static void
read_dmp_weight_array(FILE * fp, logmath_t * lmath, uint8 do_swap,
int32 counts, ngram_raw_t * raw_ngrams,
int weight_idx)
{
int32 i, k;
dmp_weight_t *tmp_weight_arr;
fread(&k, sizeof(k), 1, fp);
if (do_swap)
SWAP_INT32(&k);
tmp_weight_arr =
(dmp_weight_t *) ckd_calloc(k, sizeof(*tmp_weight_arr));
fread(tmp_weight_arr, sizeof(*tmp_weight_arr), k, fp);
for (i = 0; i < k; i++) {
if (do_swap)
SWAP_INT32(&tmp_weight_arr[i].l);
/* Convert values to log. */
tmp_weight_arr[i].f =
logmath_log10_to_log_float(lmath, tmp_weight_arr[i].f);
}
//replace indexes with real probs in raw bigrams
for (i = 0; i < counts; i++) {
raw_ngrams[i].weights[weight_idx] =
tmp_weight_arr[(int) raw_ngrams[i].weights[weight_idx]].f;
}
ckd_free(tmp_weight_arr);
}
static void
read_ngram_instance(lineiter_t ** li, hash_table_t * wid,
logmath_t * lmath, int order, int order_max,
ngram_raw_t * raw_ngram)
{
int n;
int words_expected;
int i;
char *wptr[NGRAM_MAX_ORDER + 1];
uint32 *word_out;
*li = lineiter_next(*li);
if (*li == NULL) {
E_ERROR("Unexpected end of ARPA file. Failed to read %d-gram\n",
order);
return;
}
string_trim((*li)->buf, STRING_BOTH);
words_expected = order + 1;
if ((n =
str2words((*li)->buf, wptr,
NGRAM_MAX_ORDER + 1)) < words_expected) {
if ((*li)->buf[0] != '\0') {
E_WARN("Format error; %d-gram ignored: %s\n", order,
(*li)->buf);
}
}
else {
if (order == order_max) {
raw_ngram->weights =
(float *) ckd_calloc(1, sizeof(*raw_ngram->weights));
raw_ngram->weights[0] = atof_c(wptr[0]);
if (raw_ngram->weights[0] > 0) {
E_WARN("%d-gram [%s] has positive probability. Zeroize\n",
order, wptr[1]);
raw_ngram->weights[0] = 0.0f;
}
raw_ngram->weights[0] =
logmath_log10_to_log_float(lmath, raw_ngram->weights[0]);
}
else {
float weight, backoff;
raw_ngram->weights =
(float *) ckd_calloc(2, sizeof(*raw_ngram->weights));
weight = atof_c(wptr[0]);
if (weight > 0) {
E_WARN("%d-gram [%s] has positive probability. Zeroize\n",
order, wptr[1]);
raw_ngram->weights[0] = 0.0f;
}
else {
raw_ngram->weights[0] =
logmath_log10_to_log_float(lmath, weight);
}
if (n == order + 1) {
raw_ngram->weights[1] = 0.0f;
}
else {
backoff = atof_c(wptr[order + 1]);
raw_ngram->weights[1] =
logmath_log10_to_log_float(lmath, backoff);
}
}
raw_ngram->words =
(uint32 *) ckd_calloc(order, sizeof(*raw_ngram->words));
for (word_out = raw_ngram->words + order - 1, i = 1;
word_out >= raw_ngram->words; --word_out, i++) {
hash_table_lookup_int32(wid, wptr[i], (int32 *) word_out);
}
}
}
static int
read_1grams_arpa(lineiter_t ** li, uint32 count, ngram_model_t * base,
unigram_t * unigrams)
{
uint32 i;
int n;
int n_parts;
char *wptr[3];
while (*li && strcmp((*li)->buf, "\\1-grams:") != 0) {
*li = lineiter_next(*li);
}
if (*li == NULL) {
E_ERROR_SYSTEM("Failed to read \\1-grams: mark");
return -1;
}
n_parts = 2;
for (i = 0; i < count; i++) {
*li = lineiter_next(*li);
if (*li == NULL) {
E_ERROR
("Unexpected end of ARPA file. Failed to read %dth unigram\n",
i + 1);
return -1;
}
if ((n = str2words((*li)->buf, wptr, 3)) < n_parts) {
E_ERROR("Format error at line %s, Failed to read unigrams\n", (*li)->buf);
return -1;
}
unigram_t *unigram = &unigrams[i];
unigram->prob =
logmath_log10_to_log_float(base->lmath, atof_c(wptr[0]));
if (unigram->prob > 0) {
E_WARN("Unigram '%s' has positive probability\n", wptr[1]);
unigram->prob = 0;
}
if (n == n_parts + 1) {
unigram->bo =
logmath_log10_to_log_float(base->lmath,
atof_c(wptr[2]));
}
else {
unigram->bo = 0.0f;
}
/* TODO: classify float with fpclassify and warn if bad value occurred */
base->word_str[i] = ckd_salloc(wptr[1]);
}
/* fill hash-table that maps unigram names to their word ids */
for (i = 0; i < count; i++) {
if ((hash_table_enter
(base->wid, base->word_str[i],
(void *) (long) i)) != (void *) (long) i) {
E_WARN("Duplicate word in dictionary: %s\n",
base->word_str[i]);
}
}
return 0;
}
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;
}
static int
read_ngram_instance(lineiter_t ** li, hash_table_t * wid,
logmath_t * lmath, int order, int order_max,
ngram_raw_t * raw_ngram)
{
int n;
int words_expected;
int i;
char *wptr[NGRAM_MAX_ORDER + 1];
uint32 *word_out;
if (*li)
*li = lineiter_next(*li);
if (*li == NULL) {
E_ERROR("Unexpected end of ARPA file. Failed to read %d-gram\n",
order);
return -1;
}
words_expected = order + 1;
if ((n =
str2words((*li)->buf, wptr,
NGRAM_MAX_ORDER + 1)) < words_expected) {
E_ERROR("Format error; %d-gram ignored: %s\n", order, (*li)->buf);
return -1;
}
raw_ngram->order = order;
if (order == order_max) {
raw_ngram->prob = atof_c(wptr[0]);
if (raw_ngram->prob > 0) {
E_WARN("%d-gram '%s' has positive probability\n", order, wptr[1]);
raw_ngram->prob = 0.0f;
}
raw_ngram->prob =
logmath_log10_to_log_float(lmath, raw_ngram->prob);
}
else {
float weight, backoff;
weight = atof_c(wptr[0]);
if (weight > 0) {
E_WARN("%d-gram '%s' has positive probability\n", order, wptr[1]);
raw_ngram->prob = 0.0f;
}
else {
raw_ngram->prob =
logmath_log10_to_log_float(lmath, weight);
}
if (n == order + 1) {
raw_ngram->backoff = 0.0f;
}
else {
backoff = atof_c(wptr[order + 1]);
raw_ngram->backoff =
logmath_log10_to_log_float(lmath, backoff);
}
}
raw_ngram->words =
(uint32 *) ckd_calloc(order, sizeof(*raw_ngram->words));
for (word_out = raw_ngram->words + order - 1, i = 1;
word_out >= raw_ngram->words; --word_out, i++) {
hash_table_lookup_int32(wid, wptr[i], (int32 *) word_out);
}
return 0;
}