libsent/src/phmm/gprune_beam.c

説明を見る。

/*
 * Copyright (c) 1991-2011 Kawahara Lab., Kyoto University
 * Copyright (c) 2000-2005 Shikano Lab., Nara Institute of Science and Technology
 * Copyright (c) 2005-2011 Julius project team, Nagoya Institute of Technology
 * All rights reserved
 */

#include <sent/stddefs.h>
#include <sent/htk_hmm.h>
#include <sent/htk_param.h>
#include <sent/hmm.h>
#include <sent/hmm_calc.h>

#define TEST2

/*

  best_mixtures_on_last_frame[]
  
  dim:  0 1 2 3 4 .... veclen-1    -> sum up
 ================================
  thres
 --------------------------------
  mix1  | |
  mix2  | |
  mix3  v v
  ...
  mixN  
 ================================
         \_\_ vecprob[0],vecprob[1]

  algorithm 1:
         
  foreach dim {
     foreach all_mixtures in best_mixtures_on_last_frame {
        compute score
     }
     threshold = the current lowest score + beam_width?
     foreach rest_mixtures {
        if (already marked as pruned at previous dim) {
           skip
        }
        compute score
        if (score < threshold) {
           mark as pruned
           skip
        }
        if (score > threshold) {
           update threshold
        }
     }
  }

  algorithm 2:

  foreach all_mixtures in best_mixtures_on_last_frame {
     foreach dim {
       compute score
       if (threshold[dim] < score) update
     }
     threshold[dim] += beam_width
  }
  foreach rest_mixtures {
     foreach dim {
        compute score
        if (score < threshold[dim]) skip this mixture
        update thres
     }
  }
     
*/

static void
clear_dimthres(HMMWork *wrk)
{
  int i;
  for(i=0;i<wrk->dimthres_num;i++) wrk->dimthres[i] = 0.0;
}

static void
set_dimthres(HMMWork *wrk)
{
  int i;
  for(i=0;i<wrk->dimthres_num;i++) wrk->dimthres[i] += TMBEAMWIDTH;
}

static LOGPROB
compute_g_beam_updating(HMMWork *wrk, HTK_HMM_Dens *binfo)
{
  VECT tmp, x;
  VECT *mean;
  VECT *var;
  VECT *th = wrk->dimthres;
  VECT *vec = wrk->OP_vec;
  short veclen = wrk->OP_veclen;

#ifndef TEST2
  if (binfo == NULL) return(LOG_ZERO);
#endif

  mean = binfo->mean;
  var = binfo->var->vec;

  tmp = 0.0;
  for (; veclen > 0; veclen--) {
    x = *(vec++) - *(mean++);
    tmp += x * x * *(var++);
    if ( *th < tmp) *th = tmp;
    th++;
  }
  return((tmp + binfo->gconst) * -0.5);
}

static LOGPROB
compute_g_beam_pruning(HMMWork *wrk, HTK_HMM_Dens *binfo)
{
  VECT tmp, x;
  VECT *mean;
  VECT *var;
  VECT *th = wrk->dimthres;
  VECT *vec = wrk->OP_vec;
  short veclen = wrk->OP_veclen;

#ifndef TEST2
  if (binfo == NULL) return(LOG_ZERO);
#endif
  mean = binfo->mean;
  var = binfo->var->vec;

  tmp = 0.0;
  for (; veclen > 0; veclen--) {
    x = *(vec++) - *(mean++);
    tmp += x * x * *(var++);
    if ( tmp > *(th++)) {
      return LOG_ZERO;
    }
  }
  return((tmp + binfo->gconst) * -0.5);
}


boolean
gprune_beam_init(HMMWork *wrk)
{
  int i;

  /* maximum Gaussian set size = maximum mixture size * nstream */
  wrk->OP_calced_maxnum = wrk->OP_hmminfo->maxmixturenum * wrk->OP_nstream;
  wrk->OP_calced_score = (LOGPROB *)mymalloc(sizeof(LOGPROB) * wrk->OP_calced_maxnum);
  wrk->OP_calced_id = (int *)mymalloc(sizeof(int) * wrk->OP_calced_maxnum);
  wrk->mixcalced = (boolean *)mymalloc(sizeof(int) * wrk->OP_calced_maxnum);
  for(i=0;i<wrk->OP_calced_maxnum;i++) wrk->mixcalced[i] = FALSE;
  wrk->dimthres_num = wrk->OP_hmminfo->opt.vec_size;
  wrk->dimthres = (LOGPROB *)mymalloc(sizeof(LOGPROB) * wrk->dimthres_num);

  return TRUE;
}

void
gprune_beam_free(HMMWork *wrk)
{
  free(wrk->OP_calced_score);
  free(wrk->OP_calced_id);
  free(wrk->mixcalced);
  free(wrk->dimthres);
}

void
gprune_beam(HMMWork *wrk, HTK_HMM_Dens **g, int gnum, int *last_id, int lnum)
{
  int i, j, num = 0;
  LOGPROB score, thres;

  if (last_id != NULL) {        /* compute them first to form thresholds */
    /* 1. clear dimthres */
    clear_dimthres(wrk);
    /* 2. calculate first $OP_gprune_num and set initial thresholds */
    for (j=0; j<lnum; j++) {
      i = last_id[j];
#ifdef TEST2
      if (!g[i]) {
        score = LOG_ZERO;
      } else {
        score = compute_g_beam_updating(wrk, g[i]);
      }
      num = cache_push(wrk, i, score, num);
#else
      score = compute_g_beam_updating(wrk, g[i]);
      num = cache_push(wrk, i, score, num);
#endif
      wrk->mixcalced[i] = TRUE;      /* mark them as calculated */
    }
    /* 3. set pruning thresholds for each dimension */
    set_dimthres(wrk);

    /* 4. calculate the rest with pruning*/
    for (i = 0; i < gnum; i++) {
      /* skip calced ones in 1. */
      if (wrk->mixcalced[i]) {
        wrk->mixcalced[i] = FALSE;
        continue;
      }
#ifdef TEST2
      /* compute with safe pruning */
      if (!g[i]) continue;
      score = compute_g_beam_pruning(wrk, g[i]);
      if (score > LOG_ZERO) {
        num = cache_push(wrk, i, score, num);
      }
#else
      /* compute with safe pruning */
      score = compute_g_beam_pruning(wrk, g[i]);
      if (score > LOG_ZERO) {
        num = cache_push(wrk, i, score, num);
      }
#endif
    }
  } else {                      /* in case the last_id not available */
    /* at the first 0 frame */
    /* calculate with safe pruning */
    thres = LOG_ZERO;
    for (i = 0; i < gnum; i++) {
      if (num < wrk->OP_gprune_num) {
        score = compute_g_base(wrk, g[i]);
      } else {
        score = compute_g_safe(wrk, g[i], thres);
        if (score <= thres) continue;
      }
      num = cache_push(wrk, i, score, num);
      thres = wrk->OP_calced_score[num-1];
    }
  }
  wrk->OP_calced_num = num;
}