
cfpg	Collect at max cfpg points per gaussians to estimate.
ngps	Make gaussians mixtures with ngps gaussians per state.

def spr_pylib.train.master.MasterTrainer.fvg	(		self,
			niter = 3,
			ocv = "auto"
	)

Add a single full covariance matrix to the tied gaussian models.

The iterations specified with the niter argument will alternate between two different functionalities.

• odd iteration : estimate the full covariance matrix and update the gaussian set
• even iteration : re-estimate the mixture weights and gaussian set (vitpass)

Parameters

self	The callee object.
niter	The number of training iterations.
ocv	Occupancy based culling of the gaussians weights before a vitpass.

def spr_pylib.train.master.MasterTrainer.cdtree	(		self,
			niter = 3,
			ocs = "auto",
			ocv = "auto",
			oci = "pct.4",
			dtmc = 512,
			dtlt = 512
	)

Create context dependent models by means of a descision tree.

Parameters

self	The callee object.
niter	The number of training iterations (excluding the initial segpass with untied states).
ocs	Before vitpass1 : prune gaussians weights if occupancy is less than <p4>.
ocv	Before vitpass* : prune gaussians weights if occupancy is less than <p5>.
oci	Untied model : occupancy based culling of the gaussians weights
dtmc	Descision tree : minimum count.
dtlt	Descision tree : likelihood threshold

def spr_pylib.train.master.MasterTrainer.vitalign

(

self

)

Create a new segmentation based on the model created in the previous training step.

This can not be called as first step in a training setup.

Parameters

self	The callee object.

def spr_pylib.train.master.MasterTrainer.install	(		self,
			destdir = "result",
			name = "acmod"
	)

Copy the most current model files to the specified location.

Parameters

self	The callee object
destdir	The optional result directory, relative to the base dir
name	The optional name without extension for the various files