gfsmTrain.h File Reference

transducer training from (input,output) string pairs More...

#include <gfsmAutomaton.h>
#include <gfsmUtils.h>

Include dependency graph for gfsmTrain.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Data Structures
struct	gfsmTrainer
	top-level type for automaton training data and options More...
struct	gfsmTrainConfig_
	low-level type for gfsm_automaton_train() computation state More...

Typedefs
typedef struct gfsmTrainConfig_	gfsmTrainConfig
	low-level type for gfsm_automaton_train() computation state
typedef GPtrArray	gfsmTrainPath
	low-level type for a successful training-path: GPtrArray* of gfsmTrainConfig*

Functions
gfsmTrainer *	gfsm_trainer_new (gfsmAutomaton *fst)
void	gfsm_trainer_free (gfsmTrainer *trainer, gboolean free_fst)
gfsmAutomaton *	gfsm_trainer_automaton (gfsmTrainer *trainer, gboolean steal)
void	gfsm_trainer_train (gfsmTrainer *trainer, gfsmLabelVector *input, gfsmLabelVector *output)

Detailed Description

Typedef Documentation

typedef struct gfsmTrainConfig_ gfsmTrainConfig

typedef GPtrArray gfsmTrainPath

Function Documentation

gfsmTrainer* gfsm_trainer_new

(

gfsmAutomaton *

fst

)

allocate a new gfsmTrainer for training source FST fst

Parameters

fst	automaton to be trained: it will be destructively altered by this call (weights changed to indices)

Returns

new gfsmTrainer

void gfsm_trainer_free	(	gfsmTrainer *	trainer,
		gboolean	free_fst
	)

free a gfsmTrainer

Parameters

free_fst

if true, the underlying FST will also be freed if still present.

gfsmAutomaton* gfsm_trainer_automaton	(	gfsmTrainer *	trainer,
		gboolean	steal
	)

extract an automaton representing the training counts from a gfsmTrainer trainer.

Parameters

trainer	trainer from which to extract FST topology and counts
steal	if true, trainer->fst will be used directly, invalidating trainer for any futher calls except for gfsm_trainer_free().

Returns

modified trainer->fst if steal is true, otherwise a clone whose weights are counts as stored in trainer->counts

void gfsm_trainer_train	(	gfsmTrainer *	trainer,
		gfsmLabelVector *	input,
		gfsmLabelVector *	output
	)

transducer weight-training utility: count successful paths in training FST for the i/o pair (input, output), updating trainer->counts corresponding to fst arc- and final-weights. On completion, each count is the (real)-sum of its original value and the number of occurrences of the corresponding arc (rsp. final state) in any successful path in fst with labels (input, output). May be run multiple times on the same trainer to add counts for multiple (input, output) pairs.

Parameters

trainer	transducer trainer, which should have been initialized as by gfsm_trainer_new()
input	input labels (lower)
output	output labels (upper)