| 1 |
/* -*- c-basic-offset: 4; indent-tabs-mode: nil -*- */ |
| 2 |
/* ==================================================================== |
| 3 |
* Copyright (c) 2015 Carnegie Mellon University. All rights |
| 4 |
* reserved. |
| 5 |
* |
| 6 |
* Redistribution and use in source and binary forms, with or without |
| 7 |
* modification, are permitted provided that the following conditions |
| 8 |
* are met: |
| 9 |
* |
| 10 |
* 1. Redistributions of source code must retain the above copyright |
| 11 |
* notice, this list of conditions and the following disclaimer. |
| 12 |
* |
| 13 |
* 2. Redistributions in binary form must reproduce the above copyright |
| 14 |
* notice, this list of conditions and the following disclaimer in |
| 15 |
* the documentation and/or other materials provided with the |
| 16 |
* distribution. |
| 17 |
* |
| 18 |
* This work was supported in part by funding from the Defense Advanced |
| 19 |
* Research Projects Agency and the National Science Foundation of the |
| 20 |
* United States of America, and the CMU Sphinx Speech Consortium. |
| 21 |
* |
| 22 |
* THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND |
| 23 |
* ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, |
| 24 |
* THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
| 25 |
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY |
| 26 |
* NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, |
| 27 |
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT |
| 28 |
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
| 29 |
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY |
| 30 |
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 31 |
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 32 |
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 33 |
* |
| 34 |
* ==================================================================== |
| 35 |
* |
| 36 |
*/ |
| 37 |
|
| 38 |
#include <string.h> |
| 39 |
|
| 40 |
#include <sphinxbase/err.h> |
| 41 |
#include <sphinxbase/pio.h> |
| 42 |
#include <sphinxbase/strfuncs.h> |
| 43 |
#include <sphinxbase/ckd_alloc.h> |
| 44 |
#include <sphinxbase/byteorder.h> |
| 45 |
|
| 46 |
#include "ngram_model_internal.h" |
| 47 |
#include "ngrams_raw.h" |
| 48 |
|
| 49 |
int |
| 50 |
ngram_ord_comparator(const void *a_raw, const void *b_raw) |
| 51 |
{ |
| 52 |
ngram_raw_t *a = (ngram_raw_t *) a_raw; |
| 53 |
ngram_raw_t *b = (ngram_raw_t *) b_raw; |
| 54 |
int a_w_ptr = 0; |
| 55 |
int b_w_ptr = 0; |
| 56 |
while (a_w_ptr < a->order && b_w_ptr < b->order) { |
| 57 |
if (a->words[a_w_ptr] == b->words[b_w_ptr]) { |
| 58 |
a_w_ptr++; |
| 59 |
b_w_ptr++; |
| 60 |
continue; |
| 61 |
} |
| 62 |
if (a->words[a_w_ptr] < b->words[b_w_ptr]) |
| 63 |
return -1; |
| 64 |
else |
| 65 |
return 1; |
| 66 |
} |
| 67 |
return a->order - b->order; |
| 68 |
} |
| 69 |
|
| 70 |
static int |
| 71 |
read_ngram_instance(lineiter_t ** li, hash_table_t * wid, |
| 72 |
logmath_t * lmath, int order, int order_max, |
| 73 |
ngram_raw_t * raw_ngram) |
| 74 |
{ |
| 75 |
int n; |
| 76 |
int words_expected; |
| 77 |
int i; |
| 78 |
char *wptr[NGRAM_MAX_ORDER + 1]; |
| 79 |
uint32 *word_out; |
| 80 |
|
| 81 |
if (*li) |
| 82 |
*li = lineiter_next(*li); |
| 83 |
if (*li == NULL) { |
| 84 |
E_ERROR("Unexpected end of ARPA file. Failed to read %d-gram\n", |
| 85 |
order); |
| 86 |
return -1; |
| 87 |
} |
| 88 |
words_expected = order + 1; |
| 89 |
if ((n = |
| 90 |
str2words((*li)->buf, wptr, |
| 91 |
NGRAM_MAX_ORDER + 1)) < words_expected) { |
| 92 |
E_ERROR("Format error; %d-gram ignored: %s\n", order, (*li)->buf); |
| 93 |
return -1; |
| 94 |
} |
| 95 |
|
| 96 |
raw_ngram->order = order; |
| 97 |
|
| 98 |
if (order == order_max) { |
| 99 |
raw_ngram->prob = atof_c(wptr[0]); |
| 100 |
if (raw_ngram->prob > 0) { |
| 101 |
E_WARN("%d-gram '%s' has positive probability\n", order, wptr[1]); |
| 102 |
raw_ngram->prob = 0.0f; |
| 103 |
} |
| 104 |
raw_ngram->prob = |
| 105 |
logmath_log10_to_log_float(lmath, raw_ngram->prob); |
| 106 |
} |
| 107 |
else { |
| 108 |
float weight, backoff; |
| 109 |
|
| 110 |
weight = atof_c(wptr[0]); |
| 111 |
if (weight > 0) { |
| 112 |
E_WARN("%d-gram '%s' has positive probability\n", order, wptr[1]); |
| 113 |
raw_ngram->prob = 0.0f; |
| 114 |
} |
| 115 |
else { |
| 116 |
raw_ngram->prob = |
| 117 |
logmath_log10_to_log_float(lmath, weight); |
| 118 |
} |
| 119 |
|
| 120 |
if (n == order + 1) { |
| 121 |
raw_ngram->backoff = 0.0f; |
| 122 |
} |
| 123 |
else { |
| 124 |
backoff = atof_c(wptr[order + 1]); |
| 125 |
raw_ngram->backoff = |
| 126 |
logmath_log10_to_log_float(lmath, backoff); |
| 127 |
} |
| 128 |
} |
| 129 |
raw_ngram->words = |
| 130 |
(uint32 *) ckd_calloc(order, sizeof(*raw_ngram->words)); |
| 131 |
for (word_out = raw_ngram->words + order - 1, i = 1; |
| 132 |
word_out >= raw_ngram->words; --word_out, i++) { |
| 133 |
hash_table_lookup_int32(wid, wptr[i], (int32 *) word_out); |
| 134 |
} |
| 135 |
return 0; |
| 136 |
} |
| 137 |
|
| 138 |
static int |
| 139 |
ngrams_raw_read_order(ngram_raw_t ** raw_ngrams, lineiter_t ** li, |
| 140 |
hash_table_t * wid, logmath_t * lmath, uint32 count, |
| 141 |
int order, int order_max) |
| 142 |
{ |
| 143 |
char expected_header[20]; |
| 144 |
uint32 i; |
| 145 |
|
| 146 |
sprintf(expected_header, "\\%d-grams:", order); |
| 147 |
while (*li && strcmp((*li)->buf, expected_header) != 0) { |
| 148 |
*li = lineiter_next(*li); |
| 149 |
} |
| 150 |
|
| 151 |
if (*li == NULL) { |
| 152 |
E_ERROR("Failed to find '%s', language model file truncated\n", expected_header); |
| 153 |
return -1; |
| 154 |
} |
| 155 |
|
| 156 |
*raw_ngrams = (ngram_raw_t *) ckd_calloc(count, sizeof(ngram_raw_t)); |
| 157 |
for (i = 0; i < count; i++) { |
| 158 |
if (read_ngram_instance(li, wid, lmath, order, order_max, |
| 159 |
&((*raw_ngrams)[i])) < 0) |
| 160 |
break; |
| 161 |
} |
| 162 |
|
| 163 |
qsort(*raw_ngrams, count, sizeof(ngram_raw_t), &ngram_ord_comparator); |
| 164 |
return 0; |
| 165 |
} |
| 166 |
|
| 167 |
ngram_raw_t ** |
| 168 |
ngrams_raw_read_arpa(lineiter_t ** li, logmath_t * lmath, uint32 * counts, |
| 169 |
int order, hash_table_t * wid) |
| 170 |
{ |
| 171 |
ngram_raw_t **raw_ngrams; |
| 172 |
int order_it; |
| 173 |
|
| 174 |
raw_ngrams = |
| 175 |
(ngram_raw_t **) ckd_calloc(order - 1, sizeof(*raw_ngrams)); |
| 176 |
|
| 177 |
for (order_it = 2; order_it <= order; order_it++) { |
| 178 |
if (ngrams_raw_read_order(&raw_ngrams[order_it - 2], li, wid, lmath, |
| 179 |
counts[order_it - 1], order_it, order) < 0) |
| 180 |
break; |
| 181 |
} |
| 182 |
|
| 183 |
/* Check if we found ARPA end-mark */ |
| 184 |
if (*li == NULL) { |
| 185 |
E_ERROR("ARPA file ends without end-mark\n"); |
| 186 |
ngrams_raw_free(raw_ngrams, counts, order); |
| 187 |
return NULL; |
| 188 |
} else { |
| 189 |
*li = lineiter_next(*li); |
| 190 |
if (strcmp((*li)->buf, "\\end\\") != 0) { |
| 191 |
E_WARN |
| 192 |
("Finished reading ARPA file. Expecting end mark but found '%s'\n", |
| 193 |
(*li)->buf); |
| 194 |
} |
| 195 |
} |
| 196 |
|
| 197 |
return raw_ngrams; |
| 198 |
} |
| 199 |
|
| 200 |
static void |
| 201 |
read_dmp_weight_array(FILE * fp, logmath_t * lmath, uint8 do_swap, |
| 202 |
int32 counts, ngram_raw_t * raw_ngrams, |
| 203 |
int weight_idx) |
| 204 |
{ |
| 205 |
int32 i, k; |
| 206 |
dmp_weight_t *tmp_weight_arr; |
| 207 |
|
| 208 |
fread(&k, sizeof(k), 1, fp); |
| 209 |
if (do_swap) |
| 210 |
SWAP_INT32(&k); |
| 211 |
tmp_weight_arr = |
| 212 |
(dmp_weight_t *) ckd_calloc(k, sizeof(*tmp_weight_arr)); |
| 213 |
fread(tmp_weight_arr, sizeof(*tmp_weight_arr), k, fp); |
| 214 |
for (i = 0; i < k; i++) { |
| 215 |
if (do_swap) |
| 216 |
SWAP_INT32(&tmp_weight_arr[i].l); |
| 217 |
/* Convert values to log. */ |
| 218 |
tmp_weight_arr[i].f = |
| 219 |
logmath_log10_to_log_float(lmath, tmp_weight_arr[i].f); |
| 220 |
} |
| 221 |
/* replace indexes with real probs in raw bigrams */ |
| 222 |
for (i = 0; i < counts; i++) { |
| 223 |
if (weight_idx == 0) { |
| 224 |
raw_ngrams[i].prob = |
| 225 |
tmp_weight_arr[(int) raw_ngrams[i].prob].f; |
| 226 |
} else { |
| 227 |
raw_ngrams[i].backoff = |
| 228 |
tmp_weight_arr[(int) raw_ngrams[i].backoff].f; |
| 229 |
} |
| 230 |
} |
| 231 |
ckd_free(tmp_weight_arr); |
| 232 |
} |
| 233 |
|
| 234 |
#define BIGRAM_SEGMENT_SIZE 9 |
| 235 |
|
| 236 |
ngram_raw_t ** |
| 237 |
ngrams_raw_read_dmp(FILE * fp, logmath_t * lmath, uint32 * counts, |
| 238 |
int order, uint32 * unigram_next, uint8 do_swap) |
| 239 |
{ |
| 240 |
uint32 j, ngram_idx; |
| 241 |
uint16 *bigrams_next; |
| 242 |
ngram_raw_t **raw_ngrams = |
| 243 |
(ngram_raw_t **) ckd_calloc(order - 1, sizeof(*raw_ngrams)); |
| 244 |
|
| 245 |
/* read bigrams */ |
| 246 |
raw_ngrams[0] = |
| 247 |
(ngram_raw_t *) ckd_calloc((size_t) (counts[1] + 1), |
| 248 |
sizeof(*raw_ngrams[0])); |
| 249 |
bigrams_next = |
| 250 |
(uint16 *) ckd_calloc((size_t) (counts[1] + 1), |
| 251 |
sizeof(*bigrams_next)); |
| 252 |
ngram_idx = 1; |
| 253 |
for (j = 0; j <= (int32) counts[1]; j++) { |
| 254 |
uint16 wid, prob_idx, bo_idx; |
| 255 |
ngram_raw_t *raw_ngram = &raw_ngrams[0][j]; |
| 256 |
|
| 257 |
fread(&wid, sizeof(wid), 1, fp); |
| 258 |
if (do_swap) |
| 259 |
SWAP_INT16(&wid); |
| 260 |
raw_ngram->order = 2; |
| 261 |
while (ngram_idx < counts[0] && j == unigram_next[ngram_idx]) { |
| 262 |
ngram_idx++; |
| 263 |
} |
| 264 |
|
| 265 |
if (j != counts[1]) { |
| 266 |
raw_ngram->words = |
| 267 |
(uint32 *) ckd_calloc(2, sizeof(*raw_ngram->words)); |
| 268 |
raw_ngram->words[0] = (uint32) wid; |
| 269 |
raw_ngram->words[1] = (uint32) ngram_idx - 1; |
| 270 |
} |
| 271 |
|
| 272 |
fread(&prob_idx, sizeof(prob_idx), 1, fp); |
| 273 |
fread(&bo_idx, sizeof(bo_idx), 1, fp); |
| 274 |
fread(&bigrams_next[j], sizeof(bigrams_next[j]), 1, fp); |
| 275 |
if (do_swap) { |
| 276 |
SWAP_INT16(&prob_idx); |
| 277 |
SWAP_INT16(&bo_idx); |
| 278 |
SWAP_INT16(&bigrams_next[j]); |
| 279 |
} |
| 280 |
|
| 281 |
if (j != counts[1]) { |
| 282 |
raw_ngram->prob = prob_idx + 0.5f; /* keep index in float. ugly but avoiding using extra memory */ |
| 283 |
raw_ngram->backoff = bo_idx + 0.5f; |
| 284 |
} |
| 285 |
} |
| 286 |
|
| 287 |
if (ngram_idx < counts[0]) { |
| 288 |
E_ERROR("Corrupted model, not enough unigrams %d %d\n", ngram_idx, counts[0]); |
| 289 |
ckd_free(bigrams_next); |
| 290 |
ngrams_raw_free(raw_ngrams, counts, order); |
| 291 |
return NULL; |
| 292 |
} |
| 293 |
|
| 294 |
/* read trigrams */ |
| 295 |
if (order > 2) { |
| 296 |
raw_ngrams[1] = |
| 297 |
(ngram_raw_t *) ckd_calloc((size_t) counts[2], |
| 298 |
sizeof(*raw_ngrams[1])); |
| 299 |
for (j = 0; j < (int32) counts[2]; j++) { |
| 300 |
uint16 wid, prob_idx; |
| 301 |
ngram_raw_t *raw_ngram = &raw_ngrams[1][j]; |
| 302 |
|
| 303 |
fread(&wid, sizeof(wid), 1, fp); |
| 304 |
fread(&prob_idx, sizeof(prob_idx), 1, fp); |
| 305 |
if (do_swap) { |
| 306 |
SWAP_INT16(&wid); |
| 307 |
SWAP_INT16(&prob_idx); |
| 308 |
} |
| 309 |
|
| 310 |
raw_ngram->order = 3; |
| 311 |
raw_ngram->words = |
| 312 |
(uint32 *) ckd_calloc(3, sizeof(*raw_ngram->words)); |
| 313 |
raw_ngram->words[0] = (uint32) wid; |
| 314 |
raw_ngram->prob = prob_idx + 0.5f; /* keep index in float. ugly but avoiding using extra memory */ |
| 315 |
} |
| 316 |
} |
| 317 |
|
| 318 |
/* read prob2 */ |
| 319 |
read_dmp_weight_array(fp, lmath, do_swap, (int32) counts[1], |
| 320 |
raw_ngrams[0], 0); |
| 321 |
/* read bo2 */ |
| 322 |
if (order > 2) { |
| 323 |
int32 k; |
| 324 |
int32 *tseg_base; |
| 325 |
read_dmp_weight_array(fp, lmath, do_swap, (int32) counts[1], |
| 326 |
raw_ngrams[0], 1); |
| 327 |
/* read prob3 */ |
| 328 |
read_dmp_weight_array(fp, lmath, do_swap, (int32) counts[2], |
| 329 |
raw_ngrams[1], 0); |
| 330 |
/* Read tseg_base size and tseg_base to fill trigram's first words */ |
| 331 |
fread(&k, sizeof(k), 1, fp); |
| 332 |
if (do_swap) |
| 333 |
SWAP_INT32(&k); |
| 334 |
tseg_base = (int32 *) ckd_calloc(k, sizeof(int32)); |
| 335 |
fread(tseg_base, sizeof(int32), k, fp); |
| 336 |
if (do_swap) { |
| 337 |
for (j = 0; j < (uint32) k; j++) { |
| 338 |
SWAP_INT32(&tseg_base[j]); |
| 339 |
} |
| 340 |
} |
| 341 |
ngram_idx = 0; |
| 342 |
for (j = 1; j <= counts[1]; j++) { |
| 343 |
uint32 next_ngram_idx = |
| 344 |
(uint32) (tseg_base[j >> BIGRAM_SEGMENT_SIZE] + |
| 345 |
bigrams_next[j]); |
| 346 |
while (ngram_idx < next_ngram_idx) { |
| 347 |
raw_ngrams[1][ngram_idx].words[1] = |
| 348 |
raw_ngrams[0][j - 1].words[0]; |
| 349 |
raw_ngrams[1][ngram_idx].words[2] = |
| 350 |
raw_ngrams[0][j - 1].words[1]; |
| 351 |
ngram_idx++; |
| 352 |
} |
| 353 |
} |
| 354 |
ckd_free(tseg_base); |
| 355 |
|
| 356 |
if (ngram_idx < counts[2]) { |
| 357 |
E_ERROR("Corrupted model, some trigrams have no corresponding bigram\n"); |
| 358 |
ckd_free(bigrams_next); |
| 359 |
ngrams_raw_free(raw_ngrams, counts, order); |
| 360 |
return NULL; |
| 361 |
} |
| 362 |
} |
| 363 |
ckd_free(bigrams_next); |
| 364 |
|
| 365 |
/* sort raw ngrams for reverse trie */ |
| 366 |
qsort(raw_ngrams[0], (size_t) counts[1], sizeof(*raw_ngrams[0]), |
| 367 |
&ngram_ord_comparator); |
| 368 |
if (order > 2) { |
| 369 |
qsort(raw_ngrams[1], (size_t) counts[2], sizeof(*raw_ngrams[1]), |
| 370 |
&ngram_ord_comparator); |
| 371 |
} |
| 372 |
return raw_ngrams; |
| 373 |
} |
| 374 |
|
| 375 |
void |
| 376 |
ngrams_raw_free(ngram_raw_t ** raw_ngrams, uint32 * counts, int order) |
| 377 |
{ |
| 378 |
uint32 num; |
| 379 |
int order_it; |
| 380 |
|
| 381 |
for (order_it = 0; order_it < order - 1; order_it++) { |
| 382 |
for (num = 0; num < counts[order_it + 1]; num++) { |
| 383 |
ckd_free(raw_ngrams[order_it][num].words); |
| 384 |
} |
| 385 |
ckd_free(raw_ngrams[order_it]); |
| 386 |
} |
| 387 |
ckd_free(raw_ngrams); |
| 388 |
} |