libsent/src/ngram/ngram_read_bin.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/ngram2.h>

static int file_version;  
static boolean need_swap; 
#ifdef WORDS_INT
static boolean need_conv;       
static boolean words_int_retry = FALSE; 
#endif

#define rdn(A,B,C,D) if (rdnfunc(A,B,C,D) == FALSE) return FALSE
#define rdn_wordid(A,B,C,D) if (rdn_wordid_func(A,B,C,D) == FALSE) return FALSE

static boolean
rdnfunc(FILE *fp, void *buf, size_t unitbyte, size_t unitnum)
{
  size_t tmp;
  if ((tmp = myfread(buf, unitbyte, unitnum, fp)) < unitnum) {
    jlog("Error: ngram_read_bin: failed to read %d bytes\n", unitbyte*unitnum);
    return FALSE;
  }
  if (need_swap) {
    if (unitbyte != 1) {
      swap_bytes(buf, unitbyte, unitnum);
    }
  }
  return TRUE;
}

#ifdef WORDS_INT

static boolean
rdn_wordid_func(FILE *fp, void *buf, int unitnum, boolean need_conv)
{
  int i;
  unsigned short *s;
  WORD_ID *t;
  WORD_ID d;

  if (need_conv) {
    /* read unsigned short units */
    rdn(fp, buf, sizeof(unsigned short), unitnum);
    /* convert them to WORD_ID (integer) */
    for(i=unitnum-1;i>=0;i--) {
      s = (unsigned short *)buf + i;
      t = (WORD_ID *)buf + i;
      d = *s;
      *t = d;
    }
  } else {
    /* read as usual */
    rdn(fp, buf, sizeof(WORD_ID), unitnum);
  }
  return TRUE;
}
#endif

static boolean
check_header(FILE *fp)
{
  char buf[BINGRAM_HDSIZE], *p;

  rdn(fp, buf, 1, BINGRAM_HDSIZE);
  p = buf;
#ifdef WORDS_INT
  need_conv = FALSE;
#endif

  /* version check */
  if (strnmatch(p, BINGRAM_IDSTR, strlen(BINGRAM_IDSTR))) {
    /* bingram file made by mkbingram before 3.4.2 */
    file_version = 3;
    p += strlen(BINGRAM_IDSTR) + 1;
  } else if (strnmatch(p, BINGRAM_IDSTR_V4, strlen(BINGRAM_IDSTR_V4))) {
    /* bingram file made by mkbingram later than 3.5 */
    file_version = 4;
    p += strlen(BINGRAM_IDSTR_V4) + 1;
  } else if (strnmatch(p, BINGRAM_IDSTR_V5, strlen(BINGRAM_IDSTR_V5))) {
    /* bingram file made by JuliusLib-4 and later */
    file_version = 5;
    p += strlen(BINGRAM_IDSTR_V5) + 1;
  } else {
    /* not a bingram file */
    jlog("Error: ngram_read_bin: invalid header\n");
    return FALSE;
  }
  /* word size check (for bingram build by mkbingram 3.3p5 and later */
  if (strnmatch(p, BINGRAM_SIZESTR_HEAD, strlen(BINGRAM_SIZESTR_HEAD))) {
    p += strlen(BINGRAM_SIZESTR_HEAD);
    if (! strnmatch(p, BINGRAM_SIZESTR_BODY, strlen(BINGRAM_SIZESTR_BODY))) {
      /* word size does not match (int / short) */
#ifdef WORDS_INT
      if (strnmatch(p, BINGRAM_SIZESTR_BODY_2BYTE, strlen(BINGRAM_SIZESTR_BODY_2BYTE))) {
        /* this is 2-byte word ID, will convert while reading */
        jlog("Warning: ngram_read_bin: 2-bytes bingram, converting to 4 bytes\n");
        need_conv = TRUE;
        p += strlen(BINGRAM_SIZESTR_BODY_2BYTE) + 1;
      } else {
        jlog("Error: ngram_read_bin: unknown word byte size!\n");
        return FALSE;
      }
#else
      if (strnmatch(p, BINGRAM_SIZESTR_BODY_4BYTE, strlen(BINGRAM_SIZESTR_BODY_4BYTE))) {
        /*** 4bytes to 2bytes not implemented, just terminate here... ***/
        jlog("Error: ngram_read_bin: cannot handle 4-bytes bingram\n");
        jlog("Error: ngram_read_bin: please use Julius compiled with --enable-words-int\n");
        return FALSE;
        //p += strlen(BINGRAM_SIZESTR_BODY_4BYTE) + 1;
      } else {
        jlog("Error: ngram_read_bin: unknown word byte size!\n");
        return FALSE;
      }
#endif
    } else {
      p += strlen(BINGRAM_SIZESTR_BODY) + 1;
    }

    /* byte order check (v4 (rev.3.5) and later) */
    if (file_version >= 4) {
      if (!strnmatch(p, BINGRAM_BYTEORDER_HEAD, strlen(BINGRAM_BYTEORDER_HEAD))) {
        jlog("Error: ngram_read_bin: no information for byte order??\n");
        return FALSE;
      }
      p += strlen(BINGRAM_BYTEORDER_HEAD);
      if (! strnmatch(p, BINGRAM_NATURAL_BYTEORDER, strlen(BINGRAM_NATURAL_BYTEORDER))) {
        /* file endian and running endian is different, need swapping */
        need_swap = TRUE;
      } else {
        need_swap = FALSE;
      }
      p += strlen(BINGRAM_NATURAL_BYTEORDER) + 1;
    }
  } /* if no BINGRAM_SIZESTR_HEAD found, just pass it */

  /* in case of V3 bingram file, the unit size of word_id and its byte order
     cannot be determined from the header.  In that case, we assume 
     byteorder to be a BIG ENDIAN.  The word_id unit size (2byte in normal,
     or 4byte if bingram generated with mkbingram with --enable-words-int)
     will be automagically detected.
     */

  if (file_version < 4) {
    /* assume input as big endian */
#ifdef WORDS_BIGENDIAN
    need_swap = FALSE;
#else
    need_swap = TRUE;
#endif
  }
    
  /*jlog("%s",buf);*/

  return TRUE;
}

static boolean
ngram_read_bin_v5(FILE *fp, NGRAM_INFO *ndata)
{
  int i,n,len;
  char *w, *p;
#ifdef WORDS_INT
  unsigned short *buf;
#endif
  NGRAM_TUPLE_INFO *t;

  /* read some info extended from version 5 */
  rdn(fp, &(ndata->n), sizeof(int), 1);
  rdn(fp, &(ndata->dir), sizeof(int), 1);
  rdn(fp, &(ndata->bigram_index_reversed), sizeof(boolean), 1);

  jlog("Stat: ngram_read_bin_v5: this is %s %d-gram file\n", (ndata->dir == DIR_LR) ? "forward" : "backward", ndata->n);

  /* read total info and set max_word_num */
  ndata->d = (NGRAM_TUPLE_INFO *)mymalloc(sizeof(NGRAM_TUPLE_INFO) * ndata->n);
  memset(ndata->d, 0, sizeof(NGRAM_TUPLE_INFO) * ndata->n);
  for(n=0;n<ndata->n;n++) {
    rdn(fp, &(ndata->d[n].totalnum), sizeof(NNID), 1);
  }
  ndata->max_word_num = ndata->d[0].totalnum;

  /* read wname */
  rdn(fp, &len, sizeof(int), 1);
  w = mymalloc(len);
  rdn(fp, w, 1, len);
  /* assign... */
  ndata->wname = (char **)mymalloc(sizeof(char *) * ndata->max_word_num);
  p = w; i = 0;
  while (p < w + len) {
    ndata->wname[i++] = p;
    while(*p != '\0') p++;
    p++;
  }
  if (i != ndata->max_word_num) {
    jlog("Error: ngram_read_bin_v5: wname error??\n");
    return FALSE;
  }

  /* read N-gram */
  for(n=0;n<ndata->n;n++) {
    jlog("stat: ngram_read_bin_v5: reading %d-gram\n", n+1);

    t = &(ndata->d[n]);
    
    rdn(fp, &(t->is24bit), sizeof(boolean), 1);
    rdn(fp, &(t->ct_compaction), sizeof(boolean), 1);
    rdn(fp, &(t->bgnlistlen), sizeof(NNID), 1);
    rdn(fp, &(t->context_num), sizeof(NNID), 1);

    if (n > 0) {
      if (t->is24bit) {
        t->bgn_upper = (NNID_UPPER *)mymalloc_big(sizeof(NNID_UPPER), t->bgnlistlen);
        rdn(fp, t->bgn_upper, sizeof(NNID_UPPER), t->bgnlistlen);
        t->bgn_lower = (NNID_LOWER *)mymalloc_big(sizeof(NNID_LOWER), t->bgnlistlen);
        rdn(fp, t->bgn_lower, sizeof(NNID_LOWER), t->bgnlistlen);
      } else {
        t->bgn = (NNID *)mymalloc_big(sizeof(NNID), t->bgnlistlen);
        rdn(fp, t->bgn, sizeof(NNID), t->bgnlistlen);
      }
      t->num = (WORD_ID *)mymalloc_big(sizeof(WORD_ID), t->bgnlistlen);
      rdn(fp, t->num, sizeof(WORD_ID), t->bgnlistlen);
      t->nnid2wid = (WORD_ID *)mymalloc_big(sizeof(WORD_ID), t->totalnum);
      rdn(fp, t->nnid2wid, sizeof(WORD_ID), t->totalnum);
    } else {
      t->bgn_upper = NULL;
      t->bgn_lower = NULL;
      t->bgn = NULL;
      t->num = NULL;
      t->bgnlistlen = 0;
      t->nnid2wid = NULL;
    }

    t->prob = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), t->totalnum);
    rdn(fp, t->prob, sizeof(LOGPROB), t->totalnum);

    rdn(fp, &i, sizeof(int), 1);
    if (i == 1) {
      t->bo_wt = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), t->context_num);
      rdn(fp, t->bo_wt, sizeof(LOGPROB), t->context_num);
    } else {
      t->bo_wt = NULL;
    }
    rdn(fp, &i, sizeof(int), 1);
    if (i == 1) {
      t->nnid2ctid_upper = (NNID_UPPER *)mymalloc_big(sizeof(NNID_UPPER), t->totalnum);
      t->nnid2ctid_lower = (NNID_LOWER *)mymalloc_big(sizeof(NNID_LOWER), t->totalnum);
      rdn(fp, t->nnid2ctid_upper, sizeof(NNID_UPPER), t->totalnum);
      rdn(fp, t->nnid2ctid_lower, sizeof(NNID_LOWER), t->totalnum);
    } else {
      t->nnid2ctid_upper = NULL;
      t->nnid2ctid_lower = NULL;
    }
  }
  rdn(fp, &i, sizeof(int), 1);
  if (i == 1) {
    ndata->bo_wt_1 = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), ndata->d[0].context_num);
    rdn(fp, ndata->bo_wt_1, sizeof(LOGPROB), ndata->d[0].context_num);
  } else {
    ndata->bo_wt_1 = NULL;
  }
  rdn(fp, &i, sizeof(int), 1);
  if (i == 1) {
    ndata->p_2 = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), ndata->d[1].totalnum);
    rdn(fp, ndata->p_2, sizeof(LOGPROB), ndata->d[1].totalnum);
  } else {
    ndata->p_2 = NULL;
  }

  return TRUE;
}

static boolean
ngram_read_bin_compat(FILE *fp, NGRAM_INFO *ndata, int *retry_ret)
{
  int i,n,len;
  char *w, *p;
  NNID *n3_bgn;
  NNID d, ntmp;
#ifdef WORDS_INT
  unsigned short *buf;
#endif
  NGRAM_TUPLE_INFO *t, *tt, *ttt;

  /* old binary N-gram assumes these types */
  ndata->bigram_index_reversed = TRUE;
  ndata->n = 3;
  ndata->dir = DIR_RL;

  /* read total info and set max_word_num */
  ndata->d = (NGRAM_TUPLE_INFO *)mymalloc(sizeof(NGRAM_TUPLE_INFO) * ndata->n);
  memset(ndata->d, 0, sizeof(NGRAM_TUPLE_INFO) * ndata->n);
  for(n=0;n<ndata->n;n++) {
    rdn(fp, &(ndata->d[n].totalnum), sizeof(NNID), 1);
  }
  ndata->max_word_num = ndata->d[0].totalnum;

  if (file_version == 4) {
    rdn(fp, &(ndata->d[1].context_num), sizeof(NNID), 1);
  }

  for(n=0;n<ndata->n;n++) {
    if (n < 2) {
      ndata->d[n].is24bit = FALSE;
    } else {
      if (ndata->d[n].totalnum >= NNID_MAX_24) {
        jlog("Warning: ngram_read_bin_compat: num of %d-gram exceeds 24bit, now switch to %dbit index\n", n+1, sizeof(NNID) * 8);
        ndata->d[n].is24bit = FALSE;
      } else {
        ndata->d[n].is24bit = TRUE;
      }
    }
    ndata->d[n].nnid2ctid_upper = NULL;
    ndata->d[n].nnid2ctid_lower = NULL;
  }
  /* always do back-off compaction for 3-gram and up */
  /* mark 2-gram and up */
  ndata->d[0].ct_compaction = FALSE;
  for(n=1;n<ndata->n;n++) {
    ndata->d[n].ct_compaction = TRUE;
  }

  /* read wname */
  rdn(fp, &len, sizeof(int), 1);
  w = mymalloc(len);
  rdn(fp, w, 1, len);
  /* assign... */
  ndata->wname = (char **)mymalloc(sizeof(char *) * ndata->max_word_num);
  p = w; i = 0;
  while (p < w + len) {
    ndata->wname[i++] = p;
    while(*p != '\0') p++;
    p++;
  }
  if (i != ndata->max_word_num) {
    jlog("Error: ngram_read_bin_compat: wname error??\n");
    return FALSE;
  }

  /* malloc 1-gram */
  t = &(ndata->d[0]);
  tt = &(ndata->d[1]);
  ttt = &(ndata->d[2]);

  t->bgn_upper = NULL;
  t->bgn_lower = NULL;
  t->bgn = NULL;
  t->num = NULL;
  t->bgnlistlen = 0;
  t->nnid2wid = NULL;
  t->nnid2ctid_upper = NULL;
  t->nnid2ctid_lower = NULL;

  t->context_num = t->totalnum;
  t->prob = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), t->totalnum);
  ndata->bo_wt_1 = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), t->context_num);
  t->bo_wt = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), t->context_num);
  tt->bgnlistlen = t->context_num;
  tt->bgn = (NNID *)mymalloc_big(sizeof(NNID), tt->bgnlistlen);
  tt->num = (WORD_ID *)mymalloc_big(sizeof(WORD_ID), tt->bgnlistlen);

  /* read 1-gram */
  jlog("stat: ngram_read_bin_compat: reading 1-gram\n");
  rdn(fp, t->prob, sizeof(LOGPROB), t->totalnum);
  rdn(fp, ndata->bo_wt_1, sizeof(LOGPROB), t->context_num);
  rdn(fp, t->bo_wt, sizeof(LOGPROB), t->context_num);
  rdn(fp, tt->bgn, sizeof(NNID), tt->bgnlistlen);
#ifdef WORDS_INT
  rdn_wordid(fp, tt->num, tt->bgnlistlen, need_conv);
#else
  rdn(fp, tt->num, sizeof(WORD_ID), tt->bgnlistlen);
#endif

#ifdef WORDS_INT
  {
    /* check if we are wrongly reading word_id=2byte bingram
       (if bingram version >= 4, this should not be happen because
        header correctly tells the word_id byte size.  This will 
        occur only if matches all the conditions below:
        - you run Julius with --enable-words-int,
        - you use old bingram of version <= 3, and
        - you use bingram file converted without --enable-words-int
     */
    WORD_ID w;
    for(w=0;w<ndata->max_word_num;w++) {
      if (ndata->d[1].num[w] > ndata->max_word_num) {
        if (words_int_retry) {
          jlog("Error: ngram_read_bin_compat: retry failed, wrong bingram format\n");
          return FALSE;
        }
        jlog("Warning: ngram_read_bin_compat: incorrect data, may be a 2-byte v3 bingram, retry with conversion\n");
        free(ndata->wname[0]);
        free(ndata->wname);
        free(t->prob);
        free(ndata->bo_wt_1);
        free(t->bo_wt);
        free(tt->bgn);
        free(tt->num);
        myfrewind(fp);
        words_int_retry = TRUE;
        *retry_ret = 1;
        return FALSE;
      }
    }
  }
#endif

  /* malloc the rest */
  tt->nnid2wid = (WORD_ID *)mymalloc_big(sizeof(WORD_ID), tt->totalnum);
  tt->prob = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), tt->totalnum);
  ndata->p_2 = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), tt->totalnum);
  if (file_version == 4) {      /* context compaction and 24bit */
    tt->nnid2ctid_upper = (NNID_UPPER *)mymalloc_big(sizeof(NNID_UPPER), tt->totalnum);
    tt->nnid2ctid_lower = (NNID_LOWER *)mymalloc_big(sizeof(NNID_LOWER), tt->totalnum);
    tt->bo_wt = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), tt->context_num);
    ttt->bgnlistlen = tt->context_num;
    ttt->bgn_upper = (NNID_UPPER *)mymalloc_big(sizeof(NNID_UPPER), ttt->bgnlistlen);
    ttt->bgn_lower = (NNID_LOWER *)mymalloc_big(sizeof(NNID_LOWER), ttt->bgnlistlen);
    ttt->num = (WORD_ID *)mymalloc_big(sizeof(WORD_ID), ttt->bgnlistlen);
  } else {
    tt->context_num = tt->totalnum;
    tt->bo_wt = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), tt->context_num);
    ttt->bgnlistlen = tt->context_num;
    ttt->num = (WORD_ID *)mymalloc_big(sizeof(WORD_ID), ttt->bgnlistlen);
    if (ttt->is24bit) {
      ttt->bgn_upper = (NNID_UPPER *)mymalloc_big(sizeof(NNID_UPPER), ttt->bgnlistlen);
      ttt->bgn_lower = (NNID_LOWER *)mymalloc_big(sizeof(NNID_LOWER), ttt->bgnlistlen);
      n3_bgn = (NNID *)mymalloc_big(sizeof(NNID), ttt->bgnlistlen);
    } else {
      ttt->bgn = (NNID *)mymalloc_big(sizeof(NNID), ttt->bgnlistlen);
    }
  }
      
  ttt->nnid2wid = (WORD_ID *)mymalloc_big(sizeof(WORD_ID), ttt->totalnum);
  ttt->prob = (LOGPROB *)mymalloc_big(sizeof(LOGPROB), ttt->totalnum);
  ttt->bo_wt = NULL;
  
  /* read 2-gram*/
  jlog("Stat: ngram_read_bin_compat: reading 2-gram\n");
#ifdef WORDS_INT
  rdn_wordid(fp, tt->nnid2wid, tt->totalnum, need_conv);
#else
  rdn(fp, tt->nnid2wid, sizeof(WORD_ID), tt->totalnum);
#endif
  rdn(fp, ndata->p_2, sizeof(LOGPROB), tt->totalnum);
  rdn(fp, tt->prob, sizeof(LOGPROB), tt->totalnum);
  if (file_version == 4) {
    rdn(fp, tt->nnid2ctid_upper, sizeof(NNID_UPPER), tt->totalnum);
    rdn(fp, tt->nnid2ctid_lower, sizeof(NNID_LOWER), tt->totalnum);
    rdn(fp, tt->bo_wt, sizeof(LOGPROB), tt->context_num);
    rdn(fp, ttt->bgn_upper, sizeof(NNID_UPPER), ttt->bgnlistlen);
    rdn(fp, ttt->bgn_lower, sizeof(NNID_LOWER), ttt->bgnlistlen);
#ifdef WORDS_INT
    rdn_wordid(fp, ttt->num, ttt->bgnlistlen, need_conv);
#else
    rdn(fp, ttt->num, sizeof(WORD_ID), ttt->bgnlistlen);
#endif
  } else {
    rdn(fp, tt->bo_wt, sizeof(LOGPROB), tt->context_num);
    if (ttt->is24bit) {
      rdn(fp, n3_bgn, sizeof(NNID), ttt->bgnlistlen);
      for(d=0;d<ttt->bgnlistlen;d++) {
        if (n3_bgn[d] == NNID_INVALID) {
          ttt->bgn_lower[d] = 0;
          ttt->bgn_upper[d] = NNID_INVALID_UPPER;
        } else {
          ntmp = n3_bgn[d] & 0xffff;
          ttt->bgn_lower[d] = ntmp;
          ntmp = n3_bgn[d] >> 16;
          ttt->bgn_upper[d] = ntmp;
        }
      }
    } else {
      rdn(fp, ttt->bgn, sizeof(NNID), ttt->bgnlistlen);
    }
#ifdef WORDS_INT
    rdn_wordid(fp, ttt->num, ttt->bgnlistlen, need_conv);
#else
    rdn(fp, ttt->num, sizeof(WORD_ID), ttt->bgnlistlen);
#endif
  }

  /* read 3-gram*/
  jlog("Stat: ngram_read_bin_compat: reading 3-gram\n");
#ifdef WORDS_INT
  rdn_wordid(fp, ttt->nnid2wid, ttt->totalnum, need_conv);
#else
  rdn(fp, ttt->nnid2wid, sizeof(WORD_ID), ttt->totalnum);
#endif
  rdn(fp, ttt->prob, sizeof(LOGPROB), ttt->totalnum);

  /* compact the 2-gram back-off and 3-gram links */
  if (file_version != 4) {
    if (ttt->is24bit) {
      free(n3_bgn);
      if (ngram_compact_context(ndata, 2) == FALSE) return FALSE;
    }
  }
  
  return TRUE;
}


boolean
ngram_read_bin(FILE *fp, NGRAM_INFO *ndata)
{
  int retry;

#ifdef WORDS_INT
  /* reset retry flag */
  words_int_retry = FALSE;
  /* when retrying, it restarts from here with words_int_retry = TRUE */
 ngram_read_bin_start:
#endif
  
  ndata->from_bin = TRUE;

  /* check initial header */
  if (check_header(fp) == FALSE) return FALSE;
  
#ifdef WORDS_INT
  /* in retry mode, force word_id conversion  */
  if (words_int_retry) need_conv = TRUE;
#endif
  
#ifdef WORDS_INT
  if (need_conv) jlog("Stat: ngram_read_bin: word-id size conversion enabled\n");
#endif

  if (file_version <= 4) {
    retry = 0;
    if (ngram_read_bin_compat(fp, ndata, &retry) == FALSE) {
#ifdef WORDS_INT
      if (retry == 1) {
        goto ngram_read_bin_start;
      } else {
        return FALSE;
      }
#else
      return FALSE;
#endif
    }
  } else {
    if (ngram_read_bin_v5(fp, ndata) == FALSE) return FALSE;
  }


  /* make word search tree for later lookup */
  jlog("Stat: ngram_read_bin: making entry name index\n");
  ngram_make_lookup_tree(ndata);

  bi_prob_func_set(ndata);

  return TRUE;
}