static void
build_widmap(ngram_model_t * base, logmath_t * lmath, int32 n)
{
ngram_model_set_t *set = (ngram_model_set_t *) base;
ngram_model_t **models = set->lms;
hash_table_t *vocab;
glist_t hlist;
gnode_t *gn;
int32 i;
/* Construct a merged vocabulary and a set of word-ID mappings. */
vocab = hash_table_new(models[0]->n_words, FALSE);
/* Create the set of merged words. */
for (i = 0; i < set->n_models; ++i) {
int32 j;
for (j = 0; j < models[i]->n_words; ++j) {
/* Ignore collisions. */
(void) hash_table_enter_int32(vocab, models[i]->word_str[j],
j);
}
}
/* Create the array of words, then sort it. */
if (hash_table_lookup(vocab, "<UNK>", NULL) != 0)
(void) hash_table_enter_int32(vocab, "<UNK>", 0);
/* Now we know the number of unigrams, initialize the base model. */
ngram_model_init(base, &ngram_model_set_funcs, lmath, n,
hash_table_inuse(vocab));
base->writable = FALSE; /* We will reuse the pointers from the submodels. */
i = 0;
hlist = hash_table_tolist(vocab, NULL);
for (gn = hlist; gn; gn = gnode_next(gn)) {
hash_entry_t *ent = gnode_ptr(gn);
base->word_str[i++] = (char *) ent->key;
}
glist_free(hlist);
qsort(base->word_str, base->n_words, sizeof(*base->word_str),
my_compare);
/* Now create the word ID mappings. */
if (set->widmap)
ckd_free_2d((void **) set->widmap);
set->widmap = (int32 **) ckd_calloc_2d(base->n_words, set->n_models,
sizeof(**set->widmap));
for (i = 0; i < base->n_words; ++i) {
int32 j;
/* Also create the master wid mapping. */
(void) hash_table_enter_int32(base->wid, base->word_str[i], i);
/* printf("%s: %d => ", base->word_str[i], i); */
for (j = 0; j < set->n_models; ++j) {
set->widmap[i][j] = ngram_wid(models[j], base->word_str[i]);
/* printf("%d ", set->widmap[i][j]); */
}
/* printf("\n"); */
}
hash_table_free(vocab);
}
ngram_model_t *
ngram_model_trie_read_bin(cmd_ln_t * config,
const char *path, logmath_t * lmath)
{
int32 is_pipe;
FILE *fp;
size_t hdr_size;
char *hdr;
int cmp_res;
uint8 i, order;
uint32 counts[NGRAM_MAX_ORDER];
ngram_model_trie_t *model;
ngram_model_t *base;
E_INFO("Trying to read LM in trie binary format\n");
if ((fp = fopen_comp(path, "rb", &is_pipe)) == NULL) {
E_ERROR("File %s not found\n", path);
return NULL;
}
hdr_size = strlen(trie_hdr);
hdr = (char *) ckd_calloc(hdr_size + 1, sizeof(*hdr));
fread(hdr, sizeof(*hdr), hdr_size, fp);
cmp_res = strcmp(hdr, trie_hdr);
ckd_free(hdr);
if (cmp_res) {
E_INFO("Header doesn't match\n");
fclose_comp(fp, is_pipe);
return NULL;
}
model = (ngram_model_trie_t *) ckd_calloc(1, sizeof(*model));
base = &model->base;
fread(&order, sizeof(order), 1, fp);
for (i = 0; i < order; i++) {
fread(&counts[i], sizeof(counts[i]), 1, fp);
}
ngram_model_init(base, &ngram_model_trie_funcs, lmath, order,
(int32) counts[0]);
for (i = 0; i < order; i++) {
base->n_counts[i] = counts[i];
}
model->trie = lm_trie_read_bin(counts, order, fp);
read_word_str(base, fp);
fclose_comp(fp, is_pipe);
return base;
}
ngram_model_t *
ngram_model_trie_read_dmp(cmd_ln_t * config,
const char *file_name, logmath_t * lmath)
{
uint8 do_swap;
int32 is_pipe;
int32 k;
uint32 j;
int32 vn, ts;
int32 count;
uint32 counts[3];
uint32 fixed_counts[3];
uint32 *unigram_next;
int i, order;
char str[1024];
FILE *fp;
ngram_model_trie_t *model;
ngram_model_t *base;
ngram_raw_t **raw_ngrams;
E_INFO("Trying to read LM in DMP format\n");
if ((fp = fopen_comp(file_name, "rb", &is_pipe)) == NULL) {
E_ERROR("Dump file %s not found\n", file_name);
return NULL;
}
do_swap = FALSE;
fread(&k, sizeof(k), 1, fp);
if (k != strlen(dmp_hdr) + 1) {
SWAP_INT32(&k);
if (k != strlen(dmp_hdr) + 1) {
E_ERROR
("Wrong magic header size number %x: %s is not a dump file\n",
k, file_name);
return NULL;
}
do_swap = 1;
}
if (fread(str, 1, k, fp) != (size_t) k) {
E_ERROR("Cannot read header\n");
return NULL;
}
if (strncmp(str, dmp_hdr, k) != 0) {
E_ERROR("Wrong header %s: %s is not a dump file\n", dmp_hdr);
return NULL;
}
if (fread(&k, sizeof(k), 1, fp) != 1)
return NULL;
if (do_swap)
SWAP_INT32(&k);
if (fread(str, 1, k, fp) != (size_t) k) {
E_ERROR("Cannot read LM filename in header\n");
return NULL;
}
/* read version#, if present (must be <= 0) */
if (fread(&vn, sizeof(vn), 1, fp) != 1)
return NULL;
if (do_swap)
SWAP_INT32(&vn);
if (vn <= 0) {
/* read and don't compare timestamps (we don't care) */
if (fread(&ts, sizeof(ts), 1, fp) != 1)
return NULL;
if (do_swap)
SWAP_INT32(&ts);
/* read and skip format description */
for (;;) {
if (fread(&k, sizeof(k), 1, fp) != 1)
return NULL;
if (do_swap)
SWAP_INT32(&k);
if (k == 0)
break;
if (fread(str, 1, k, fp) != (size_t) k) {
E_ERROR("Failed to read word\n");
return NULL;
}
}
/* read model->ucount */
if (fread(&count, sizeof(count), 1, fp) != 1)
return NULL;
if (do_swap)
SWAP_INT32(&count);
counts[0] = count;
}
else {
counts[0] = vn;
}
/* read model->bcount, tcount */
if (fread(&count, sizeof(count), 1, fp) != 1)
return NULL;
if (do_swap)
SWAP_INT32(&count);
counts[1] = count;
if (fread(&count, sizeof(count), 1, fp) != 1)
return NULL;
if (do_swap)
SWAP_INT32(&count);
counts[2] = count;
E_INFO("ngrams 1=%d, 2=%d, 3=%d\n", counts[0], counts[1], counts[2]);
model = (ngram_model_trie_t *) ckd_calloc(1, sizeof(*model));
base = &model->base;
if (counts[2] > 0)
order = 3;
else if (counts[1] > 0)
order = 2;
else
order = 1;
ngram_model_init(base, &ngram_model_trie_funcs, lmath, order,
(int32) counts[0]);
model->trie = lm_trie_create(counts[0], order);
unigram_next =
(uint32 *) ckd_calloc((int32) counts[0] + 1, sizeof(unigram_next));
for (j = 0; j <= (int32) counts[0]; j++) {
int32 bigrams;
dmp_weight_t weight;
/* Skip over the mapping ID, we don't care about it. */
fread(&bigrams, sizeof(int32), 1, fp);
/* Read the weights from actual unigram structure. */
fread(&weight, sizeof(weight), 1, fp);
if (do_swap)
SWAP_INT32(&weight.l);
weight.f = logmath_log10_to_log_float(lmath, weight.f);
model->trie->unigrams[j].prob = weight.f;
fread(&weight, sizeof(weight), 1, fp);
if (do_swap)
SWAP_INT32(&weight.l);
weight.f = logmath_log10_to_log_float(lmath, weight.f);
model->trie->unigrams[j].bo = weight.f;
//store pointer to dmp next to recognize wid
fread(&bigrams, sizeof(int32), 1, fp);
if (do_swap)
SWAP_INT32(&bigrams);
model->trie->unigrams[j].next = bigrams;
unigram_next[j] = bigrams;
}
if (order > 1) {
raw_ngrams =
ngrams_raw_read_dmp(fp, lmath, counts, order, unigram_next,
do_swap);
ngrams_raw_fix_counts(raw_ngrams, counts, fixed_counts, order);
for (i = 0; i < order; i++) {
base->n_counts[i] = fixed_counts[i];
}
//build reversed trie
lm_trie_alloc_ngram(model->trie, order > 2 ? fixed_counts : counts,
order);
lm_trie_build(model->trie, raw_ngrams, counts, order);
counts[1]++;
//free raw ngrams
ngrams_raw_free(raw_ngrams, counts, order);
}
ckd_free(unigram_next);
/* read ascii word strings */
read_word_str(base, fp);
fclose_comp(fp, is_pipe);
return base;
}
ngram_model_t *
ngram_model_trie_read_arpa(cmd_ln_t * config,
const char *path, logmath_t * lmath)
{
FILE *fp;
lineiter_t *li;
ngram_model_trie_t *model;
ngram_model_t *base;
ngram_raw_t **raw_ngrams;
int32 is_pipe;
uint32 counts[NGRAM_MAX_ORDER];
uint32 fixed_counts[NGRAM_MAX_ORDER];
int order;
int i;
E_INFO("Trying to read LM in arpa format\n");
if ((fp = fopen_comp(path, "r", &is_pipe)) == NULL) {
E_ERROR("File %s not found\n", path);
return NULL;
}
model = (ngram_model_trie_t *) ckd_calloc(1, sizeof(*model));
li = lineiter_start_clean(fp);
/* Read n-gram counts from file */
if (read_counts_arpa(&li, counts, &order) == -1) {
ckd_free(model);
lineiter_free(li);
fclose_comp(fp, is_pipe);
return NULL;
}
E_INFO("LM of order %d\n", order);
for (i = 0; i < order; i++) {
E_INFO("#%d-grams: %d\n", i + 1, counts[i]);
}
base = &model->base;
ngram_model_init(base, &ngram_model_trie_funcs, lmath, order,
(int32) counts[0]);
base->writable = TRUE;
model->trie = lm_trie_create(counts[0], order);
if (read_1grams_arpa(&li, counts[0], base, model->trie->unigrams) < 0) {
ckd_free(model);
lineiter_free(li);
fclose_comp(fp, is_pipe);
return NULL;
}
if (order > 1) {
raw_ngrams =
ngrams_raw_read_arpa(&li, base->lmath, counts, order,
base->wid);
ngrams_raw_fix_counts(raw_ngrams, counts, fixed_counts, order);
for (i = 0; i < order; i++) {
base->n_counts[i] = fixed_counts[i];
}
lm_trie_alloc_ngram(model->trie, fixed_counts, order);
lm_trie_build(model->trie, raw_ngrams, counts, order);
ngrams_raw_free(raw_ngrams, counts, order);
}
lineiter_free(li);
fclose_comp(fp, is_pipe);
return base;
}
