int32 subvq_mgau_eval(mgau_model_t * g, subvq_t * vq, int32 m, int32 n, int32 * active) { mgau_t *mgau; int32 *map; int32 i, v, sv_id; int32 c; int32 *vqdist; int32 score; int32 last_active; float64 f; f = 1.0 / log(logmath_get_base(g->logmath)); vqdist = vq->vqdist[0]; score = S3_LOGPROB_ZERO; mgau = &(g->mgau[m]); map = vq->map[m][0]; if (!active) { for (i = 0; i < n; i++) { v = 0; for (sv_id = 0; sv_id < vq->n_sv; sv_id++) { v += vqdist[*(map++)]; } score = logmath_add(g->logmath, score, v + mgau->mixw[i]); } } else { last_active = 0; for (i = 0; active[i] >= 0; i++) { c = active[i]; } for (i = 0; active[i] >= 0; i++) { c = active[i]; map += (c - last_active) * vq->n_sv; v = 0; for (sv_id = 0; sv_id < vq->n_sv; sv_id++) { v += vqdist[*(map++)]; } last_active = c + 1; score = logmath_add(g->logmath, score, v + mgau->mixw[i]); } } if (score == S3_LOGPROB_ZERO) { E_INFO("Warning!! Score is S3_LOGPROB_ZERO\n"); } return score; }
/* * Normalize density values. */ int32 gauden_dist_norm(gauden_t * g, int32 n_top, gauden_dist_t *** dist, uint8 * active) { int32 gid, f, t; int32 sum, scale; if (g->n_mgau > 1) { /* Normalize by subtracting max(density values) from each density */ return (gauden_dist_norm_global(g, n_top, dist, active)); } /* Normalize by subtracting log(sum of density values) from each density */ gid = 0; scale = 0; for (f = 0; f < g->n_feat; f++) { sum = dist[gid][f][0].dist; for (t = 1; t < n_top; t++) sum = logmath_add(g->logmath, sum, dist[gid][f][t].dist); for (t = 0; t < n_top; t++) dist[gid][f][t].dist -= sum; scale += sum; } return scale; /* Scale factor applied to EVERY senone score */ }
static int32 ngram_model_set_add_ug(ngram_model_t * base, int32 wid, int32 lweight) { ngram_model_set_t *set = (ngram_model_set_t *) base; int32 *newwid; int32 i, prob; /* At this point the word has already been added to the master model and we have a new word ID for it. Add it to active submodels and track the word IDs. */ newwid = ckd_calloc(set->n_models, sizeof(*newwid)); prob = base->log_zero; for (i = 0; i < set->n_models; ++i) { int32 wprob, n_hist; /* Only add to active models. */ if (set->cur == -1 || set->cur == i) { /* Did this word already exist? */ newwid[i] = ngram_wid(set->lms[i], base->word_str[wid]); if (newwid[i] == NGRAM_INVALID_WID) { /* Add it to the submodel. */ newwid[i] = ngram_model_add_word(set->lms[i], base->word_str[wid], (float32) logmath_exp(base->lmath, lweight)); if (newwid[i] == NGRAM_INVALID_WID) { ckd_free(newwid); return base->log_zero; } } /* Now get the unigram probability for the new word and either * interpolate it or use it (if this is the current model). */ wprob = ngram_ng_prob(set->lms[i], newwid[i], NULL, 0, &n_hist); if (set->cur == i) prob = wprob; else if (set->cur == -1) prob = logmath_add(base->lmath, prob, set->lweights[i] + wprob); } else { newwid[i] = NGRAM_INVALID_WID; } } /* Okay we have the word IDs for this in all the submodels. Now do some complicated memory mangling to add this to the widmap. */ set->widmap = ckd_realloc(set->widmap, base->n_words * sizeof(*set->widmap)); set->widmap[0] = ckd_realloc(set->widmap[0], base->n_words * set->n_models * sizeof(**set->widmap)); for (i = 0; i < base->n_words; ++i) set->widmap[i] = set->widmap[0] + i * set->n_models; memcpy(set->widmap[wid], newwid, set->n_models * sizeof(*newwid)); ckd_free(newwid); return prob; }
int32 interp_cd_ci(interp_t * ip, int32 * senscr, int32 cd, int32 ci) { assert((ci >= 0) && (ci < ip->n_sen)); assert((cd >= 0) && (cd < ip->n_sen)); senscr[cd] = logmath_add(ip->logmath, senscr[cd] + ip->wt[cd].cd, senscr[ci] + ip->wt[cd].ci); return 0; }
static int32 ngram_model_set_raw_score(ngram_model_t * base, int32 wid, int32 * history, int32 n_hist, int32 * n_used) { ngram_model_set_t *set = (ngram_model_set_t *) base; int32 mapwid; int32 score; int32 i; /* Truncate the history. */ if (n_hist > base->n - 1) n_hist = base->n - 1; /* Interpolate if there is no current. */ if (set->cur == -1) { score = base->log_zero; for (i = 0; i < set->n_models; ++i) { int32 j; /* Map word and history IDs for each model. */ mapwid = set->widmap[wid][i]; for (j = 0; j < n_hist; ++j) { if (history[j] == NGRAM_INVALID_WID) set->maphist[j] = NGRAM_INVALID_WID; else set->maphist[j] = set->widmap[history[j]][i]; } score = logmath_add(base->lmath, score, set->lweights[i] + ngram_ng_prob(set->lms[i], mapwid, set->maphist, n_hist, n_used)); } } else { int32 j; /* Map word and history IDs (FIXME: do this in a function?) */ mapwid = set->widmap[wid][set->cur]; for (j = 0; j < n_hist; ++j) { if (history[j] == NGRAM_INVALID_WID) set->maphist[j] = NGRAM_INVALID_WID; else set->maphist[j] = set->widmap[history[j]][set->cur]; } score = ngram_ng_prob(set->lms[set->cur], mapwid, set->maphist, n_hist, n_used); } return score; }
int32 interp_all(interp_t * ip, int32 * senscr, s3senid_t * cimap, int32 n_ci_sen) { int32 ci, cd; assert(n_ci_sen <= ip->n_sen); for (cd = n_ci_sen; cd < ip->n_sen; cd++) { ci = cimap[cd]; senscr[cd] = logmath_add(ip->logmath, senscr[cd] + ip->wt[cd].cd, senscr[ci] + ip->wt[cd].ci); } return 0; }
int test_decode(ps_decoder_t *ps) { FILE *rawfh; int16 buf[2048]; size_t nread; int16 const *bptr; int nfr; ps_lattice_t *dag; acmod_t *acmod; ngram_search_t *ngs; int i, j; ps_latlink_t *link; ps_latnode_t *node; latlink_list_t *x; int32 norm, post; ngs = (ngram_search_t *)ps->search; acmod = ps->acmod; /* Decode stuff and build a DAG. */ TEST_ASSERT(rawfh = fopen(DATADIR "/goforward.raw", "rb")); TEST_EQUAL(0, acmod_start_utt(acmod)); ngram_fwdtree_start(ngs); while (!feof(rawfh)) { nread = fread(buf, sizeof(*buf), 2048, rawfh); bptr = buf; while ((nfr = acmod_process_raw(acmod, &bptr, &nread, FALSE)) > 0) { while (acmod->n_feat_frame > 0) { ngram_fwdtree_search(ngs, acmod->output_frame); acmod_advance(acmod); } } } ngram_fwdtree_finish(ngs); printf("FWDTREE: %s\n", ngram_search_bp_hyp(ngs, ngram_search_find_exit(ngs, -1, NULL, NULL))); TEST_ASSERT(acmod_end_utt(acmod) >= 0); fclose(rawfh); dag = ngram_search_lattice(ps->search); if (dag == NULL) { E_ERROR("Failed to build DAG!\n"); return -1; } /* Write lattice to disk. */ TEST_EQUAL(0, ps_lattice_write(dag, "test_posterior.lat")); /* Do a bunch of checks on the DAG generation and traversal code: */ /* Verify that forward and backward iteration give the same number of edges. */ i = j = 0; for (link = ps_lattice_traverse_edges(dag, NULL, NULL); link; link = ps_lattice_traverse_next(dag, NULL)) { ++i; } for (link = ps_lattice_reverse_edges(dag, NULL, NULL); link; link = ps_lattice_reverse_next(dag, NULL)) { ++j; } printf("%d forward edges, %d reverse edges\n", i, j); TEST_EQUAL(i,j); /* Verify that the same links are reachable via entries and exits. */ for (node = dag->nodes; node; node = node->next) { for (x = node->exits; x; x = x->next) x->link->alpha = -42; } for (node = dag->nodes; node; node = node->next) { for (x = node->entries; x; x = x->next) TEST_EQUAL(x->link->alpha, -42); } /* Verify that forward iteration is properly ordered. */ for (link = ps_lattice_traverse_edges(dag, NULL, NULL); link; link = ps_lattice_traverse_next(dag, NULL)) { link->alpha = 0; for (x = link->from->entries; x; x = x->next) { TEST_EQUAL(x->link->alpha, 0); } } /* Verify that backward iteration is properly ordered. */ for (node = dag->nodes; node; node = node->next) { for (x = node->exits; x; x = x->next) x->link->alpha = -42; } for (link = ps_lattice_reverse_edges(dag, NULL, NULL); link; link = ps_lattice_reverse_next(dag, NULL)) { link->alpha = 0; for (x = link->to->exits; x; x = x->next) { TEST_EQUAL(x->link->alpha, 0); } } /* Find and print best path. */ link = ps_lattice_bestpath(dag, ngs->lmset, 1.0, 1.0/20.0); printf("BESTPATH: %s\n", ps_lattice_hyp(dag, link)); /* Calculate betas. */ post = ps_lattice_posterior(dag, ngs->lmset, 1.0/20.0); printf("Best path score: %d\n", link->path_scr + dag->final_node_ascr); printf("P(S|O) = %d\n", post); /* Verify that sum of final alphas and initial alphas+betas is * sufficiently similar. */ norm = logmath_get_zero(acmod->lmath); for (x = dag->start->exits; x; x = x->next) norm = logmath_add(acmod->lmath, norm, x->link->beta + x->link->alpha); E_INFO("Sum of final alphas+betas = %d\n", dag->norm); E_INFO("Sum of initial alphas+betas = %d\n", norm); TEST_EQUAL_LOG(dag->norm, norm); /* Print posterior probabilities for each link in best path. */ while (link) { printf("P(%s,%d) = %d = %f\n", dict_wordstr(ps->search->dict, link->from->wid), link->ef, link->alpha + link->beta - dag->norm, logmath_exp(acmod->lmath, link->alpha + link->beta - dag->norm)); link = link->best_prev; } return 0; }
int main(int argc, char *argv[]) { logmath_t *lmath; int32 rv; lmath = logmath_init(1.0001, 0, 1); TEST_ASSERT(lmath); printf("log(1e-150) = %d\n", logmath_log(lmath, 1e-150)); TEST_EQUAL_LOG(logmath_log(lmath, 1e-150), -3454050); printf("exp(log(1e-150)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-150))); TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-150)), 1e-150); printf("log(1e-48) = %d\n", logmath_log(lmath, 1e-48)); printf("exp(log(1e-48)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-48))); TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-48)), 1e-48); printf("log(42) = %d\n", logmath_log(lmath, 42)); TEST_EQUAL_LOG(logmath_log(lmath, 42), 37378); printf("exp(log(42)) = %f\n",logmath_exp(lmath, logmath_log(lmath, 42))); TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 42)), 42); printf("log(1e-3 + 5e-3) = %d l+ %d = %d\n", logmath_log(lmath, 1e-3), logmath_log(lmath, 5e-3), logmath_add(lmath, logmath_log(lmath, 1e-3), logmath_log(lmath, 5e-3))); printf("log(1e-3 + 5e-3) = %e + %e = %e\n", logmath_exp(lmath, logmath_log(lmath, 1e-3)), logmath_exp(lmath, logmath_log(lmath, 5e-3)), logmath_exp(lmath, logmath_add(lmath, logmath_log(lmath, 1e-3), logmath_log(lmath, 5e-3)))); TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48), logmath_log(lmath, 5e-48)), logmath_log(lmath, 6e-48)); TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48), logmath_log(lmath, 42)), logmath_log(lmath, 42)); rv = logmath_write(lmath, "tmp.logadd"); TEST_EQUAL(rv, 0); logmath_free(lmath); lmath = logmath_read("tmp.logadd"); TEST_ASSERT(lmath); printf("log(1e-150) = %d\n", logmath_log(lmath, 1e-150)); TEST_EQUAL_LOG(logmath_log(lmath, 1e-150), -3454050); printf("exp(log(1e-150)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-150))); TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-150)), 1e-150); printf("log(1e-48) = %d\n", logmath_log(lmath, 1e-48)); printf("exp(log(1e-48)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-48))); TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-48)), 1e-48); printf("log(42) = %d\n", logmath_log(lmath, 42)); TEST_EQUAL_LOG(logmath_log(lmath, 42), 37378); printf("exp(log(42)) = %f\n",logmath_exp(lmath, logmath_log(lmath, 42))); TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 42)), 41.99); printf("log(1e-3 + 5e-3) = %d l+ %d = %d\n", logmath_log(lmath, 1e-3), logmath_log(lmath, 5e-3), logmath_add(lmath, logmath_log(lmath, 1e-3), logmath_log(lmath, 5e-3))); printf("log(1e-3 + 5e-3) = %e + %e = %e\n", logmath_exp(lmath, logmath_log(lmath, 1e-3)), logmath_exp(lmath, logmath_log(lmath, 5e-3)), logmath_exp(lmath, logmath_add(lmath, logmath_log(lmath, 1e-3), logmath_log(lmath, 5e-3)))); TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48), logmath_log(lmath, 5e-48)), logmath_log(lmath, 6e-48)); TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48), logmath_log(lmath, 42)), logmath_log(lmath, 42)); return 0; }