cont_mgau.h File Reference

Interface of full GMM computation with integer value of log likelihood. More...

#include <s3types.h>
#include <logmath.h>

Go to the source code of this file.

Classes

struct  mgau_t
 A single mixture-Gaussian model for one senone (see above comment). More...
struct  mgau_model_t
 The set of mixture-Gaussians in an acoustic model used in Sphinx 3.X family of tool. More...

Defines

#define CONTHMM   10001
#define SEMIHMM   10002
#define FULL_INT_COMP   20001
#define FULL_FLOAT_COMP   20002
#define MIX_INT_FLOAT_COMP   20003
#define MGAU_MEAN   1
#define MGAU_VAR   2
#define MGAU_FULLVAR   3
#define NO_BSTIDX   -1
#define NOT_UPDATED   -100
#define mgau_n_mgau(g)   ((g)->n_mgau)
#define mgau_max_comp(g)   ((g)->max_comp)
#define mgau_veclen(g)   ((g)->veclen)
#define mgau_n_comp(g, m)   ((g)->mgau[m].n_comp)
#define mgau_mean(g, m, c)   ((g)->mgau[m].mean[c])
#define mgau_var(g, m, c)   ((g)->mgau[m].var[c])
#define mgau_fullvar(g, m, c)   ((g)->mgau[m].fullvar[c])
#define mgau_mixw(g, m, c)   ((g)->mgau[m].mixw[c])
#define mgau_mixw_f(g, m, c)   ((g)->mgau[m].mixw_f[c])
#define mgau_lrd(g, m, c)   ((g)->mgau[m].lrd[c])
#define mgau_lrdi(g, m, c)   ((g)->mgau[m].lrdi[c])
#define mgau_frm_sen_eval(g)   ((g)->frm_sen_eval)
#define mgau_frm_gau_eval(g)   ((g)->frm_gau_eval)
#define mgau_frm_cisen_eval(g)   ((g)->frm_ci_sen_eval)
#define mgau_frm_cigau_eval(g)   ((g)->frm_ci_gau_eval)

Functions

S3DECODER_EXPORT mgau_model_tmgau_init (const char *meanfile, const char *varfile, float64 varfloor, const char *mixwfile, float64 mixwfloor, int32 precomp, const char *senmgau, int32 comp_type, logmath_t *logmath)
int32 mgau_var_nzvec_floor (mgau_model_t *g, float64 floor)
S3DECODER_EXPORT int32 mgau_eval (mgau_model_t *g, int32 m, int32 *active_comp, float32 *x, int32 fr, int32 bUpdBstIdx)
S3DECODER_EXPORT int32 mgau_comp_eval (mgau_model_t *g, int32 s, float32 *x, int32 *score)
int32 mgau_dump (mgau_model_t *g, int32 type)
S3DECODER_EXPORT void mgau_free (mgau_model_t *g)
S3DECODER_EXPORT int32 mgau_mean_reload (mgau_model_t *g, const char *mean_file_name)

Detailed Description

Interface of full GMM computation with integer value of log likelihood.


Define Documentation

#define CONTHMM   10001

Specification of sets of GMM

Referenced by mgau_init().

#define FULL_FLOAT_COMP   20002

Referenced by mgau_init().

#define FULL_INT_COMP   20001

Referenced by mgau_init().

#define mgau_frm_cigau_eval (  )     ((g)->frm_ci_gau_eval)

Number of CI gaussians evaluated in this frame

#define mgau_frm_cisen_eval (  )     ((g)->frm_ci_sen_eval)

Number of CI senones evaluated in this frame

#define mgau_frm_gau_eval (  )     ((g)->frm_gau_eval)

Number of gaussians evaluated in this frame

#define mgau_frm_sen_eval (  )     ((g)->frm_sen_eval)

Number of senones evaluated in this frame

#define mgau_fullvar ( g,
m,
 )     ((g)->mgau[m].fullvar[c])
#define MGAU_FULLVAR   3
#define mgau_lrd ( g,
m,
 )     ((g)->mgau[m].lrd[c])

Access the floating point version of the Gaussian constant

Access the integer version of the Gaussian constant

#define mgau_lrdi ( g,
m,
 )     ((g)->mgau[m].lrdi[c])
#define mgau_max_comp (  )     ((g)->max_comp)

Access the maximum number of components. It can be different across GMMs

Referenced by main().

#define mgau_mean ( g,
m,
 )     ((g)->mgau[m].mean[c])

Access the mean

#define MGAU_MEAN   1
#define mgau_mixw ( g,
m,
 )     ((g)->mgau[m].mixw[c])
#define mgau_mixw_f ( g,
m,
 )     ((g)->mgau[m].mixw_f[c])
#define mgau_n_comp ( g,
 )     ((g)->mgau[m].n_comp)

Access the number of component for a particular mixture in a GMM

Referenced by approx_cont_mgau_ci_eval(), main(), mgau_dump(), and mgau_var_nzvec_floor().

#define mgau_n_mgau (  )     ((g)->n_mgau)

Access macros Access number of GMMs

Referenced by main(), mgau_dump(), and mgau_var_nzvec_floor().

#define mgau_var ( g,
m,
 )     ((g)->mgau[m].var[c])

Access the variance

Referenced by main().

#define MGAU_VAR   2

Referenced by mgau_dump(), and mgau_init().

#define mgau_veclen (  )     ((g)->veclen)

Access an integer array contains the size of each stream

Referenced by main(), mgau_comp_eval(), mgau_dump(), mgau_eval(), and mgau_var_nzvec_floor().

#define MIX_INT_FLOAT_COMP   20003

Referenced by main(), mgau_eval(), and mgau_init().

#define NO_BSTIDX   -1
#define NOT_UPDATED   -100
#define SEMIHMM   10002

Referenced by mgau_init().


Function Documentation

S3DECODER_EXPORT int32 mgau_comp_eval ( mgau_model_t g,
int32  s,
float32 *  x,
int32 *  score 
)

Like mgau_eval(), but return the scores of the individual components, instead of combining them into a senone score.

Returns:
: Best component score.
Parameters:
g In: Set of mixture Gaussians
s In: Mixture being considered
x In: Input vector being compared to the components
score Out: Array of scores for each component

References mgau_model_t::distfloor, mgau_model_t::logmath, mgau_t::lrd, mgau_t::mean, mgau_model_t::mgau, mgau_veclen, mgau_t::n_comp, and mgau_t::var.

int32 mgau_dump ( mgau_model_t g,
int32  type 
)

A routine that dump all mean and variance parameters of a set of gaussian distribution.

Returns:
always 0
Parameters:
g In: Set of mixture Gaussians
type In: type of output, MGAU_MEAN for mean or MGAU_VAR for variance.

References mgau_t::mean, mgau_model_t::mgau, MGAU_MEAN, mgau_n_comp, mgau_n_mgau, MGAU_VAR, mgau_veclen, and mgau_t::var.

S3DECODER_EXPORT int32 mgau_eval ( mgau_model_t g,
int32  m,
int32 *  active,
float32 *  x,
int32  fr,
int32  update_best_id 
)

Evaluate a single mixture Gaussian at the given vector x; i.e., compute the Mahalanobis distance of x from each mean in the mixture, and combine them using the mixture weights. Return value: The final score from this evaluation (a logs3 domain value). NOTE: if the specified mixture is empty, S3_LOGPROB_ZERO is returned (see libmisc/libmisc.h).

Returns:
the senone score.

Compute the log likelihood of a Gaussian mixture model in fixed-point. Notice that within the program, the Gaussian distribution is computed using floating point. But log-add is done in fixed point (usually by table lookup).

Parameters:
g In: The entire mixture Gaussian model
m In: The chosen mixture in the model (i.e., g->mgau[m])
active In: An optional, -1 terminated list of active component indices; if non-NULL, only the specified components are used in the evaluation.
x In: Input observation vector (of length g->veclen).
fr In: Frame number where GMM m is updated
update_best_id In: Whether the best index for the GMM will be updated or not

References mgau_t::bstidx, mgau_t::bstscr, mgau_model_t::comp_type, mgau_model_t::distfloor, mgau_model_t::logmath, mgau_model_t::mgau, mgau_veclen, MIX_INT_FLOAT_COMP, NO_BSTIDX, S3_LOGPROB_ZERO, and mgau_t::updatetime.

Referenced by approx_cont_mgau_ci_eval(), and approx_cont_mgau_frame_eval().

S3DECODER_EXPORT void mgau_free ( mgau_model_t g  ) 

RAH Free memory allocated by mgau_init

ARCHAN, I noticed this program because of Ricky's comment. In 2004, a very useful tool called valgrind started to be available for Linux. This tool allows me to pick up a lot of memory problems easily.

Parameters:
g In: A set of model to free

References mgau_t::fullvar, mgau_t::lrd, mgau_t::mean, mgau_model_t::mgau, mgau_t::mixw, mgau_t::mixw_f, and mgau_t::var.

Referenced by main().

S3DECODER_EXPORT mgau_model_t* mgau_init ( const char *  meanfile,
const char *  varfile,
float64  varfloor,
const char *  mixwfile,
float64  mixwfloor,
int32  precomp,
const char *  senmgau,
int32  comp_type,
logmath_t *  logmath 
)

Create a new mixture Gaussian model from the given files (Sphinx3 format). Optionally, apply the precomputations mentioned in the main comment above.

See also:
mgau_file_read
mgau_mixw_read
mgau_uninit_compact
mgau_var_floor
mgau_precomp
Returns:
pointer to the model created if successful; NULL if error.

At the moment, S3 models have the same # of means in each codebook and 1 var/mean

Parameters:
meanfile In: File containing means of mixture gaussians
varfile In: File containing variances of mixture gaussians
varfloor In: Floor value applied to variances; e.g., 0.0001
mixwfile In: File containing mixture weights
mixwfloor In: Floor value for mixture weights; e.g., 0.0000001
precomp In: If TRUE, create and precompute mgau_t.lrd and also transform each var value to 1/(2*var). (If FALSE, one cannot use the evaluation routines provided here.)
senmgau In: type of the gaussians distribution, .cont. or .semi. FIX me! This is confusing!
comp_type In: Type of computation in this set of gaussian mixtures.

References mgau_model_t::comp_type, CONTHMM, mgau_model_t::distfloor, FULL_FLOAT_COMP, FULL_INT_COMP, mgau_model_t::gau_type, mgau_model_t::logmath, mgau_model_t::mgau, MGAU_MEAN, MGAU_VAR, MIX_INT_FLOAT_COMP, S3_LOGPROB_ZERO, S3_LOGPROB_ZERO_F, SEMIHMM, mgau_t::var, and mgau_model_t::verbose.

Referenced by main().

S3DECODER_EXPORT int32 mgau_mean_reload ( mgau_model_t g,
const char *  mean_file_name 
)

Reloading the means. This is particularly useful for speaker adaptation.

reload the mean file

Parameters:
g In/Out: The GMM
mean_file_name In: file name for the mean file.

References mgau_model_t::mgau, and MGAU_MEAN.

int32 mgau_var_nzvec_floor ( mgau_model_t g,
float64  floor 
)

Floor any variance vector that is non-zero (vector).

Returns:
No. of variance VALUES floored.
Parameters:
g In: A mixture of Gaussian components
floor In: The floor value

References mgau_model_t::mgau, mgau_n_comp, mgau_n_mgau, mgau_veclen, mgau_t::var, vector_is_zero(), and mgau_model_t::verbose.


Generated on 7 Mar 2010 by  doxygen 1.6.1