int
srch_FSG_init(kb_t * kb, /**< The KB */
void *srch /**< The pointer to a search structure */
)
{
srch_t *s;
fsg_search_t *fsgsrch;
word_fsg_t *wordfsg;
s = (srch_t *) srch;
/* This is very strange */
fsgsrch = fsg_search_init(NULL, s);
s->grh->graph_struct = fsgsrch;
s->grh->graph_type = GRAPH_STRUCT_GENGRAPH;
if ((wordfsg = srch_FSG_read_fsgfile(s, cmd_ln_str("-fsg"))) == NULL) {
E_INFO("Could not read wordfsg with file name %s\n",
cmd_ln_str("-fsg"));
return SRCH_FAILURE;
}
if (!fsg_search_set_current_fsg(fsgsrch, wordfsg->name)) {
E_INFO("Could not set the current fsg with name %s\n",
wordfsg->name);
return SRCH_FAILURE;
}
return SRCH_SUCCESS;
}
static void
gst_pocketsphinx_get_property (GObject * object, guint prop_id,
GValue * value, GParamSpec * pspec)
{
switch (prop_id) {
case PROP_HMM_DIR:
g_value_set_string(value, cmd_ln_str("-hmm"));
break;
case PROP_LM_FILE:
g_value_set_string(value, cmd_ln_str("-lm"));
break;
case PROP_DICT_FILE:
g_value_set_string(value, cmd_ln_str("-dict"));
break;
case PROP_FSG_FILE:
g_value_set_string(value, cmd_ln_str("-fsg"));
break;
case PROP_FWDFLAT:
g_value_set_boolean(value, cmd_ln_boolean("-fwdflat"));
break;
case PROP_BESTPATH:
g_value_set_boolean(value, cmd_ln_boolean("-bestpath"));
break;
default:
G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
break;
}
}
int
main(int argc, char *argv[])
{
cmd_ln_parse(defs, orig_argc, orig_argv, TRUE);
cmd_ln_parse(defs, argc, argv, FALSE);
printf("%d %s %d %f\n",
cmd_ln_int32("-a"),
cmd_ln_str("-b") ? cmd_ln_str("-b") : "(null)",
cmd_ln_boolean("-c"),
cmd_ln_float64("-d"));
return 0;
}
static int
wr_parm()
{
if (omixw) {
if(cmd_ln_str("-omixwfn") == NULL) {
E_INFO("Please specify -omixwfn\n");
return S3_ERROR;
}
wr_mixw(cmd_ln_str("-omixwfn"));
}
if (ogau) {
if(cmd_ln_str("-ogaufn") == NULL) {
E_INFO("Please specify -ogaufn\n");
return S3_ERROR;
}
wr_gau(cmd_ln_str("-ogaufn"));
}
if (ogau_full) {
if(cmd_ln_str("-ofullgaufn") == NULL) {
E_INFO("Please specify -ofullgaufn\n");
return S3_ERROR;
}
wr_gau_full(cmd_ln_str("-ofullgaufn"));
}
if (otmat) {
if(cmd_ln_str("-otmatfn") == NULL) {
E_INFO("Please specify -otmatfn\n");
return S3_ERROR;
}
wr_tmat(cmd_ln_str("-otmatfn"));
}
return S3_SUCCESS;
}
int
main(int argc, char *argv[])
{
int32 i;
int32 n_map=0;
int32 n_class=0;
int32 *mllr_map;
char line[128];
parse_cmd_ln(argc, argv);
if (cmd_ln_str("-nmap")) {
n_map = cmd_ln_int32("-nmap");
}
else {
E_FATAL("Specify # of state -> MLLR class mappings using -nmap\n");
}
if (cmd_ln_str("-nclass")) {
n_class = cmd_ln_int32("-nclass");
}
else {
E_FATAL("Specify # of MLLR class mappings using -nclass\n");
}
if (cmd_ln_str("-cb2mllrfn") == NULL) {
E_FATAL("Specify output file using -cb2mllrfn\n");
}
mllr_map = (int32 *)ckd_calloc(n_map, sizeof(int32));
for (i = 0; i < n_map; i++) {
if (fgets(line, 128, stdin) == NULL) {
E_FATAL("Ran out of mappings at %d, but expected %d\n",
i, n_map);
}
mllr_map[i] = atoi(line);
}
if (fgets(line, 128, stdin) != NULL) {
E_WARN("Expected EOF after %d mappings, but still more data\n", n_map);
}
if (s3cb2mllr_write(cmd_ln_str("-cb2mllrfn"),
mllr_map,
n_map,
n_class) != S3_SUCCESS) {
return 1;
}
return 0;
}
int
main(int argc, char *argv[])
{
cmd_ln_parse(defn, argc, argv, TRUE);
/* Run a control file if requested. */
if (cmd_ln_str("-c")) {
if (run_control_file(cmd_ln_str("-c")) < 0)
return 1;
}
else {
if (extract_pitch(cmd_ln_str("-i"), cmd_ln_str("-o")) < 0)
return 1;
}
cmd_ln_free();
return 0;
}
float64
cluster(int32 ts,
uint32 n_stream,
uint32 n_in_frame,
uint32 *veclen,
uint32 blksize,
vector_t **mean,
uint32 n_density,
codew_t **out_label)
{
float64 sum_sqerr, sqerr=0;
uint32 s, n_frame;
const char *meth;
*out_label = NULL;
k_means_set_get_obs(&get_obs);
for (s = 0, sum_sqerr = 0; s < n_stream; s++, sum_sqerr += sqerr) {
meth = cmd_ln_str("-method");
n_frame = setup_obs(ts, s, n_in_frame, n_stream, veclen, blksize);
if (strcmp(meth, "rkm") == 0) {
sqerr = random_kmeans(cmd_ln_int32("-ntrial"),
n_frame,
veclen[s],
mean[s],
n_density,
cmd_ln_float32("-minratio"),
cmd_ln_int32("-maxiter"),
out_label);
if (sqerr < 0) {
E_ERROR("Too few observations for kmeans\n");
return -1.0;
}
}
else if (strcmp(meth, "fnkm") == 0) {
sqerr = furthest_neighbor_kmeans(n_frame,
veclen[s],
mean[s],
n_density,
cmd_ln_float32("-minratio"),
cmd_ln_int32("-maxiter"));
}
else {
E_ERROR("I don't know how to do method '%s'. Sorry.\n", meth);
}
}
return sum_sqerr;
}
int
main(int argc, char *argv[])
{
uint32 *spd;
model_def_t *mdef;
const char *tying_type;
uint32 i;
uint32 n_cb=0;
parse_cmd_ln(argc, argv);
E_INFO("Reading model definition file %s\n", cmd_ln_str("-moddeffn"));
if (model_def_read(&mdef, cmd_ln_str("-moddeffn")) != S3_SUCCESS) {
exit(1);
}
E_INFO("%d tied states defined\n", mdef->n_tied_state);
tying_type = cmd_ln_str("-tyingtype");
E_INFO("Generating state parameter definitions for %s tying\n", tying_type);
spd = ckd_calloc(mdef->n_tied_state, sizeof(uint32));
if (strcmp(tying_type, "semi") == 0) {
n_cb = 1;
}
else if (strcmp(tying_type, "pd") == 0) {
E_INFO("Phone dependent codebooks not yet implemented\n");
exit(1);
}
else if (strcmp(tying_type, "cont") == 0) {
n_cb = mdef->n_tied_state;
for (i = 0; i < mdef->n_tied_state; i++)
spd[i] = i;
}
else {
E_FATAL("Unknown tying type %s given\n", tying_type);
}
E_INFO("Writing %s\n", cmd_ln_str("-ts2cbfn"));
if (s3ts2cb_write(cmd_ln_str("-ts2cbfn"),
spd,
mdef->n_tied_state,
n_cb) != S3_SUCCESS) {
E_FATAL_SYSTEM("Unable to write %s\n", cmd_ln_str("-ts2cbfn"));
}
return 0;
}
int
srch_FSG_dump_vithist(void *srch)
{
FILE *latfp;
char file[8192];
srch_t *s;
fsg_search_t *fsgsrch;
s = (srch_t *) srch;
fsgsrch = (fsg_search_t *) s->grh->graph_struct;
sprintf(file, "%s/%s.hist", cmd_ln_str("-bptbldir"), fsgsrch->uttid);
if ((latfp = fopen(file, "w")) == NULL)
E_ERROR("fopen(%s,w) failed\n", file);
else {
fsg_history_dump(fsgsrch->history, fsgsrch->uttid, latfp,
fsgsrch->dict);
fclose(latfp);
}
return SRCH_SUCCESS;
}
/* lgalescu 2004/08/22 */
int32 save_cm_to_file(float32 *cep_means, int32 ceplen)
{
int32 i;
FILE *cmfp;
char *cmfn = cmd_ln_str("-cmsave");
/* be sure we only call this function when it makes sense */
if (cmfn == NULL)
return 0;
if ((cmfp = fopen(cmfn, "w")) == NULL) {
E_WARN("Could not open file %s. Cepstral means not saved.\n", cmfn);
return 0;
}
for (i = 0; i < ceplen; i++) {
fprintf(cmfp, "cur_mean[%d] = %.2f\n", i, cep_means[i]);
}
fclose(cmfp);
return 1;
}
static int
rd_parm()
{
if(cmd_ln_str("-imixwfn") ==NULL&&
cmd_ln_str("-igaufn") ==NULL&&
cmd_ln_str("-ifullgaufn")==NULL&&
cmd_ln_str("-itmatfn") ==NULL
) {
E_INFO("Please specify one of the following: -imixwfn, -igaufn, -ifullgaufn, -itmatfn\n");
return S3_ERROR;
}
if (cmd_ln_str("-imixwfn")) {
if(cmd_ln_str("-nmixwout")==NULL){
E_INFO("Please specify -nmixwout\n");
return S3_ERROR;
}
rd_mixw(cmd_ln_str("-imixwfn"),
cmd_ln_int32("-nmixwout"));
}
if (cmd_ln_str("-igaufn")) {
if(cmd_ln_str("-ncbout")==NULL){
E_INFO("Please specify -ncbout\n");
return S3_ERROR;
}
rd_gau(cmd_ln_str("-igaufn"),
cmd_ln_int32("-ncbout"));
}
if (cmd_ln_str("-ifullgaufn")) {
if(cmd_ln_str("-ncbout")==NULL){
E_INFO("Please specify -ncbout\n");
return S3_ERROR;
}
rd_gau_full(cmd_ln_str("-ifullgaufn"),
cmd_ln_int32("-ncbout"));
}
if (cmd_ln_str("-itmatfn")) {
if(cmd_ln_str("-ntmatout")==NULL){
E_INFO("Please specify -ntmatout\n");
return S3_ERROR;
}
rd_tmat(cmd_ln_str("-itmatfn"),
cmd_ln_int32("-ntmatout"));
}
return S3_SUCCESS;
}
int
mk_node(dtree_node_t *node,
uint32 node_id,
uint32 *id,
uint32 n_id,
float32 ****mixw,
float32 ****means,
float32 ****vars,
uint32 *veclen,
uint32 n_model,
uint32 n_state,
uint32 n_stream,
uint32 n_density,
float32 *stwt,
float32 mwfloor)
{
float32 ***mixw_occ, **dist;
uint32 mm, m, s, j, k;
float64 *dnom, norm, wt_ent, s_wt_ent, occ;
float32 mx_wt;
uint32 *l_id;
float32 ***lmeans=0,***lvars=0;
float32 varfloor=0;
uint32 continuous, sumveclen;
char* type;
type = (char *)cmd_ln_str("-ts2cbfn");
if (strcmp(type,".semi.")!=0 && strcmp(type,".cont.") != 0)
E_FATAL("Type %s unsupported; trees can only be built on types .semi. or .cont.\n",type);
if (strcmp(type,".cont.") == 0)
continuous = 1;
else
continuous = 0;
if (continuous == 1) {
varfloor = cmd_ln_float32("-varfloor");
/* Sumveclen is overallocation, but coding is simpler */
for (j=0,sumveclen=0; j < n_stream; j++) sumveclen += veclen[j];
lmeans = (float32 ***) ckd_calloc_3d(n_state,n_stream,sumveclen,sizeof(float32));
lvars = (float32 ***) ckd_calloc_3d(n_state,n_stream,sumveclen,sizeof(float32));
}
mixw_occ = (float32 ***)ckd_calloc_3d(n_state, n_stream, n_density, sizeof(float32));
dist = (float32 **)ckd_calloc_2d(n_stream, n_density, sizeof(float32));
dnom = (float64 *)ckd_calloc(n_stream, sizeof(float64));
/* Merge distributions of all the elements in a cluster for combined
distribution */
for (s = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
float32 *lmeanvec=0, *lvarvec=0;
if (continuous == 1) {
lmeanvec = lmeans[s][j];
lvarvec = lvars[s][j];
}
for (mm = 0; mm < n_id; mm++) {
m = id[mm];
for (k = 0; k < n_density; k++) {
mixw_occ[s][j][k] += mixw[m][s][j][k];
}
/* For continuous hmms we have only one gaussian per state */
if (continuous == 1) {
for (k = 0; k < veclen[j]; k++) {
lmeanvec[k] += mixw[m][s][j][0] * means[m][s][j][k];
lvarvec[k] += mixw[m][s][j][0] * (vars[m][s][j][k] +
means[m][s][j][k] * means[m][s][j][k]);
}
}
}
if (continuous == 1) {
if (mixw_occ[s][j][0] != 0) {
for (k = 0; k < veclen[j]; k++) {
lmeanvec[k] /= mixw_occ[s][j][0];
lvarvec[k] = lvarvec[k]/mixw_occ[s][j][0] -
lmeanvec[k]*lmeanvec[k];
if (lvarvec[k] < varfloor) lvarvec[k] = varfloor;
}
}
else {
for (k = 0; k < veclen[j]; k++)
if (lmeanvec[k] != 0)
E_FATAL("denominator = 0, but numerator = %f at k = %d\n",lmeanvec[k],k);
}
}
}
}
/* Find out which state is under consideration */
for (j = 0, mx_wt = 0, s = 0; s < n_state; s++) {
if (stwt[s] > mx_wt) {
mx_wt = stwt[s];
j = s;
}
}
/* occ is the same for each independent feature, so just choose 0 */
for (k = 0, occ = 0; k < n_density; k++) {
occ += mixw_occ[j][0][k];
}
for (s = 0, wt_ent = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
for (k = 0, dnom[j] = 0; k < n_density; k++) {
dnom[j] += mixw_occ[s][j][k];
}
}
for (j = 0, s_wt_ent = 0; j < n_stream; j++) {
norm = 1.0 / dnom[j];
/* discrete_entropy for discrete case, continuous entropy for
continuous HMMs */
if (continuous != 1) {
for (k = 0; k < n_density; k++) {
dist[j][k] = mixw_occ[s][j][k] * norm;
if (dist[j][k] < mwfloor)
dist[j][k] = mwfloor;
}
s_wt_ent += dnom[j] * ent_d(dist[j], n_density);
}
else {
s_wt_ent += dnom[j] * ent_cont(lmeans[s][j], lvars[s][j], veclen[j]);
}
}
wt_ent += stwt[s] * s_wt_ent;
}
node->node_id = node_id;
l_id = ckd_calloc(n_id, sizeof(uint32));
for (j = 0; j < n_id; j++) {
l_id[j] = id[j];
}
node->id = l_id;
node->n_id = n_id;
node->mixw_occ = mixw_occ;
if (continuous == 1) {
node->means = lmeans;
node->vars = lvars;
}
node->occ = occ;
node->wt_ent = wt_ent;
ckd_free_2d((void **)dist);
ckd_free((void *)dnom);
return S3_SUCCESS;
}
int
main(int argc, char *argv[])
{
vector_t ***mean;
vector_t ***var = NULL;
vector_t ****fullvar = NULL;
vector_t ***new_mean;
vector_t ***new_var = NULL;
vector_t ****new_fullvar = NULL;
float32 ***dnom;
float32 ***mixw;
float32 ***new_mixw;
uint32 n_mixw;
uint32 n_mgau;
uint32 n_dnom;
uint32 n_feat;
uint32 n_density;
uint32 n_inc;
uint32 *veclen;
int32 var_is_full;
parse_cmd_ln(argc, argv);
E_INFO("Reading mixing weight file %s.\n",
cmd_ln_str("-inmixwfn"));
if (s3mixw_read(cmd_ln_str("-inmixwfn"),
&mixw,
&n_mixw,
&n_feat,
&n_density) != S3_SUCCESS) {
return 1;
}
n_inc = cmd_ln_int32("-ninc");
if (n_inc > n_density) {
E_WARN("# of densities to split (== %u) > total # of densities/mixture (== %u); # split <- %u # den/mix\n",
n_inc, n_density, n_density);
n_inc = n_density;
}
if (s3gau_read(cmd_ln_str("-inmeanfn"),
&mean,
&n_mgau,
&n_feat,
&n_density,
&veclen) != S3_SUCCESS) {
return 1;
}
var_is_full = cmd_ln_int32("-fullvar");
if (var_is_full) {
if (s3gau_read_full(cmd_ln_str("-invarfn"),
&fullvar,
&n_mgau,
&n_feat,
&n_density,
&veclen) != S3_SUCCESS) {
return 1;
}
}
else {
if (s3gau_read(cmd_ln_str("-invarfn"),
&var,
&n_mgau,
&n_feat,
&n_density,
&veclen) != S3_SUCCESS) {
return 1;
}
}
if (s3gaudnom_read(cmd_ln_str("-dcountfn"),
&dnom,
&n_dnom,
&n_feat,
&n_density) != S3_SUCCESS) {
return 1;
}
new_mean = gauden_alloc_param(n_mgau, n_feat, n_density+n_inc, veclen);
if (var_is_full)
new_fullvar = gauden_alloc_param_full(n_mgau, n_feat, n_density+n_inc, veclen);
else
new_var = gauden_alloc_param(n_mgau, n_feat, n_density+n_inc, veclen);
new_mixw = (float32 ***)ckd_calloc_3d(n_mixw, n_feat, n_density+n_inc,
sizeof(float32));
E_INFO("output n_density == %u\n", n_density+n_inc);
inc_densities(new_mixw,
new_mean,
new_var,
new_fullvar,
mixw,
mean,
var,
fullvar,
dnom,
n_mixw,
n_mgau,
n_dnom,
n_feat,
n_density,
veclen,
n_inc);
if (cmd_ln_str("-outmixwfn") != NULL) {
if (s3mixw_write(cmd_ln_str("-outmixwfn"),
new_mixw,
n_mixw,
n_feat,
n_density+n_inc) != S3_SUCCESS) {
return 1;
}
}
else {
E_FATAL("You must use the -outmixwfn argument\n");
}
if (cmd_ln_str("-outmeanfn") != NULL) {
if (s3gau_write(cmd_ln_str("-outmeanfn"),
(const vector_t ***)new_mean,
n_mgau,
n_feat,
n_density+n_inc,
veclen) != S3_SUCCESS) {
return 1;
}
}
else {
E_FATAL("You must use the -outmeanfn argument\n");
}
if (cmd_ln_str("-outvarfn") != NULL) {
if (var_is_full) {
if (s3gau_write_full(cmd_ln_str("-outvarfn"),
(const vector_t ****)new_fullvar,
n_mgau,
n_feat,
n_density+n_inc,
veclen) != S3_SUCCESS) {
return 1;
}
}
else {
if (s3gau_write(cmd_ln_str("-outvarfn"),
(const vector_t ***)new_var,
n_mgau,
n_feat,
n_density+n_inc,
veclen) != S3_SUCCESS) {
return 1;
}
}
}
else {
E_FATAL("You must use the -outvarfn argument\n");
}
return 0;
}
int
init(model_def_t **out_imdef,
pset_t **out_pset,
uint32 *out_n_pset,
dtree_t ****out_tree,
uint32 *out_n_seno)
{
model_def_t *imdef;
uint32 p, s;
uint32 n_ci, n_state;
char fn[MAXPATHLEN+1];
const char *a_fn;
FILE *fp;
dtree_t ***tree, *tr;
pset_t *pset;
uint32 n_pset;
uint32 n_seno;
const char *treedir;
uint32 ts_id;
int allphones;
a_fn = cmd_ln_str("-imoddeffn");
if (a_fn == NULL)
E_FATAL("Specify -imoddeffn\n");
if (model_def_read(&imdef, a_fn) != S3_SUCCESS) {
return S3_ERROR;
}
*out_imdef = imdef;
a_fn = cmd_ln_str("-psetfn");
E_INFO("Reading: %s\n", a_fn);
*out_pset = pset = read_pset_file(a_fn, imdef->acmod_set, &n_pset);
*out_n_pset = n_pset;
allphones = cmd_ln_int32("-allphones");
if (allphones)
n_ci = 1;
else
n_ci = acmod_set_n_ci(imdef->acmod_set);
treedir = cmd_ln_str("-treedir");
tree = (dtree_t ***)ckd_calloc(n_ci, sizeof(dtree_t **));
*out_tree = tree;
ts_id = imdef->n_tied_ci_state;
for (p = 0, n_seno = 0; p < n_ci; p++) {
if (allphones || !acmod_set_has_attrib(imdef->acmod_set, p, "filler")) {
const char *pname;
if (allphones) {
n_state = imdef->defn[acmod_set_n_ci(imdef->acmod_set)].n_state;
pname = "ALLPHONES";
}
else {
n_state = imdef->defn[p].n_state;
pname = acmod_set_id2name(imdef->acmod_set, p);
}
tree[p] = (dtree_t **)ckd_calloc(n_state, sizeof(dtree_t *));
for (s = 0; s < n_state-1; s++) {
E_INFO("%s-%u: offset %u\n",
pname, s, ts_id);
sprintf(fn, "%s/%s-%u.dtree",
treedir, pname, s);
fp = fopen(fn, "r");
if (fp == NULL) {
E_FATAL_SYSTEM("Unable to open %s for reading", fn);
}
tree[p][s] = tr = read_final_tree(fp, pset, n_pset);
label_leaves(&tr->node[0], &ts_id);
fclose(fp);
n_seno += cnt_leaf(&tr->node[0]);
}
}
}
assert(n_seno == (ts_id - imdef->n_tied_ci_state));
E_INFO("n_seno= %u\n", ts_id);
*out_n_seno = n_seno;
return S3_SUCCESS;
}
void
cluster_leaves(dtree_t *tr,
uint32 *veclen,
float64 *wt_ent_dec,
uint32 *out_n_a,
uint32 *out_n_b,
pset_t *pset,
uint32 n_state,
uint32 n_stream,
uint32 n_density,
float32 *stwt,
float32 mwfloor)
{
uint32 n_leaf;
float32 ****mixw_occ;
uint32 *clust, n_a, n_b;
uint32 *node_id;
dtree_node_t *root;
uint32 i;
float32 ****means=0;
float32 ****vars=0;
const char* type;
uint32 continuous, sumveclen;
type = cmd_ln_str("-ts2cbfn");
if (strcmp(type,".semi.")!=0 && strcmp(type,".cont.") != 0)
E_FATAL("Type %s unsupported; trees can only be built on types .semi. or .cont.\n",type);
if (strcmp(type,".cont.") == 0)
continuous = 1;
else
continuous = 0;
root = &tr->node[0];
/* determine the # of leaf nodes in the simple tree */
n_leaf = cnt_leaf(root);
if (continuous == 1) {
for (i=0,sumveclen=0; i < n_stream; i++) sumveclen += veclen[i];
means = (float32 ****)ckd_calloc_4d(n_leaf, n_state, n_stream, sumveclen, sizeof(float32));
vars = (float32 ****)ckd_calloc_4d(n_leaf, n_state, n_stream, sumveclen, sizeof(float32));
}
/* Alloc space for:
* - leaf node distribution array
* - leaf node cluster id array
* - leaf node id array
*/
mixw_occ = (float32 ****)ckd_calloc_4d(n_leaf, n_state, n_stream, n_density, sizeof(float32));
clust = (uint32 *)ckd_calloc(n_leaf, sizeof(uint32));
node_id = (uint32 *)ckd_calloc(n_leaf, sizeof(uint32));
/* compute the density occupancies of the leaves */
leaf_mixw_occ(root, pset, mixw_occ, node_id, n_state, n_stream, n_density, 0);
if (continuous == 1) {
/* compute means and variances of the leaves */
leaf_mean_vars(root, pset, means, vars, node_id, n_state, n_stream, veclen, 0);
}
/* Cluster the leaf nodes into two classes */
*wt_ent_dec = two_class(mixw_occ, means, vars, veclen, n_leaf, n_state, n_stream, n_density, stwt, clust, mwfloor);
for (i = 0; i < n_leaf; i++) {
tr->node[node_id[i]].clust = clust[i];
}
/* Simplify the tree based on the two classes
* (i.e. if siblings belong to the same class,
* delete the node) */
prune_leaves(root, pset);
/* Determine how many leaf nodes in class A and B
* in the simplified tree */
n_a = n_b = 0;
cnt_class(root, &n_a, &n_b);
#if 0
fprintf(stderr, "Pruned tree %u/%u:\n", n_a, n_b);
print_tree(stderr, "|", root, pset, 1);
fprintf(stderr, "\n");
#endif
*out_n_a = n_a;
*out_n_b = n_b;
}
int32
mmi_viterbi_update(vector_t **feature,
uint32 n_obs,
state_t *state_seq,
uint32 n_state,
model_inventory_t *inv,
float64 a_beam,
int32 mean_reest,
int32 var_reest,
float64 arc_gamma,
feat_t *fcb)
{
float64 *scale = NULL;
float64 **dscale = NULL;
float64 **active_alpha;
uint32 **active_astate;
uint32 **bp;
uint32 *n_active_astate;
gauden_t *g;/* Gaussian density parameters and reestimation sums */
float32 ***mixw;/* all mixing weights */
float64 ***now_den = NULL;/* Short for den[t] */
uint32 ***now_den_idx = NULL;/* Short for den_idx[t] */
uint32 *active_cb;
uint32 n_active_cb;
float32 ***denacc = NULL;/* mean/var reestimation accumulators for time t */
size_t denacc_size;/* Total size of data references in denacc. Allows
for quick clears between time frames */
uint32 n_lcl_cb;
uint32 *cb_inv;
uint32 i, j, q;
int32 t;
uint32 n_feat;
uint32 n_density;
uint32 n_top;
int ret;
uint32 n_cb;
static float64 *p_op = NULL;
static float64 *p_ci_op = NULL;
static float64 **d_term = NULL;
static float64 **d_term_ci = NULL;
/* caller must ensure that there is some non-zero amount
of work to be done here */
assert(n_obs > 0);
assert(n_state > 0);
g = inv->gauden;
n_feat = gauden_n_feat(g);
n_density = gauden_n_density(g);
n_top = gauden_n_top(g);
n_cb = gauden_n_mgau(g);
if (p_op == NULL) {
p_op = ckd_calloc(n_feat, sizeof(float64));
p_ci_op = ckd_calloc(n_feat, sizeof(float64));
}
if (d_term == NULL) {
d_term = (float64 **)ckd_calloc_2d(n_feat, n_top, sizeof(float64));
d_term_ci = (float64 **)ckd_calloc_2d(n_feat, n_top, sizeof(float64));
}
scale = (float64 *)ckd_calloc(n_obs, sizeof(float64));
dscale = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
n_active_astate = (uint32 *)ckd_calloc(n_obs, sizeof(uint32));
active_alpha = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
active_astate = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
active_cb = ckd_calloc(2*n_state, sizeof(uint32));
bp = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
/* Run forward algorithm, which has embedded Viterbi. */
ret = forward(active_alpha, active_astate, n_active_astate, bp,
scale, dscale,
feature, n_obs, state_seq, n_state,
inv, a_beam, NULL, 1);
if (cmd_ln_str("-outphsegdir")) {
E_FATAL("current MMI implementation don't support -outphsegdir\n");
}
if (ret != S3_SUCCESS) {
/* Some problem with the utterance, release per utterance storage and
* forget about adding the utterance accumulators to the global accumulators */
goto all_done;
}
mixw = inv->mixw;
n_lcl_cb = inv->n_cb_inverse;
cb_inv = inv->cb_inverse;
/* Allocate local accumulators for mean, variance reestimation
sums if necessary */
gauden_alloc_l_acc(g, n_lcl_cb,
mean_reest, var_reest,
FALSE);
n_active_cb = 0;
now_den = (float64 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_top,
sizeof(float64));
now_den_idx = (uint32 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_top,
sizeof(uint32));
if (mean_reest || var_reest) {
/* allocate space for the per frame density counts */
denacc = (float32 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_density,
sizeof(float32));
/* # of bytes required to store all weighted vectors */
denacc_size = n_lcl_cb * n_feat * n_density * sizeof(float32);
}
else {
denacc = NULL;
denacc_size = 0;
}
/* Okay now run through the backtrace and accumulate counts. */
/* Find the non-emitting ending state */
for (q = 0; q < n_active_astate[n_obs-1]; ++q) {
if (active_astate[n_obs-1][q] == n_state-1)
break;
}
if (q == n_active_astate[n_obs-1]) {
E_ERROR("Failed to align audio to trancript: final state of the search is not reached\n");
ret = S3_ERROR;
goto all_done;
}
for (t = n_obs-1; t >= 0; --t) {
uint32 l_cb;
uint32 l_ci_cb;
float64 op, p_reest_term;
uint32 prev;
j = active_astate[t][q];
/* Follow any non-emitting states at time t first. */
while (state_seq[j].mixw == TYING_NON_EMITTING) {
prev = active_astate[t][bp[t][q]];
q = bp[t][q];
j = prev;
}
/* Now accumulate statistics for the real state. */
l_cb = state_seq[j].l_cb;
l_ci_cb = state_seq[j].l_ci_cb;
n_active_cb = 0;
gauden_compute_log(now_den[l_cb],
now_den_idx[l_cb],
feature[t],
g,
state_seq[j].cb,
NULL);
active_cb[n_active_cb++] = l_cb;
if (l_cb != l_ci_cb) {
gauden_compute_log(now_den[l_ci_cb],
now_den_idx[l_ci_cb],
feature[t],
g,
state_seq[j].ci_cb,
NULL);
active_cb[n_active_cb++] = l_ci_cb;
}
ret = gauden_scale_densities_bwd(now_den, now_den_idx,
&dscale[t],
active_cb, n_active_cb, g);
if (ret != S3_SUCCESS)
goto all_done;
assert(state_seq[j].mixw != TYING_NON_EMITTING);
/* Now calculate mixture densities. */
/* This is the normalizer sum_m c_{jm} p(o_t|\lambda_{jm}) */
op = gauden_mixture(now_den[l_cb], now_den_idx[l_cb],
mixw[state_seq[j].mixw], g);
/* Make up this bogus value to be consistent with backward.c */
p_reest_term = 1.0 / op;
/* Compute the output probability excluding the contribution
* of each feature stream. i.e. p_op[0] is the output
* probability excluding feature stream 0 */
partial_op(p_op,
op,
now_den[l_cb],
now_den_idx[l_cb],
mixw[state_seq[j].mixw],
n_feat,
n_top);
/* compute the probability of each (of possibly topn) density */
den_terms(d_term,
p_reest_term,
p_op,
now_den[l_cb],
now_den_idx[l_cb],
mixw[state_seq[j].mixw],
n_feat,
n_top);
if (l_cb != l_ci_cb) {
/* For each feature stream f, compute:
* sum_k(mixw[f][k] den[f][k])
* and store the results in p_ci_op */
partial_ci_op(p_ci_op,
now_den[l_ci_cb],
now_den_idx[l_ci_cb],
mixw[state_seq[j].ci_mixw],
n_feat,
n_top);
/* For each feature stream and density compute the terms:
* w[f][k] den[f][k] / sum_k(w[f][k] den[f][k]) * post_j
* and store results in d_term_ci */
den_terms_ci(d_term_ci,
1.0, /* post_j = 1.0 */
p_ci_op,
now_den[l_ci_cb],
now_den_idx[l_ci_cb],
mixw[state_seq[j].ci_mixw],
n_feat,
n_top);
}
/* accumulate the probability for each density in the
* density reestimation accumulators */
if (mean_reest || var_reest) {
accum_den_terms(denacc[l_cb], d_term,
now_den_idx[l_cb], n_feat, n_top);
if (l_cb != l_ci_cb) {
accum_den_terms(denacc[l_ci_cb], d_term_ci,
now_den_idx[l_ci_cb], n_feat, n_top);
}
}
/* Note that there is only one state/frame so this is kind of
redundant */
if (mean_reest || var_reest) {
/* Update the mean and variance reestimation accumulators */
mmi_accum_gauden(denacc,
cb_inv,
n_lcl_cb,
feature[t],
now_den_idx,
g,
mean_reest,
var_reest,
arc_gamma,
fcb);
memset(&denacc[0][0][0], 0, denacc_size);
}
if (t > 0) {
prev = active_astate[t-1][bp[t][q]];
q = bp[t][q];
j = prev;
}
}
/* If no error was found, add the resulting utterance reestimation
* accumulators to the global reestimation accumulators */
accum_global(inv, state_seq, n_state,
FALSE, FALSE, mean_reest, var_reest,
FALSE);
all_done:
ckd_free((void *)scale);
for (i = 0; i < n_obs; i++) {
if (dscale[i])
ckd_free((void *)dscale[i]);
}
ckd_free((void **)dscale);
ckd_free(n_active_astate);
for (i = 0; i < n_obs; i++) {
ckd_free((void *)active_alpha[i]);
ckd_free((void *)active_astate[i]);
ckd_free((void *)bp[i]);
}
ckd_free((void *)active_alpha);
ckd_free((void *)active_astate);
ckd_free((void *)active_cb);
ckd_free((void **)bp);
if (denacc)
ckd_free_3d((void ***)denacc);
if (now_den)
ckd_free_3d((void ***)now_den);
if (now_den_idx)
ckd_free_3d((void ***)now_den_idx);
if (ret != S3_SUCCESS)
E_ERROR("viterbi update error in sentence %s\n", corpus_utt_brief_name());
return ret;
}
int
main(int argc, char *argv[])
{
model_def_t *imdef;
model_def_t *omdef;
pset_t *pset;
uint32 n_pset;
dtree_t ***tree;
uint32 n_seno;
uint32 n_ci;
uint32 n_acmod;
uint32 p;
uint32 s;
model_def_entry_t *idefn, *odefn;
acmod_id_t b, l, r;
word_posn_t wp;
int allphones;
parse_cmd_ln(argc, argv);
if (init(&imdef, &pset, &n_pset, &tree, &n_seno) != S3_SUCCESS)
return 1;
omdef = (model_def_t *)ckd_calloc(1, sizeof(model_def_t));
omdef->acmod_set = imdef->acmod_set; /* same set of acoustic models */
omdef->n_total_state = imdef->n_total_state;
omdef->n_tied_ci_state = imdef->n_tied_ci_state;
omdef->n_tied_state = imdef->n_tied_ci_state + n_seno;
omdef->n_tied_tmat = imdef->n_tied_tmat;
omdef->defn = (model_def_entry_t *)ckd_calloc(imdef->n_defn,
sizeof(model_def_entry_t));
/*
* Define the context-independent models
*/
n_ci = acmod_set_n_ci(imdef->acmod_set);
for (p = 0; p < n_ci; p++) {
idefn = &imdef->defn[p];
odefn = &omdef->defn[p];
odefn->p = idefn->p;
odefn->tmat = idefn->tmat;
odefn->state = ckd_calloc(idefn->n_state, sizeof(uint32));
odefn->n_state = idefn->n_state;
for (s = 0; s < idefn->n_state; s++) {
if (idefn->state[s] == NO_ID)
odefn->state[s] = NO_ID;
else {
odefn->state[s] = idefn->state[s];
}
}
}
/*
* Define the rest of the models
*/
allphones = cmd_ln_int32("-allphones");
n_acmod = acmod_set_n_acmod(omdef->acmod_set);
for (; p < n_acmod; p++) {
b = acmod_set_base_phone(omdef->acmod_set, p);
assert(p != b);
idefn = &imdef->defn[p];
odefn = &omdef->defn[p];
odefn->p = idefn->p;
odefn->tmat = idefn->tmat;
odefn->state = ckd_calloc(idefn->n_state, sizeof(uint32));
odefn->n_state = idefn->n_state;
for (s = 0; s < idefn->n_state; s++) {
if (idefn->state[s] == NO_ID)
/* Non-emitting state */
odefn->state[s] = NO_ID;
else {
uint32 bb;
/* emitting state: find the tied state */
acmod_set_id2tri(omdef->acmod_set,
&b, &l, &r, &wp,
p);
#ifdef HORRIBLY_VERBOSE
fprintf(stderr, "%s %u ",
acmod_set_id2name(omdef->acmod_set, p), s);
#endif
bb = allphones ? 0 : b;
odefn->state[s] = tied_state(&tree[bb][s]->node[0],
b, l, r, wp,
pset);
#ifdef HORRIBLY_VERBOSE
fprintf(stderr, "\t-> %u\n", odefn->state[s]);
fprintf(stderr, "\n");
#endif
}
}
}
if (model_def_write(omdef, cmd_ln_str("-omoddeffn")) != S3_SUCCESS) {
return 1;
}
return 0;
}
int
main(int argc, char *argv[])
{
lexicon_t *lex;
model_def_t *omdef;
model_def_t *dmdef;
feat_t *feat;
uint32 n_stream, blksize;
uint32 *veclen;
uint32 ts_off;
uint32 ts_cnt;
FILE *fp;
if (main_initialize(argc, argv, &lex, &omdef, &dmdef, &feat) != S3_SUCCESS) {
return -1;
}
n_stream = feat_dimension1(feat);
veclen = feat_stream_lengths(feat);
blksize = feat_dimension(feat);
if (strcmp(cmd_ln_str("-gthobj"), "state") == 0) {
ts_off = cmd_ln_int32("-tsoff");
if (cmd_ln_str("-tscnt") == NULL) {
ts_cnt = omdef->n_tied_state - ts_off;
}
else {
ts_cnt = cmd_ln_int32("-tscnt");
}
if (ts_off + ts_cnt > omdef->n_tied_state) {
E_FATAL("Too many tied states specified\n");
}
n_tot_frame = 0;
ptmr_reset(&all_timer);
ptmr_reset(&km_timer);
ptmr_reset(&var_timer);
ptmr_reset(&em_timer);
ptmr_start(&all_timer);
if (init_state(cmd_ln_str("-segdmpfn"),
cmd_ln_str("-segidxfn"),
cmd_ln_int32("-ndensity"),
n_stream,
veclen,
blksize,
cmd_ln_int32("-reest"),
cmd_ln_str("-mixwfn"),
cmd_ln_str("-meanfn"),
cmd_ln_str("-varfn"),
ts_off,
ts_cnt,
omdef->n_tied_state,
(dmdef != NULL ? dmdef->n_tied_state : omdef->n_tied_state))
!= S3_SUCCESS) {
E_ERROR("Unable to train [%u %u]\n", ts_off, ts_off+ts_cnt-1);
}
ptmr_stop(&all_timer);
if (n_tot_frame > 0) {
E_INFO("TOTALS:");
E_INFOCONT(" km %4.3fx %4.3e",
km_timer.t_cpu / (n_tot_frame * 0.01),
(km_timer.t_cpu > 0 ?
km_timer.t_elapsed / km_timer.t_cpu : 0.0));
E_INFOCONT(" var %4.3fx %4.3e",
var_timer.t_cpu / (n_tot_frame * 0.01),
(var_timer.t_cpu > 0 ?
var_timer.t_elapsed / var_timer.t_cpu : 0.0));
E_INFOCONT(" em %4.3fx %4.3e",
em_timer.t_cpu / (n_tot_frame * 0.01),
(em_timer.t_cpu > 0 ?
em_timer.t_elapsed / em_timer.t_cpu : 0.0));
E_INFOCONT(" all %4.3fx %4.3e",
all_timer.t_cpu / (n_tot_frame * 0.01),
(all_timer.t_cpu > 0 ?
all_timer.t_elapsed / all_timer.t_cpu : 0.0));
E_INFOCONT("\n");
}
if (cmd_ln_str("-tsrngfn") != NULL) {
fp = fopen(cmd_ln_str("-tsrngfn"),
"w");
if (fp == NULL) {
E_FATAL_SYSTEM("Unable to open %s for reading",
cmd_ln_str("-tsrngfn"));
}
fprintf(fp, "%d %d\n", ts_off, ts_cnt);
}
else if (ts_cnt != omdef->n_tied_state) {
E_WARN("Subset of tied states specified, but no -tsrngfn arg");
}
}
else if (strcmp(cmd_ln_str("-gthobj"), "single") == 0) {
n_tot_frame = 0;
ptmr_reset(&all_timer);
ptmr_reset(&km_timer);
ptmr_reset(&var_timer);
ptmr_reset(&em_timer);
ptmr_start(&all_timer);
if (init_state(cmd_ln_str("-segdmpfn"),
NULL, /* No index -> single class dump file */
cmd_ln_int32("-ndensity"),
n_stream,
veclen,
blksize,
cmd_ln_int32("-reest"),
cmd_ln_str("-mixwfn"),
cmd_ln_str("-meanfn"),
cmd_ln_str("-varfn"),
0,
1,
1,
1) != S3_SUCCESS) {
E_ERROR("Unable to train\n");
}
ptmr_stop(&all_timer);
if (n_tot_frame > 0) {
E_INFO("TOTALS:");
E_INFOCONT(" km %4.3fx %4.3e",
km_timer.t_cpu / (n_tot_frame * 0.01),
(km_timer.t_cpu > 0 ?
km_timer.t_elapsed / km_timer.t_cpu : 0.0));
E_INFOCONT(" var %4.3fx %4.3e",
var_timer.t_cpu / (n_tot_frame * 0.01),
(var_timer.t_cpu > 0 ?
var_timer.t_elapsed / var_timer.t_cpu : 0.0));
E_INFOCONT(" em %4.3fx %4.3e",
em_timer.t_cpu / (n_tot_frame * 0.01),
(em_timer.t_cpu > 0 ?
em_timer.t_elapsed / em_timer.t_cpu : 0.0));
E_INFOCONT(" all %4.3fx %4.3e",
all_timer.t_cpu / (n_tot_frame * 0.01),
(all_timer.t_cpu > 0 ?
all_timer.t_elapsed / all_timer.t_cpu : 0.0));
E_INFOCONT("\n");
}
}
return 0;
}
int
main_initialize(int argc,
char *argv[],
lexicon_t **out_lex,
model_def_t **out_omdef,
model_def_t **out_dmdef,
feat_t** out_feat)
{
model_def_t *dmdef = NULL;
model_def_t *omdef = NULL;
lexicon_t *lex = NULL;
feat_t *feat;
const char *fn;
uint32 n_ts;
uint32 n_cb;
const char *ts2cbfn;
parse_cmd_ln(argc, argv);
feat =
feat_init(cmd_ln_str("-feat"),
cmn_type_from_str(cmd_ln_str("-cmn")),
cmd_ln_boolean("-varnorm"),
agc_type_from_str(cmd_ln_str("-agc")),
1, cmd_ln_int32("-ceplen"));
if (cmd_ln_str("-lda")) {
E_INFO("Reading linear feature transformation from %s\n",
cmd_ln_str("-lda"));
if (feat_read_lda(feat,
cmd_ln_str("-lda"),
cmd_ln_int32("-ldadim")) < 0)
return -1;
}
if (cmd_ln_str("-svspec")) {
int32 **subvecs;
E_INFO("Using subvector specification %s\n",
cmd_ln_str("-svspec"));
if ((subvecs = parse_subvecs(cmd_ln_str("-svspec"))) == NULL)
return -1;
if ((feat_set_subvecs(feat, subvecs)) < 0)
return -1;
}
if (cmd_ln_exists("-agcthresh")
&& 0 != strcmp(cmd_ln_str("-agc"), "none")) {
agc_set_threshold(feat->agc_struct,
cmd_ln_float32("-agcthresh"));
}
if (feat->cmn_struct
&& cmd_ln_exists("-cmninit")) {
char *c, *cc, *vallist;
int32 nvals;
vallist = ckd_salloc(cmd_ln_str("-cmninit"));
c = vallist;
nvals = 0;
while (nvals < feat->cmn_struct->veclen
&& (cc = strchr(c, ',')) != NULL) {
*cc = '\0';
feat->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof(c));
c = cc + 1;
++nvals;
}
if (nvals < feat->cmn_struct->veclen && *c != '\0') {
feat->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof(c));
}
ckd_free(vallist);
}
*out_feat = feat;
if (cmd_ln_str("-omoddeffn")) {
E_INFO("Reading output model definitions: %s\n", cmd_ln_str("-omoddeffn"));
/* Read in the model definitions. Defines the set of
CI phones and context dependent phones. Defines the
transition matrix tying and state level tying. */
if (model_def_read(&omdef,
cmd_ln_str("-omoddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
if (cmd_ln_str("-dmoddeffn")) {
E_INFO("Reading dump model definitions: %s\n", cmd_ln_str("-dmoddeffn"));
if (model_def_read(&dmdef,
cmd_ln_str("-dmoddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
setup_d2o_map(dmdef, omdef);
}
else {
E_INFO("Assuming dump and output model definitions are identical\n");
}
ts2cbfn = cmd_ln_str("-ts2cbfn");
if (ts2cbfn) {
if (strcmp(SEMI_LABEL, ts2cbfn) == 0) {
omdef->cb = semi_ts2cb(omdef->n_tied_state);
n_ts = omdef->n_tied_state;
n_cb = 1;
}
else if (strcmp(CONT_LABEL, ts2cbfn) == 0) {
omdef->cb = cont_ts2cb(omdef->n_tied_state);
n_ts = omdef->n_tied_state;
n_cb = omdef->n_tied_state;
}
else if (strcmp(PTM_LABEL, ts2cbfn) == 0) {
omdef->cb = ptm_ts2cb(omdef);
n_ts = omdef->n_tied_state;
n_cb = omdef->acmod_set->n_ci;
}
else if (s3ts2cb_read(cmd_ln_str("-ts2cbfn"),
&omdef->cb,
&n_ts,
&n_cb) != S3_SUCCESS) {
return S3_ERROR;
}
if (omdef->n_tied_state != n_ts) {
E_FATAL("Model definition file n_tied_state = %u, but %u mappings in ts2cb\n",
omdef->n_tied_state, n_ts);
}
}
}
else {
E_INFO("No mdef files. Assuming 1-class init\n");
}
*out_omdef = omdef;
*out_dmdef = dmdef;
fn = cmd_ln_str("-dictfn");
if (fn) {
E_INFO("Reading main lexicon: %s\n", fn);
lex = lexicon_read(NULL,
fn,
omdef->acmod_set);
if (lex == NULL)
return S3_ERROR;
}
fn = cmd_ln_str("-fdictfn");
if (fn) {
E_INFO("Reading filler lexicon: %s\n", fn);
(void)lexicon_read(lex,
fn,
omdef->acmod_set);
}
*out_lex = lex;
stride = cmd_ln_int32("-stride");
return S3_SUCCESS;
}
/*ARCHAN, to allow backward compatibility -lm, -lmctlfn coexists. This makes the current implmentation more complicated than necessary. */
void kb_init (kb_t *kb)
{
kbcore_t *kbcore;
mdef_t *mdef;
dict_t *dict;
dict2pid_t *d2p;
lm_t *lm;
lmset_t *lmset;
s3cipid_t sil, ci;
s3wid_t w;
int32 i, n, n_lc;
wordprob_t *wp;
s3cipid_t *lc;
bitvec_t lc_active;
char *str;
int32 cisencnt;
int32 j;
/* Initialize the kb structure to zero, just in case */
memset(kb, 0, sizeof(*kb));
kb->kbcore = NULL;
kb->kbcore = kbcore_init (cmd_ln_float32 ("-logbase"),
cmd_ln_str("-feat"),
cmd_ln_str("-cmn"),
cmd_ln_str("-varnorm"),
cmd_ln_str("-agc"),
cmd_ln_str("-mdef"),
cmd_ln_str("-dict"),
cmd_ln_str("-fdict"),
"", /* Hack!! Hardwired constant
for -compsep argument */
cmd_ln_str("-lm"),
cmd_ln_str("-lmctlfn"),
cmd_ln_str("-lmdumpdir"),
cmd_ln_str("-fillpen"),
cmd_ln_str("-senmgau"),
cmd_ln_float32("-silprob"),
cmd_ln_float32("-fillprob"),
cmd_ln_float32("-lw"),
cmd_ln_float32("-wip"),
cmd_ln_float32("-uw"),
cmd_ln_str("-mean"),
cmd_ln_str("-var"),
cmd_ln_float32("-varfloor"),
cmd_ln_str("-mixw"),
cmd_ln_float32("-mixwfloor"),
cmd_ln_str("-subvq"),
cmd_ln_str("-gs"),
cmd_ln_str("-tmat"),
cmd_ln_float32("-tmatfloor"));
if(kb->kbcore==NULL){
E_FATAL("Initialization of kb failed\n");
}
kbcore = kb->kbcore;
mdef = kbcore_mdef(kbcore);
dict = kbcore_dict(kbcore);
lm = kbcore_lm(kbcore);
lmset=kbcore_lmset(kbcore);
d2p = kbcore_dict2pid(kbcore);
if (NOT_S3WID(dict_startwid(dict)) || NOT_S3WID(dict_finishwid(dict)))
E_FATAL("%s or %s not in dictionary\n", S3_START_WORD, S3_FINISH_WORD);
if(lmset){
for(i=0;i<kbcore_nlm(kbcore);i++){
if (NOT_S3LMWID(lm_startwid(lmset[i].lm)) || NOT_S3LMWID(lm_finishwid(lmset[i].lm)))
E_FATAL("%s or %s not in LM %s\n", S3_START_WORD, S3_FINISH_WORD,lmset[i].name);
}
}else if(lm){
if (NOT_S3LMWID(lm_startwid(lm)) || NOT_S3LMWID(lm_finishwid(lm)))
E_FATAL("%s or %s not in LM\n", S3_START_WORD, S3_FINISH_WORD);
}
/* Check that HMM topology restrictions are not violated */
if (tmat_chk_1skip (kbcore->tmat) < 0)
E_FATAL("Tmat contains arcs skipping more than 1 state\n");
/*
* Unlink <s> and </s> between dictionary and LM, to prevent their
* recognition. They are merely dummy words (anchors) at the beginning
* and end of each utterance.
*/
if(lmset){
for(i=0;i<kbcore_nlm(kbcore);i++){
lm_lmwid2dictwid(lmset[i].lm, lm_startwid(lmset[i].lm)) = BAD_S3WID;
lm_lmwid2dictwid(lmset[i].lm, lm_finishwid(lmset[i].lm)) = BAD_S3WID;
for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
lmset[i].lm->dict2lmwid[w] = BAD_S3LMWID;
for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
lmset[i].lm->dict2lmwid[w] = BAD_S3LMWID;
}
}else if(lm){ /* No LM is set at this point*/
lm_lmwid2dictwid(lm, lm_startwid(lm)) = BAD_S3WID;
lm_lmwid2dictwid(lm, lm_finishwid(lm)) = BAD_S3WID;
for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
kbcore->dict2lmwid[w] = BAD_S3LMWID;
for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
kbcore->dict2lmwid[w] = BAD_S3LMWID;
}
sil = mdef_silphone (kbcore_mdef (kbcore));
if (NOT_S3CIPID(sil))
E_FATAL("Silence phone '%s' not in mdef\n", S3_SILENCE_CIPHONE);
kb->sen_active = (int32 *) ckd_calloc (mdef_n_sen(mdef), sizeof(int32));
kb->rec_sen_active = (int32 *) ckd_calloc (mdef_n_sen(mdef), sizeof(int32));
kb->ssid_active = (int32 *) ckd_calloc (mdef_n_sseq(mdef), sizeof(int32));
kb->comssid_active = (int32 *) ckd_calloc (dict2pid_n_comsseq(d2p), sizeof(int32));
/* Build set of all possible left contexts */
lc = (s3cipid_t *) ckd_calloc (mdef_n_ciphone(mdef) + 1, sizeof(s3cipid_t));
lc_active = bitvec_alloc (mdef_n_ciphone (mdef));
for (w = 0; w < dict_size (dict); w++) {
ci = dict_pron (dict, w, dict_pronlen(dict, w) - 1);
if (! mdef_is_fillerphone (mdef, (int)ci))
bitvec_set (lc_active, ci);
}
ci = mdef_silphone(mdef);
bitvec_set (lc_active, ci);
for (ci = 0, n_lc = 0; ci < mdef_n_ciphone(mdef); ci++) {
if (bitvec_is_set (lc_active, ci))
lc[n_lc++] = ci;
}
lc[n_lc] = BAD_S3CIPID;
E_INFO("Building lextrees\n");
/* Get the number of lexical tree*/
kb->n_lextree = cmd_ln_int32 ("-Nlextree");
if (kb->n_lextree < 1) {
E_ERROR("No. of ugtrees specified: %d; will instantiate 1 ugtree\n",
kb->n_lextree);
kb->n_lextree = 1;
}
/* ARCHAN: This code was rearranged in s3.4 implementation of dynamic LM */
/* Build active word list */
wp = (wordprob_t *) ckd_calloc (dict_size(dict), sizeof(wordprob_t));
if(lmset){
kb->ugtreeMulti = (lextree_t **) ckd_calloc (kbcore_nlm(kbcore)*kb->n_lextree, sizeof(lextree_t *));
/* Just allocate pointers*/
kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
for(i=0;i<kbcore_nlm(kbcore);i++){
E_INFO("Creating Unigram Table for lm %d name %s\n",i,lmset[i].name);
n=0;
for(j=0;j<dict_size(dict);j++){ /*try to be very careful again */
wp[j].wid=-1;
wp[j].prob=-1;
}
n = lm_ug_wordprob (lmset[i].lm, dict,MAX_NEG_INT32, wp);
E_INFO("Size of word table after unigram + words in class: %d.\n",n);
if (n < 1)
E_FATAL("%d active words in %s\n", n,lmset[i].name);
n = wid_wordprob2alt(dict,wp,n);
E_INFO("Size of word table after adding alternative prons: %d.\n",n);
if (cmd_ln_int32("-treeugprob") == 0) {
for (i = 0; i < n; i++)
wp[i].prob = -1; /* Flatten all initial probabilities */
}
for (j = 0; j < kb->n_lextree; j++) {
kb->ugtreeMulti[i*kb->n_lextree+j] = lextree_build (kbcore, wp, n, lc);
lextree_type (kb->ugtreeMulti[i*kb->n_lextree+j]) = 0;
E_INFO("Lextrees (%d) for lm %d name %s, %d nodes(ug)\n",
kb->n_lextree, i, lmset[i].name,lextree_n_node(kb->ugtreeMulti[i*kb->n_lextree+j]));
}
}
}else if (lm){
E_INFO("Creating Unigram Table\n");
n=0;
n = lm_ug_wordprob (lm, dict,MAX_NEG_INT32, wp);
E_INFO("Size of word table after unigram + words in class: %d\n",n);
if (n < 1)
E_FATAL("%d active words\n", n);
n = wid_wordprob2alt (dict, wp, n); /* Add alternative pronunciations */
/* Retain or remove unigram probs from lextree, depending on option */
if (cmd_ln_int32("-treeugprob") == 0) {
for (i = 0; i < n; i++)
wp[i].prob = -1; /* Flatten all initial probabilities */
}
/* Create the desired no. of unigram lextrees */
kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
for (i = 0; i < kb->n_lextree; i++) {
kb->ugtree[i] = lextree_build (kbcore, wp, n, lc);
lextree_type (kb->ugtree[i]) = 0;
}
E_INFO("Lextrees(%d), %d nodes(ug)\n",
kb->n_lextree, lextree_n_node(kb->ugtree[0]));
}
/* Create filler lextrees */
/* ARCHAN : only one filler tree is supposed to be build even for dynamic LMs */
n = 0;
for (i = dict_filler_start(dict); i <= dict_filler_end(dict); i++) {
if (dict_filler_word(dict, i)) {
wp[n].wid = i;
wp[n].prob = fillpen (kbcore->fillpen, i);
n++;
}
}
kb->fillertree = (lextree_t **)ckd_calloc(kb->n_lextree,sizeof(lextree_t*));
for (i = 0; i < kb->n_lextree; i++) {
kb->fillertree[i] = lextree_build (kbcore, wp, n, NULL);
lextree_type (kb->fillertree[i]) = -1;
}
ckd_free ((void *) wp);
ckd_free ((void *) lc);
bitvec_free (lc_active);
E_INFO("Lextrees(%d), %d nodes(filler)\n",
kb->n_lextree,
lextree_n_node(kb->fillertree[0]));
if (cmd_ln_int32("-lextreedump")) {
if(lmset){
E_FATAL("Currently, doesn't support -lextreedump for multiple-LMs\n");
}
for (i = 0; i < kb->n_lextree; i++) {
fprintf (stderr, "UGTREE %d\n", i);
lextree_dump (kb->ugtree[i], dict, stderr);
}
for (i = 0; i < kb->n_lextree; i++) {
fprintf (stderr, "FILLERTREE %d\n", i);
lextree_dump (kb->fillertree[i], dict, stderr);
}
fflush (stderr);
}
kb->ascr = ascr_init (mgau_n_mgau(kbcore_mgau(kbcore)),
kbcore->dict2pid->n_comstate);
kb->beam = beam_init (cmd_ln_float64("-subvqbeam"),
cmd_ln_float64("-beam"),
cmd_ln_float64("-pbeam"),
cmd_ln_float64("-wbeam"));
E_INFO("Beam= %d, PBeam= %d, WBeam= %d, SVQBeam= %d\n",
kb->beam->hmm, kb->beam->ptrans, kb->beam->word, kb->beam->subvq);
/*Sections of optimization related parameters*/
kb->ds_ratio=cmd_ln_int32("-ds");
E_INFO("Down Sampling Ratio = %d\n",kb->ds_ratio);
kb->rec_bstcid=-1;
kb->skip_count=0;
kb->cond_ds=cmd_ln_int32("-cond_ds");
E_INFO("Conditional Down Sampling Parameter = %d\n",kb->cond_ds);
if(kb->cond_ds>0&&kb->kbcore->gs==NULL) E_FATAL("Conditional Down Sampling require the use of Gaussian Selection map\n");
kb->gs4gs=cmd_ln_int32("-gs4gs");
E_INFO("GS map would be used for Gaussian Selection? = %d\n",kb->gs4gs);
kb->svq4svq=cmd_ln_int32("-svq4svq");
E_INFO("SVQ would be used as Gaussian Score ?= %d\n",kb->svq4svq);
kb->ci_pbeam=-1*logs3(cmd_ln_float32("-ci_pbeam"));
E_INFO("CI phone beam to prune the number of parent CI phones in CI-base GMM Selection = %d\n",kb->ci_pbeam);
if(kb->ci_pbeam>10000000){
E_INFO("Virtually no CI phone beam is applied now. (ci_pbeam>1000000)\n");
}
kb->wend_beam=-1*logs3(cmd_ln_float32("-wend_beam"));
E_INFO("Word-end pruning beam: %d\n",kb->wend_beam);
kb->pl_window=cmd_ln_int32("-pl_window");
E_INFO("Phoneme look-ahead window size = %d\n",kb->pl_window);
kb->pl_window_start=0;
kb->pl_beam=logs3(cmd_ln_float32("-pl_beam"));
E_INFO("Phoneme look-ahead beam = %d\n",kb->pl_beam);
for(cisencnt=0;cisencnt==mdef->cd2cisen[cisencnt];cisencnt++)
;
kb->cache_ci_senscr=(int32**)ckd_calloc_2d(kb->pl_window,cisencnt,sizeof(int32));
kb->cache_best_list=(int32*)ckd_calloc(kb->pl_window,sizeof(int32));
kb->phn_heur_list=(int32*)ckd_calloc(mdef_n_ciphone (mdef),sizeof(int32));
if ((kb->feat = feat_array_alloc(kbcore_fcb(kbcore),S3_MAX_FRAMES)) == NULL)
E_FATAL("feat_array_alloc() failed\n");
kb->vithist = vithist_init(kbcore, kb->beam->word, cmd_ln_int32("-bghist"));
ptmr_init (&(kb->tm_sen));
ptmr_init (&(kb->tm_srch));
ptmr_init (&(kb->tm_ovrhd));
kb->tot_fr = 0;
kb->tot_sen_eval = 0.0;
kb->tot_gau_eval = 0.0;
kb->tot_hmm_eval = 0.0;
kb->tot_wd_exit = 0.0;
kb->hmm_hist_binsize = cmd_ln_int32("-hmmhistbinsize");
if(lmset)
n = ((kb->ugtreeMulti[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
else
n = ((kb->ugtree[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
n /= kb->hmm_hist_binsize;
kb->hmm_hist_bins = n+1;
kb->hmm_hist = (int32 *) ckd_calloc (n+1, sizeof(int32)); /* Really no need for +1 */
/* Open hypseg file if specified */
str = cmd_ln_str("-hypseg");
kb->matchsegfp = NULL;
if (str) {
#ifdef SPEC_CPU_WINDOWS
if ((kb->matchsegfp = fopen(str, "wt")) == NULL)
#else
if ((kb->matchsegfp = fopen(str, "w")) == NULL)
#endif
E_ERROR("fopen(%s,w) failed; use FWDXCT: from std logfile\n", str);
}
str = cmd_ln_str("-hyp");
kb->matchfp = NULL;
if (str) {
#ifdef SPEC_CPU_WINDOWS
if ((kb->matchfp = fopen(str, "wt")) == NULL)
#else
if ((kb->matchfp = fopen(str, "w")) == NULL)
#endif
E_ERROR("fopen(%s,w) failed; use FWDXCT: from std logfile\n", str);
}
}
static int
init_state(const char *obsdmp,
const char *obsidx,
uint32 n_density,
uint32 n_stream,
uint32 *veclen,
uint32 blksize,
int reest,
const char *mixwfn,
const char *meanfn,
const char *varfn,
uint32 ts_off,
uint32 ts_cnt,
uint32 n_ts,
uint32 n_d_ts)
{
vector_t ***mean;
vector_t ***var = NULL;
vector_t ****fullvar = NULL;
float32 ***mixw = NULL;
uint32 n_frame;
uint32 ignore = 0;
codew_t *label;
uint32 n_corpus = 0;
float64 sqerr;
float64 tot_sqerr;
segdmp_type_t t;
uint32 i, j, ts, n;
int32 full_covar;
full_covar = cmd_ln_int32("-fullvar");
/* fully-continuous for now */
mean = gauden_alloc_param(ts_cnt, n_stream, n_density, veclen);
if (full_covar)
fullvar = gauden_alloc_param_full(ts_cnt, n_stream, n_density, veclen);
else
var = gauden_alloc_param(ts_cnt, n_stream, n_density, veclen);
if (mixwfn)
mixw = (float32 ***)ckd_calloc_3d(ts_cnt,
n_stream,
n_density,
sizeof(float32));
if (cmd_ln_str("-segidxfn")) {
E_INFO("Multi-class dump\n");
if (segdmp_open_read(cmd_ln_str_list("-segdmpdirs"),
cmd_ln_str("-segdmpfn"),
cmd_ln_str("-segidxfn"),
&n,
&t,
n_stream, veclen, blksize) != S3_SUCCESS) {
E_FATAL("Unable to open dumps\n");
}
if (n != n_d_ts) {
E_FATAL("Expected %u tied-states in dump, but apparently %u\n",
n_d_ts, n);
}
if (t != SEGDMP_TYPE_FEAT) {
E_FATAL("Expected feature dump, but instead saw %u\n", t);
}
multiclass = TRUE;
}
else {
E_INFO("1-class dump file\n");
multiclass = FALSE;
dmp_fp = s3open(cmd_ln_str("-segdmpfn"), "rb",
&dmp_swp);
if (dmp_fp == NULL) {
E_ERROR_SYSTEM("Unable to open dump file %s for reading\n",
cmd_ln_str("-segdmpfn"));
return S3_ERROR;
}
if (s3read(&n_frame, sizeof(uint32), 1, dmp_fp, dmp_swp, &ignore) != 1) {
E_ERROR_SYSTEM("Unable to open dump file %s for reading\n",
cmd_ln_str("-segdmpfn"));
return S3_ERROR;
}
data_offset = ftell(dmp_fp);
}
tot_sqerr = 0;
for (i = 0; i < ts_cnt; i++) {
ts = ts_off + i;
/* stride not accounted for yet */
if (o2d == NULL) {
if (multiclass)
n_frame = segdmp_n_seg(ts);
}
else {
for (j = 0, n_frame = 0; j < n_o2d[ts]; j++) {
n_frame += segdmp_n_seg(o2d[ts][j]);
}
}
E_INFO("Corpus %u: sz==%u frames%s\n",
ts, n_frame,
(n_frame > cmd_ln_int32("-vartiethr") ? "" : " tied var"));
if (n_frame == 0) {
continue;
}
E_INFO("Convergence ratios are abs(cur - prior) / abs(prior)\n");
/* Do some variety of k-means clustering */
ptmr_start(&km_timer);
sqerr = cluster(ts, n_stream, n_frame, veclen, blksize, mean[i], n_density, &label);
ptmr_stop(&km_timer);
if (sqerr < 0) {
E_ERROR("Unable to do k-means for state %u; skipping...\n", ts);
continue;
}
/* Given the k-means and assuming equal prior liklihoods
* compute the variances */
ptmr_start(&var_timer);
if (full_covar)
full_variances(ts, mean[i], fullvar[i], n_density, n_stream, veclen, blksize,
n_frame, label);
else
variances(ts, mean[i], var[i], n_density, n_stream, veclen, blksize, n_frame, label);
ptmr_stop(&var_timer);
if (mixwfn) {
/* initialize the mixing weights by counting # of occurrances
* of the top codeword over the corpus and normalizing */
init_mixw(mixw[i], mean[i], n_density, veclen, n_frame, n_stream, label);
ckd_free(label);
if (reest == TRUE && full_covar)
E_ERROR("EM re-estimation is not yet supported for full covariances\n");
else if (reest == TRUE) {
ptmr_start(&em_timer);
/* Do iterations of EM to estimate the mixture densities */
reest_sum(ts, mean[i], var[i], mixw[i], n_density, n_stream,
n_frame, veclen, blksize,
cmd_ln_int32("-niter"),
FALSE,
cmd_ln_int32("-vartiethr"));
ptmr_stop(&em_timer);
}
}
++n_corpus;
tot_sqerr += sqerr;
E_INFO("sqerr [%u] == %e\n", ts, sqerr);
}
if (n_corpus > 0) {
E_INFO("sqerr = %e tot %e rms\n", tot_sqerr, sqrt(tot_sqerr/n_corpus));
}
if (!multiclass)
s3close(dmp_fp);
if (meanfn) {
if (s3gau_write(meanfn,
(const vector_t ***)mean,
ts_cnt,
n_stream,
n_density,
veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_INFO("No mean file given; none written\n");
}
if (varfn) {
if (full_covar) {
if (s3gau_write_full(varfn,
(const vector_t ****)fullvar,
ts_cnt,
n_stream,
n_density,
veclen) != S3_SUCCESS)
return S3_ERROR;
}
else {
if (s3gau_write(varfn,
(const vector_t ***)var,
ts_cnt,
n_stream,
n_density,
veclen) != S3_SUCCESS)
return S3_ERROR;
}
}
else {
E_INFO("No variance file given; none written\n");
}
if (mixwfn) {
if (s3mixw_write(mixwfn,
mixw,
ts_cnt,
n_stream,
n_density) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_INFO("No mixing weight file given; none written\n");
}
return S3_SUCCESS;
}
int
main_initialize(int argc,
char *argv[],
lexicon_t **out_lex,
model_def_t **out_omdef,
model_def_t **out_dmdef)
{
model_def_t *dmdef = NULL;
model_def_t *omdef = NULL;
lexicon_t *lex = NULL;
const char *fn;
uint32 n_ts;
uint32 n_cb;
const char *ts2cbfn;
parse_cmd_ln(argc, argv);
timing_bind_name("km", timing_new());
timing_bind_name("var", timing_new());
timing_bind_name("em", timing_new());
timing_bind_name("all", timing_new());
if (cmd_ln_access("-feat") != NULL) {
feat_set(cmd_ln_str("-feat"));
feat_set_in_veclen(cmd_ln_int32("-ceplen"));
feat_set_subvecs(cmd_ln_str("-svspec"));
}
else {
E_FATAL("You need to set a feature extraction config using -feat\n");
}
if (cmd_ln_access("-ldafn") != NULL) {
if (feat_read_lda(cmd_ln_access("-ldafn"), cmd_ln_int32("-ldadim"))) {
E_FATAL("Failed to read LDA matrix\n");
}
}
if (cmd_ln_access("-omoddeffn")) {
E_INFO("Reading output model definitions: %s\n", cmd_ln_access("-omoddeffn"));
/* Read in the model definitions. Defines the set of
CI phones and context dependent phones. Defines the
transition matrix tying and state level tying. */
if (model_def_read(&omdef,
cmd_ln_access("-omoddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
if (cmd_ln_access("-dmoddeffn")) {
E_INFO("Reading dump model definitions: %s\n", cmd_ln_access("-dmoddeffn"));
if (model_def_read(&dmdef,
cmd_ln_access("-dmoddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
setup_d2o_map(dmdef, omdef);
}
else {
E_INFO("Assuming dump and output model definitions are identical\n");
}
ts2cbfn = cmd_ln_access("-ts2cbfn");
if (ts2cbfn) {
if (strcmp(SEMI_LABEL, ts2cbfn) == 0) {
omdef->cb = semi_ts2cb(omdef->n_tied_state);
n_ts = omdef->n_tied_state;
n_cb = 1;
}
else if (strcmp(CONT_LABEL, ts2cbfn) == 0) {
omdef->cb = cont_ts2cb(omdef->n_tied_state);
n_ts = omdef->n_tied_state;
n_cb = omdef->n_tied_state;
}
else if (strcmp(PTM_LABEL, ts2cbfn) == 0) {
omdef->cb = ptm_ts2cb(omdef);
n_ts = omdef->n_tied_state;
n_cb = omdef->acmod_set->n_ci;
}
else if (s3ts2cb_read(cmd_ln_access("-ts2cbfn"),
&omdef->cb,
&n_ts,
&n_cb) != S3_SUCCESS) {
return S3_ERROR;
}
if (omdef->n_tied_state != n_ts) {
E_FATAL("Model definition file n_tied_state = %u, but %u mappings in ts2cb\n",
omdef->n_tied_state, n_ts);
}
}
}
else {
E_INFO("No mdef files. Assuming 1-class init\n");
}
*out_omdef = omdef;
*out_dmdef = dmdef;
fn = cmd_ln_access("-dictfn");
if (fn) {
E_INFO("Reading main lexicon: %s\n", fn);
lex = lexicon_read(NULL,
fn,
omdef->acmod_set);
if (lex == NULL)
return S3_ERROR;
}
fn = cmd_ln_access("-fdictfn");
if (fn) {
E_INFO("Reading filler lexicon: %s\n", fn);
(void)lexicon_read(lex,
fn,
omdef->acmod_set);
}
*out_lex = lex;
stride = *(int32 *)cmd_ln_access("-stride");
return S3_SUCCESS;
}
/* the following function is used for MMIE training
lqin 2010-03 */
static int
mmi_normalize()
{
uint32 i;
uint32 n_mgau;
uint32 n_stream;
uint32 n_density;
vector_t ***in_mean = NULL;
vector_t ***in_var = NULL;
vector_t ***wt_mean = NULL;
vector_t ***wt_var = NULL;
const uint32 *veclen = NULL;
const char **accum_dir;
const char *in_mean_fn;
const char *out_mean_fn;
const char *in_var_fn;
const char *out_var_fn;
vector_t ***wt_num_mean = NULL;
vector_t ***wt_den_mean = NULL;
vector_t ***wt_num_var = NULL;
vector_t ***wt_den_var = NULL;
float32 ***num_dnom = NULL;
float32 ***den_dnom = NULL;
uint32 n_num_mgau;
uint32 n_den_mgau;
uint32 n_num_stream;
uint32 n_den_stream;
uint32 n_num_density;
uint32 n_den_density;
float32 constE;
uint32 n_temp_mgau;
uint32 n_temp_stream;
uint32 n_temp_density;
const uint32 *temp_veclen = NULL;
accum_dir = cmd_ln_str_list("-accumdir");
/* the following variables are used for mmie training */
out_mean_fn = cmd_ln_str("-meanfn");
out_var_fn = cmd_ln_str("-varfn");
in_mean_fn = cmd_ln_str("-inmeanfn");
in_var_fn = cmd_ln_str("-invarfn");
constE = cmd_ln_float32("-constE");
/* get rid of some unnecessary parameters */
if (cmd_ln_int32("-fullvar")) {
E_FATAL("Current MMIE training can not be done for full variance, set -fulllvar as no\n");
}
if (cmd_ln_int32("-tiedvar")) {
E_FATAL("Current MMIE training can not be done for tied variance, set -tiedvar as no\n");
}
if (cmd_ln_str("-mixwfn")) {
E_FATAL("Current MMIE training does not support mixture weight update, remove -mixwfn \n");
}
if (cmd_ln_str("-inmixwfn")) {
E_FATAL("Current MMIE training does not support mixture weight update, remove -inmixwfn \n");
}
if (cmd_ln_str("-tmatfn")) {
E_FATAL("Current MMIE training does not support transition matrix update, remove -tmatfn \n");
}
if (cmd_ln_str("-regmatfn")) {
E_FATAL("Using norm for computing regression matrix is obsolete, please use mllr_transform \n");
}
/* must be at least one accum dir */
if (accum_dir[0] == NULL) {
E_FATAL("No accumulated reestimation path is specified, use -accumdir \n");
}
/* at least update mean or variance parameters */
if (out_mean_fn == NULL && out_var_fn == NULL) {
E_FATAL("Neither -meanfn nor -varfn is specified, at least do mean or variance update \n");
}
else if (out_mean_fn == NULL) {
E_INFO("No -meanfn specified, will skip if any\n");
}
else if (out_var_fn == NULL) {
E_INFO("No -varfn specified, will skip if any\n");
}
/* read input mean */
if (in_mean_fn != NULL) {
E_INFO("read original density mean parameters from %s\n", in_mean_fn);
if (s3gau_read(in_mean_fn,
&in_mean,
&n_mgau,
&n_stream,
&n_density,
&veclen) != S3_SUCCESS) {
E_FATAL_SYSTEM("Couldn't read %s", in_mean_fn);
}
ckd_free((void *)veclen);
veclen = NULL;
}
/* read input variance */
if (in_var_fn != NULL) {
E_INFO("read original density variance parameters from %s\n", in_var_fn);
if (s3gau_read(in_var_fn,
&in_var,
&n_mgau,
&n_stream,
&n_density,
&veclen) != S3_SUCCESS) {
E_FATAL_SYSTEM("Couldn't read %s", in_var_fn);
}
ckd_free((void *)veclen);
veclen = NULL;
}
/* read accumulated numerator and denominator counts */
for (i = 0; accum_dir[i]; i++) {
E_INFO("Reading and accumulating counts from %s\n", accum_dir[i]);
rdacc_mmie_den(accum_dir[i],
"numlat",
&wt_num_mean,
&wt_num_var,
&num_dnom,
&n_num_mgau,
&n_num_stream,
&n_num_density,
&veclen);
rdacc_mmie_den(accum_dir[i],
"denlat",
&wt_den_mean,
&wt_den_var,
&den_dnom,
&n_den_mgau,
&n_den_stream,
&n_den_density,
&veclen);
if (n_num_mgau != n_den_mgau)
E_FATAL("number of gaussians inconsistent between num and den lattice\n");
else if (n_num_mgau != n_mgau)
E_FATAL("number of gaussians inconsistent between imput model and accumulator (%u != %u)\n", n_mgau, n_num_mgau);
if (n_num_stream != n_den_stream)
E_FATAL("number of gaussian streams inconsistent between num and den lattice\n");
else if (n_num_stream != n_stream)
E_FATAL("number of gaussian streams inconsistent between imput model and accumulator (%u != %u)\n", n_stream, n_num_stream);
if (n_num_density != n_den_density)
E_FATAL("number of gaussian densities inconsistent between num and den lattice\n");
else if (n_num_density != n_density)
E_FATAL("number of gaussian densities inconsistent between imput model and accumulator (%u != %u)\n", n_density, n_num_density);
}
/* initialize update parameters as the input parameters */
if (out_mean_fn) {
if (s3gau_read(in_mean_fn,
&wt_mean,
&n_temp_mgau,
&n_temp_stream,
&n_temp_density,
&temp_veclen) != S3_SUCCESS) {
E_FATAL_SYSTEM("Couldn't read %s", in_mean_fn);
}
ckd_free((void *)temp_veclen);
temp_veclen = NULL;
}
if (out_var_fn) {
if (s3gau_read(in_var_fn,
&wt_var,
&n_temp_mgau,
&n_temp_stream,
&n_temp_density,
&temp_veclen) != S3_SUCCESS) {
E_FATAL_SYSTEM("Couldn't read %s", in_var_fn);
}
ckd_free((void *)temp_veclen);
temp_veclen = NULL;
}
/* update mean parameters */
if (wt_mean) {
if (out_mean_fn) {
E_INFO("Normalizing mean for n_mgau= %u, n_stream= %u, n_density= %u\n",
n_mgau, n_stream, n_density);
gauden_norm_wt_mmie_mean(in_mean, wt_mean, wt_num_mean, wt_den_mean,
in_var, wt_num_var, wt_den_var, num_dnom, den_dnom,
n_mgau, n_stream, n_density, veclen, constE);
}
else {
E_INFO("Ignoring means since -meanfn not specified\n");
}
}
else {
E_INFO("No means to normalize\n");
}
/* update variance parameters */
if (wt_var) {
if (out_var_fn) {
E_INFO("Normalizing variance for n_mgau= %u, n_stream= %u, n_density= %u\n",
n_mgau, n_stream, n_density);
gauden_norm_wt_mmie_var(in_var, wt_var, wt_num_var, wt_den_var, num_dnom, den_dnom,
in_mean, wt_mean, wt_num_mean, wt_den_mean,
n_mgau, n_stream, n_density, veclen, constE);
}
else {
E_INFO("Ignoring variances since -varfn not specified\n");
}
}
else {
E_INFO("No variances to normalize\n");
}
/* write the updated mean parameters to files */
if (out_mean_fn) {
if (wt_mean) {
if (s3gau_write(out_mean_fn,
(const vector_t ***)wt_mean,
n_mgau,
n_stream,
n_density,
veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_WARN("NO reestimated means seen, but -meanfn specified\n");
}
}
else {
if (wt_mean) {
E_INFO("Reestimated means seen, but -meanfn NOT specified\n");
}
}
/* write the updated variance parameters to files */
if (out_var_fn) {
if (wt_var) {
if (s3gau_write(out_var_fn,
(const vector_t ***)wt_var,
n_mgau,
n_stream,
n_density,
veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_WARN("NO reestimated variances seen, but -varfn specified\n");
}
}
else {
if (wt_var) {
E_INFO("Reestimated variances seen, but -varfn NOT specified\n");
}
}
if (veclen)
ckd_free((void *)veclen);
if (temp_veclen)
ckd_free((void *)temp_veclen);
return S3_SUCCESS;
}
int32 main (int32 argc, char *argv[])
{
FILE *fpout;
mgau_model_t *mgau;
int32 **subvec;
int32 max_datarows, datarows, datacols, svqrows, svqcols;
float32 **data, **vqmean;
int32 *datamap, *vqmap;
float64 sqerr;
int32 stdev;
int32 i, j, v, m, c;
cmd_ln_parse (arg, argc, argv);
logs3_init ((float64) 1.0003);
/* Load means/vars but DO NOT precompute variance inverses or determinants */
mgau = mgau_init (cmd_ln_str("-mean"), cmd_ln_str("-var"), 0.0 /* no varfloor */,
cmd_ln_str("-mixw"), cmd_ln_float64 ("-mixwfloor"),
FALSE);
mgau_var_nzvec_floor (mgau, cmd_ln_float64 ("-varfloor"));
/* Parse subvector spec argument; subvec is null terminated; subvec[x] is -1 terminated */
subvec = parse_subvecs (cmd_ln_str("-svspec"));
if (cmd_ln_str ("-subvq"))
fpout = myfopen (cmd_ln_str ("-subvq"), "w");
else
fpout = stdout;
/* Echo command line to output file */
for (i = 0; i < argc-1; i++)
fprintf (fpout, "# %s \\\n", argv[i]);
fprintf (fpout, "# %s\n#\n", argv[argc-1]);
/* Print input and output configurations to output file */
for (v = 0; subvec[v]; v++); /* No. of subvectors */
svqrows = cmd_ln_int32 ("-svqrows");
fprintf (fpout, "VQParam %d %d -> %d %d\n",
mgau_n_mgau(mgau), mgau_max_comp(mgau), v, svqrows);
for (v = 0; subvec[v]; v++) {
for (i = 0; subvec[v][i] >= 0; i++);
fprintf (fpout, "Subvector %d length %d ", v, i);
for (i = 0; subvec[v][i] >= 0; i++)
fprintf (fpout, " %2d", subvec[v][i]);
fprintf (fpout, "\n");
}
fflush (fpout);
/*
* datamap[] for identifying non-0 input vectors that take part in the clustering process:
* datamap[m*max_mean + c] = row index of data[][] containing the copy.
* vqmap[] for mapping vq input data to vq output.
*/
max_datarows = mgau_n_mgau(mgau) * mgau_max_comp(mgau);
datamap = (int32 *) ckd_calloc (max_datarows, sizeof(int32));
vqmap = (int32 *) ckd_calloc (max_datarows, sizeof(int32));
stdev = cmd_ln_int32 ("-stdev");
/* Copy and cluster each subvector */
for (v = 0; subvec[v]; v++) {
E_INFO("Clustering subvector %d\n", v);
for (datacols = 0; subvec[v][datacols] >= 0; datacols++); /* Input subvec length */
svqcols = datacols * 2; /* subvec length after concatenating mean + var */
/* Allocate input/output data areas */
data = (float32 **) ckd_calloc_2d (max_datarows, svqcols, sizeof(float32));
vqmean = (float32 **) ckd_calloc_2d (svqrows, svqcols, sizeof(float32));
/* Make a copy of the subvectors from the input data, and initialize maps */
for (i = 0; i < max_datarows; i++)
datamap[i] = -1;
datarows = 0;
for (m = 0; m < mgau_n_mgau(mgau); m++) { /* For each mixture m */
for (c = 0; c < mgau_n_comp(mgau, m); c++) { /* For each component c in m */
if (vector_is_zero (mgau_var(mgau, m, c), mgau_veclen(mgau)))
continue;
for (i = 0; i < datacols; i++) { /* Copy specified dimensions, mean+var */
data[datarows][i*2] = mgau->mgau[m].mean[c][subvec[v][i]];
data[datarows][i*2+1] = (! stdev) ?
mgau->mgau[m].var[c][subvec[v][i]] :
sqrt(mgau->mgau[m].var[c][subvec[v][i]]);
}
datamap[m * mgau_max_comp(mgau) + c] = datarows++;
}
}
E_INFO("Sanity check: input data[0]:\n");
vector_print (stderr, data[0], svqcols);
for (i = 0; i < max_datarows; i++)
vqmap[i] = -1;
#if 0
{
int32 **in;
printf ("Input data: %d x %d\n", datarows, svqcols);
in = (int32 **)data;
for (i = 0; i < datarows; i++) {
printf ("%8d:", i);
for (j = 0; j < svqcols; j++)
printf (" %08x", in[i][j]);
printf ("\n");
}
for (i = 0; i < datarows; i++) {
printf ("%15d:", i);
for (j = 0; j < svqcols; j++)
printf (" %15.7e", data[i][j]);
printf ("\n");
}
fflush (stdout);
}
#endif
/* VQ the subvector copy built above */
sqerr = vector_vqgen (data, datarows, svqcols, svqrows,
cmd_ln_float64("-eps"), cmd_ln_int32("-iter"),
vqmean, vqmap);
/* Output VQ */
fprintf (fpout, "Codebook %d Sqerr %e\n", v, sqerr);
for (i = 0; i < svqrows; i++) {
if (stdev) {
/* Convert clustered stdev back to var */
for (j = 1; j < svqcols; j += 2)
vqmean[i][j] *= vqmean[i][j];
}
vector_print (fpout, vqmean[i], svqcols);
}
fprintf (fpout, "Map %d\n", v);
for (i = 0; i < max_datarows; i += mgau_max_comp(mgau)) {
for (j = 0; j < mgau_max_comp(mgau); j++) {
if (datamap[i+j] < 0)
fprintf (fpout, " -1");
else
fprintf (fpout, " %d", vqmap[datamap[i+j]]);
}
fprintf (fpout, "\n");
}
fflush (fpout);
/* Cleanup */
ckd_free_2d ((void **) data);
ckd_free_2d ((void **) vqmean);
}
fprintf (fpout, "End\n");
fclose (fpout);
exit(0);
}
static int
cp_parm()
{
FILE *fp;
uint32 i, o;
uint32 max=0;
/* Open the file first to see whether command-line parameters
match
*/
if(cmd_ln_str("-cpopsfn")==NULL) {
E_INFO("Please specify -cpopsfn\n");
return S3_ERROR;
}
fp = fopen(cmd_ln_str("-cpopsfn"), "r");
if (fp == NULL) {
E_INFO("Unable to open cpops file\n");
return S3_ERROR;
}
while (fscanf(fp, "%u %u", &o, &i) == 2) {
if(o+1>max) {
max=o+1;
}
}
if (omixw) {
if(max != n_mixw_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -nmixout (%d)\n",max, n_mixw_o);
return S3_ERROR;
}
}
if (ogau) {
if(max != n_cb_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -ncbout (%d)\n",max, n_cb_o);
return S3_ERROR;
}
}
if (ogau_full) {
if(max != n_cb_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -ncbout (%d)\n",max, n_cb_o);
return S3_ERROR;
}
}
if (otmat) {
if(max != n_tmat_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -ntmatout (%d)\n",max, n_tmat_o);
return S3_ERROR;
}
}
fclose(fp);
fp = fopen(cmd_ln_str("-cpopsfn"), "r");
while (fscanf(fp, "%u %u", &o, &i) == 2) {
if (omixw) {
cp_mixw(o, i);
}
if (ogau) {
cp_gau(o, i);
}
if (ogau_full) {
cp_gau_full(o, i);
}
if (otmat) {
cp_tmat(o, i);
}
}
fclose(fp);
return S3_SUCCESS;
}
void kb_setlm(char* lmname,kb_t* kb)
{
lmset_t* lms;
kbcore_t* kbc=NULL;
int i;
int j;
int n;
/* s3wid_t dictid;*/
kbc=kb->kbcore;
lms=kbc->lmset;
E_INFO("Inside kb_setlm\n");
kbc->lm=NULL;
for(j=0;j<kb->n_lextree;j++){
kb->ugtree[j]=NULL;
}
E_INFO("Inside kb_setlm\n");
if(lms!=NULL || cmd_ln_str("-lmctlfn")){
for(i=0;i<kbc->n_lm;i++){
if(!strcmp(lmname,lms[i].name)){
/* Point the current lm to a particular lm */
kbc->lm=lms[i].lm;
for(j=0;j<kb->n_lextree;j++){
kb->ugtree[j]=kb->ugtreeMulti[i*kb->n_lextree+j];
}
break;
}
}
if(kbc->lm==NULL){
E_ERROR("LM name %s cannot be found! Fall back to use base LM model\n",lmname);
kbc->lm=lms[0].lm;
for(j=0;j<kb->n_lextree;j++){
kb->ugtree[j]=kb->ugtreeMulti[j];
}
}
}
E_INFO("Current LM name %s.\n",lms[i].name);
/* if((kb->vithist->lms2vh_root=
(vh_lms2vh_t**)ckd_realloc(kb->vithist->lms2vh_root,
lm_n_ug(kbc->lm)*sizeof(vh_lms2vh_t *)
))==NULL)
{
E_FATAL("failed to allocate memory for vithist\n");
}*/
n = ((kb->ugtree[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
n /= kb->hmm_hist_binsize;
kb->hmm_hist_bins = n+1;
kb->hmm_hist = (int32 *) ckd_realloc (kb->hmm_hist,(n+1)*sizeof(int32)); /* Really no need for +1 */
E_INFO("Current LM name %s\n",lms[i].name);
E_INFO("LM ug size %d\n",kbc->lm->n_ug);
E_INFO("LM bg size %d\n",kbc->lm->n_bg);
E_INFO("LM tg size %d\n",kbc->lm->n_tg);
E_INFO("HMM history bin size %d\n", n+1);
for(j=0;j<kb->n_lextree;j++){
E_INFO("Lextrees(%d), %d nodes(ug)\n",
kb->n_lextree, lextree_n_node(kb->ugtree[j]));
}
/* for (n = 0; n < dict_size(kbcore_dict(kbc)); n++){
E_INFO("Index %d, map %d word %s\n",n,kbc->lm->dict2lmwid[n],dict_wordstr(kbcore_dict(kbc),n));
}*/
}
static int
normalize()
{
char file_name[MAXPATHLEN+1];
float32 ***mixw_acc = NULL;
float32 ***in_mixw = NULL;
float64 s;
uint32 n_mixw;
uint32 n_stream;
uint32 n_mllr_class;
uint32 n_density;
float32 ***tmat_acc = NULL;
uint32 n_tmat;
uint32 n_state_pm;
uint32 i, j, k;
vector_t ***in_mean = NULL;
vector_t ***wt_mean = NULL;
vector_t ***in_var = NULL;
vector_t ***wt_var = NULL;
vector_t ****in_fullvar = NULL;
vector_t ****wt_fullvar = NULL;
int32 pass2var = FALSE;
int32 var_is_full = FALSE;
float32 ***dnom = NULL;
uint32 n_mgau;
uint32 n_gau_stream;
uint32 n_gau_density;
const uint32 *veclen = NULL;
const char **accum_dir;
const char *oaccum_dir;
const char *in_mixw_fn;
const char *out_mixw_fn;
const char *out_tmat_fn;
const char *in_mean_fn;
const char *out_mean_fn;
const char *in_var_fn;
const char *out_var_fn;
const char *out_dcount_fn;
int err;
uint32 mllr_mult;
uint32 mllr_add;
float32 *****regl = NULL;
float32 ****regr = NULL;
uint32 no_retries=0;
accum_dir = cmd_ln_str_list("-accumdir");
oaccum_dir = cmd_ln_str("-oaccumdir");
out_mixw_fn = cmd_ln_str("-mixwfn");
out_tmat_fn = cmd_ln_str("-tmatfn");
out_mean_fn = cmd_ln_str("-meanfn");
out_var_fn = cmd_ln_str("-varfn");
in_mixw_fn = cmd_ln_str("-inmixwfn");
in_mean_fn = cmd_ln_str("-inmeanfn");
in_var_fn = cmd_ln_str("-invarfn");
out_dcount_fn = cmd_ln_str("-dcountfn");
var_is_full = cmd_ln_int32("-fullvar");
/* must be at least one accum dir */
assert(accum_dir[0] != NULL);
if (out_mixw_fn == NULL) {
E_INFO("No -mixwfn specified, will skip if any\n");
}
if (out_tmat_fn == NULL) {
E_INFO("No -tmatfn specified, will skip if any\n");
}
if (out_mean_fn == NULL) {
E_INFO("No -meanfn specified, will skip if any\n");
}
if (out_var_fn == NULL) {
E_INFO("No -varfn specified, will skip if any\n");
}
if (in_mixw_fn != NULL) {
E_INFO("Selecting unseen mixing weight parameters from %s\n",
in_mixw_fn);
}
if (in_mean_fn != NULL) {
E_INFO("Selecting unseen density mean parameters from %s\n",
in_mean_fn);
if (s3gau_read(in_mean_fn,
&in_mean,
&n_mgau,
&n_gau_stream,
&n_gau_density,
&veclen) != S3_SUCCESS) {
E_FATAL_SYSTEM("Couldn't read %s", in_mean_fn);
}
ckd_free((void *)veclen);
veclen = NULL;
}
if (in_var_fn != NULL) {
E_INFO("Selecting unseen density variance parameters from %s\n",
in_var_fn);
if (var_is_full) {
if (s3gau_read_full(in_var_fn,
&in_fullvar,
&n_mgau,
&n_gau_stream,
&n_gau_density,
&veclen) != S3_SUCCESS) {
E_FATAL_SYSTEM("Couldn't read %s", in_var_fn);
}
}
else {
if (s3gau_read(in_var_fn,
&in_var,
&n_mgau,
&n_gau_stream,
&n_gau_density,
&veclen) != S3_SUCCESS) {
E_FATAL_SYSTEM("Couldn't read %s", in_var_fn);
}
}
ckd_free((void *)veclen);
veclen = NULL;
}
n_stream = 0;
for (i = 0; accum_dir[i]; i++) {
E_INFO("Reading and accumulating counts from %s\n", accum_dir[i]);
if (out_mixw_fn) {
rdacc_mixw(accum_dir[i],
&mixw_acc, &n_mixw, &n_stream, &n_density);
}
if (out_tmat_fn) {
rdacc_tmat(accum_dir[i],
&tmat_acc, &n_tmat, &n_state_pm);
}
if (out_mean_fn || out_var_fn) {
if (var_is_full)
rdacc_den_full(accum_dir[i],
&wt_mean,
&wt_fullvar,
&pass2var,
&dnom,
&n_mgau,
&n_gau_stream,
&n_gau_density,
&veclen);
else
rdacc_den(accum_dir[i],
&wt_mean,
&wt_var,
&pass2var,
&dnom,
&n_mgau,
&n_gau_stream,
&n_gau_density,
&veclen);
if (out_mixw_fn) {
if (n_stream != n_gau_stream) {
E_ERROR("mixw inconsistent w/ densities WRT # "
"streams (%u != %u)\n",
n_stream, n_gau_stream);
}
if (n_density != n_gau_density) {
E_ERROR("mixw inconsistent w/ densities WRT # "
"den/mix (%u != %u)\n",
n_density, n_gau_density);
}
}
else {
n_stream = n_gau_stream;
n_density = n_gau_density;
}
}
}
if (oaccum_dir && mixw_acc) {
/* write the total mixing weight reest. accumulators */
err = 0;
sprintf(file_name, "%s/mixw_counts", oaccum_dir);
if (in_mixw_fn) {
if (s3mixw_read(in_mixw_fn,
&in_mixw,
&i,
&j,
&k) != S3_SUCCESS) {
E_FATAL_SYSTEM("Unable to read %s", in_mixw_fn);
}
if (i != n_mixw) {
E_FATAL("# mixw in input mixw file != # mixw in output mixw file\n");
}
if (j != n_stream) {
E_FATAL("# stream in input mixw file != # stream in output mixw file\n");
}
if (k != n_density) {
E_FATAL("# density in input mixw file != # density in output mixw file\n");
}
for (i = 0; i < n_mixw; i++) {
for (j = 0; j < n_stream; j++) {
for (k = 0, s = 0; k < n_density; k++) {
s += mixw_acc[i][j][k];
}
if ((s == 0) && in_mixw) {
for (k = 0, s = 0; k < n_density; k++) {
mixw_acc[i][j][k] = in_mixw[i][j][k];
}
E_INFO("set mixw %u stream %u to input mixw value\n", i, j);
}
}
}
}
do {
/* Write out the accumulated reestimation sums */
if (s3mixw_write(file_name,
mixw_acc,
n_mixw,
n_stream,
n_density) != S3_SUCCESS) {
if (err == 0) {
E_ERROR("Unable to write %s; Retrying...\n", file_name);
}
++err;
sleep(3);
no_retries++;
if(no_retries>10){
E_FATAL("Failed to get the files after 10 retries(about 30 seconds).\n ");
}
}
} while (err > 1);
}
if (pass2var)
E_INFO("-2passvar yes\n");
if (oaccum_dir && (wt_mean || wt_var || wt_fullvar)) {
/* write the total mixing Gau. den reest. accumulators */
err = 0;
sprintf(file_name, "%s/gauden_counts", oaccum_dir);
do {
int32 rv;
if (var_is_full)
rv = s3gaucnt_write_full(file_name,
wt_mean,
wt_fullvar,
pass2var,
dnom,
n_mgau,
n_gau_stream,
n_gau_density,
veclen);
else
rv = s3gaucnt_write(file_name,
wt_mean,
wt_var,
pass2var,
dnom,
n_mgau,
n_gau_stream,
n_gau_density,
veclen);
if (rv != S3_SUCCESS) {
if (err == 0) {
E_ERROR("Unable to write %s; Retrying...\n", file_name);
}
++err;
sleep(3);
no_retries++;
if(no_retries>10){
E_FATAL("Failed to get the files after 10 retries(about 5 minutes).\n ");
}
}
} while (err > 1);
}
if (oaccum_dir && tmat_acc) {
/* write the total transition matrix reest. accumulators */
err = 0;
sprintf(file_name, "%s/tmat_counts", oaccum_dir);
do {
if (s3tmat_write(file_name,
tmat_acc,
n_tmat,
n_state_pm) != S3_SUCCESS) {
if (err == 0) {
E_ERROR("Unable to write %s; Retrying...\n", file_name);
}
++err;
sleep(3);
no_retries++;
if(no_retries>10){
E_FATAL("Failed to get the files after 10 retries(about 5 minutes).\n ");
}
}
} while (err > 1);
}
if (oaccum_dir && regr && regl) {
/* write the total MLLR regression matrix accumulators */
err = 0;
sprintf(file_name, "%s/regmat_counts", oaccum_dir);
do {
if (s3regmatcnt_write(file_name,
regr,
regl,
n_mllr_class,
n_stream,
veclen,
mllr_mult,
mllr_add) != S3_SUCCESS) {
if (err == 0) {
E_ERROR("Unable to write %s; Retrying...\n", file_name);
}
++err;
sleep(3);
no_retries++;
if(no_retries>10){
E_FATAL("Failed to get the files after 10 retries(about 5 minutes).\n ");
}
}
} while (err > 1);
}
if (wt_mean || wt_var || wt_fullvar) {
if (out_mean_fn) {
E_INFO("Normalizing mean for n_mgau= %u, n_stream= %u, n_density= %u\n",
n_mgau, n_stream, n_density);
gauden_norm_wt_mean(in_mean, wt_mean, dnom,
n_mgau, n_stream, n_density, veclen);
}
else {
if (wt_mean) {
E_INFO("Ignoring means since -meanfn not specified\n");
}
}
if (out_var_fn) {
if (var_is_full) {
if (wt_fullvar) {
E_INFO("Normalizing fullvar\n");
gauden_norm_wt_fullvar(in_fullvar, wt_fullvar, pass2var, dnom,
wt_mean, /* wt_mean now just mean */
n_mgau, n_stream, n_density, veclen,
cmd_ln_boolean("-tiedvar"));
}
}
else {
if (wt_var) {
E_INFO("Normalizing var\n");
gauden_norm_wt_var(in_var, wt_var, pass2var, dnom,
wt_mean, /* wt_mean now just mean */
n_mgau, n_stream, n_density, veclen,
cmd_ln_boolean("-tiedvar"));
}
}
}
else {
if (wt_var || wt_fullvar) {
E_INFO("Ignoring variances since -varfn not specified\n");
}
}
}
else {
E_INFO("No means or variances to normalize\n");
}
/*
* Write the parameters to files
*/
if (out_mixw_fn) {
if (mixw_acc) {
if (s3mixw_write(out_mixw_fn,
mixw_acc,
n_mixw,
n_stream,
n_density) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_WARN("NO mixing weight accumulators seen, but -mixwfn specified.\n");
}
}
else {
if (mixw_acc) {
E_INFO("Mixing weight accumulators seen, but -mixwfn NOT specified.\n");
}
}
if (out_tmat_fn) {
if (tmat_acc) {
if (s3tmat_write(out_tmat_fn,
tmat_acc,
n_tmat,
n_state_pm) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_WARN("NO transition matrix accumulators seen, but -tmatfn specified.\n");
}
}
else {
if (tmat_acc)
E_INFO("Transition matrix accumulators seen, but -tmatfn NOT specified\n");
}
if (out_mean_fn) {
if (wt_mean) {
if (s3gau_write(out_mean_fn,
(const vector_t ***)wt_mean,
n_mgau,
n_stream,
n_density,
veclen) != S3_SUCCESS)
return S3_ERROR;
if (out_dcount_fn) {
if (s3gaudnom_write(out_dcount_fn,
dnom,
n_mgau,
n_stream,
n_density) != S3_SUCCESS)
return S3_ERROR;
}
}
else
E_WARN("NO reestimated means seen, but -meanfn specified\n");
}
else {
if (wt_mean) {
E_INFO("Reestimated means seen, but -meanfn NOT specified\n");
}
}
if (out_var_fn) {
if (var_is_full) {
if (wt_fullvar) {
if (s3gau_write_full(out_var_fn,
(const vector_t ****)wt_fullvar,
n_mgau,
n_stream,
n_density,
veclen) != S3_SUCCESS)
return S3_ERROR;
}
else
E_WARN("NO reestimated variances seen, but -varfn specified\n");
}
else {
if (wt_var) {
if (s3gau_write(out_var_fn,
(const vector_t ***)wt_var,
n_mgau,
n_stream,
n_density,
veclen) != S3_SUCCESS)
return S3_ERROR;
}
else
E_WARN("NO reestimated variances seen, but -varfn specified\n");
}
}
else {
if (wt_var) {
E_INFO("Reestimated variances seen, but -varfn NOT specified\n");
}
}
if (veclen)
ckd_free((void *)veclen);
return S3_SUCCESS;
}
int32
viterbi_update(float64 *log_forw_prob,
vector_t **feature,
uint32 n_obs,
state_t *state_seq,
uint32 n_state,
model_inventory_t *inv,
float64 a_beam,
float32 spthresh,
s3phseg_t *phseg,
int32 mixw_reest,
int32 tmat_reest,
int32 mean_reest,
int32 var_reest,
int32 pass2var,
int32 var_is_full,
FILE *pdumpfh,
feat_t *fcb)
{
float64 *scale = NULL;
float64 **dscale = NULL;
float64 **active_alpha;
uint32 **active_astate;
uint32 **bp;
uint32 *n_active_astate;
gauden_t *g; /* Gaussian density parameters and
reestimation sums */
float32 ***mixw; /* all mixing weights */
float64 ***now_den = NULL; /* Short for den[t] */
uint32 ***now_den_idx = NULL;/* Short for den_idx[t] */
uint32 *active_cb;
uint32 n_active_cb;
float32 **tacc; /* Transition matrix reestimation sum accumulators
for the utterance. */
float32 ***wacc; /* mixing weight reestimation sum accumulators
for the utterance. */
float32 ***denacc = NULL; /* mean/var reestimation accumulators for time t */
size_t denacc_size; /* Total size of data references in denacc. Allows
for quick clears between time frames */
uint32 n_lcl_cb;
uint32 *cb_inv;
uint32 i, j, q;
int32 t;
uint32 n_feat;
uint32 n_density;
uint32 n_top;
int ret;
timing_t *fwd_timer = NULL;
timing_t *rstu_timer = NULL;
timing_t *gau_timer = NULL;
timing_t *rsts_timer = NULL;
timing_t *rstf_timer = NULL;
float64 log_fp; /* accumulator for the log of the probability
* of observing the input given the model */
uint32 max_n_next = 0;
uint32 n_cb;
static float64 *p_op = NULL;
static float64 *p_ci_op = NULL;
static float64 **d_term = NULL;
static float64 **d_term_ci = NULL;
/* caller must ensure that there is some non-zero amount
of work to be done here */
assert(n_obs > 0);
assert(n_state > 0);
/* Get the forward estimation CPU timer */
fwd_timer = timing_get("fwd");
/* Get the per utterance reestimation CPU timer */
rstu_timer = timing_get("rstu");
/* Get the Gaussian density evaluation CPU timer */
gau_timer = timing_get("gau");
/* Get the per state reestimation CPU timer */
rsts_timer = timing_get("rsts");
/* Get the per frame reestimation CPU timer */
rstf_timer = timing_get("rstf");
g = inv->gauden;
n_feat = gauden_n_feat(g);
n_density = gauden_n_density(g);
n_top = gauden_n_top(g);
n_cb = gauden_n_mgau(g);
if (p_op == NULL) {
p_op = ckd_calloc(n_feat, sizeof(float64));
p_ci_op = ckd_calloc(n_feat, sizeof(float64));
}
if (d_term == NULL) {
d_term = (float64 **)ckd_calloc_2d(n_feat, n_top, sizeof(float64));
d_term_ci = (float64 **)ckd_calloc_2d(n_feat, n_top, sizeof(float64));
}
scale = (float64 *)ckd_calloc(n_obs, sizeof(float64));
dscale = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
n_active_astate = (uint32 *)ckd_calloc(n_obs, sizeof(uint32));
active_alpha = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
active_astate = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
active_cb = ckd_calloc(2*n_state, sizeof(uint32));
bp = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
/* Run forward algorithm, which has embedded Viterbi. */
if (fwd_timer)
timing_start(fwd_timer);
ret = forward(active_alpha, active_astate, n_active_astate, bp,
scale, dscale,
feature, n_obs, state_seq, n_state,
inv, a_beam, phseg, 0);
/* Dump a phoneme segmentation if requested */
if (cmd_ln_str("-outphsegdir")) {
const char *phsegdir;
char *segfn, *uttid;
phsegdir = cmd_ln_str("-outphsegdir");
uttid = (cmd_ln_int32("-outputfullpath")
? corpus_utt_full_name() : corpus_utt());
segfn = ckd_calloc(strlen(phsegdir) + 1
+ strlen(uttid)
+ strlen(".phseg") + 1, 1);
strcpy(segfn, phsegdir);
strcat(segfn, "/");
strcat(segfn, uttid);
strcat(segfn, ".phseg");
write_phseg(segfn, inv, state_seq, active_astate, n_active_astate,
n_state, n_obs, active_alpha, scale, bp);
ckd_free(segfn);
}
if (fwd_timer)
timing_stop(fwd_timer);
if (ret != S3_SUCCESS) {
/* Some problem with the utterance, release per utterance storage and
* forget about adding the utterance accumulators to the global accumulators */
goto all_done;
}
mixw = inv->mixw;
if (mixw_reest) {
/* Need to reallocate mixing accumulators for utt */
if (inv->l_mixw_acc) {
ckd_free_3d((void ***)inv->l_mixw_acc);
inv->l_mixw_acc = NULL;
}
inv->l_mixw_acc = (float32 ***)ckd_calloc_3d(inv->n_mixw_inverse,
n_feat,
n_density,
sizeof(float32));
}
wacc = inv->l_mixw_acc;
n_lcl_cb = inv->n_cb_inverse;
cb_inv = inv->cb_inverse;
/* Allocate local accumulators for mean, variance reestimation
sums if necessary */
gauden_alloc_l_acc(g, n_lcl_cb,
mean_reest, var_reest,
var_is_full);
if (tmat_reest) {
if (inv->l_tmat_acc) {
ckd_free_2d((void **)inv->l_tmat_acc);
inv->l_tmat_acc = NULL;
}
for (i = 0; i < n_state; i++) {
if (state_seq[i].n_next > max_n_next)
max_n_next = state_seq[i].n_next;
}
inv->l_tmat_acc = (float32 **)ckd_calloc_2d(n_state,
max_n_next,
sizeof(float32));
}
/* transition matrix reestimation sum accumulators
for the utterance */
tacc = inv->l_tmat_acc;
n_active_cb = 0;
now_den = (float64 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_top,
sizeof(float64));
now_den_idx = (uint32 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_top,
sizeof(uint32));
if (mean_reest || var_reest) {
/* allocate space for the per frame density counts */
denacc = (float32 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_density,
sizeof(float32));
/* # of bytes required to store all weighted vectors */
denacc_size = n_lcl_cb * n_feat * n_density * sizeof(float32);
}
else {
denacc = NULL;
denacc_size = 0;
}
/* Okay now run through the backtrace and accumulate counts. */
/* Find the non-emitting ending state */
for (q = 0; q < n_active_astate[n_obs-1]; ++q) {
if (active_astate[n_obs-1][q] == n_state-1)
break;
}
if (q == n_active_astate[n_obs-1]) {
E_ERROR("Failed to align audio to trancript: final state of the search is not reached\n");
ret = S3_ERROR;
goto all_done;
}
for (t = n_obs-1; t >= 0; --t) {
uint32 l_cb;
uint32 l_ci_cb;
float64 op, p_reest_term;
uint32 prev;
j = active_astate[t][q];
/* Follow any non-emitting states at time t first. */
while (state_seq[j].mixw == TYING_NON_EMITTING) {
prev = active_astate[t][bp[t][q]];
#if VITERBI_DEBUG
printf("Following non-emitting state at time %d, %u => %u\n",
t, j, prev);
#endif
/* Backtrace and accumulate transition counts. */
if (tmat_reest) {
assert(tacc != NULL);
tacc[prev][j - prev] += 1.0;
}
q = bp[t][q];
j = prev;
}
/* Now accumulate statistics for the real state. */
l_cb = state_seq[j].l_cb;
l_ci_cb = state_seq[j].l_ci_cb;
n_active_cb = 0;
if (gau_timer)
timing_start(gau_timer);
gauden_compute_log(now_den[l_cb],
now_den_idx[l_cb],
feature[t],
g,
state_seq[j].cb,
NULL);
active_cb[n_active_cb++] = l_cb;
if (l_cb != l_ci_cb) {
gauden_compute_log(now_den[l_ci_cb],
now_den_idx[l_ci_cb],
feature[t],
g,
state_seq[j].ci_cb,
NULL);
active_cb[n_active_cb++] = l_ci_cb;
}
gauden_scale_densities_bwd(now_den, now_den_idx,
&dscale[t],
active_cb, n_active_cb, g);
assert(state_seq[j].mixw != TYING_NON_EMITTING);
/* Now calculate mixture densities. */
/* This is the normalizer sum_m c_{jm} p(o_t|\lambda_{jm}) */
op = gauden_mixture(now_den[l_cb], now_den_idx[l_cb],
mixw[state_seq[j].mixw], g);
if (gau_timer)
timing_stop(gau_timer);
if (rsts_timer)
timing_start(rsts_timer);
/* Make up this bogus value to be consistent with backward.c */
p_reest_term = 1.0 / op;
/* Compute the output probability excluding the contribution
* of each feature stream. i.e. p_op[0] is the output
* probability excluding feature stream 0 */
partial_op(p_op,
op,
now_den[l_cb],
now_den_idx[l_cb],
mixw[state_seq[j].mixw],
n_feat,
n_top);
/* compute the probability of each (of possibly topn) density */
den_terms(d_term,
p_reest_term,
p_op,
now_den[l_cb],
now_den_idx[l_cb],
mixw[state_seq[j].mixw],
n_feat,
n_top);
if (l_cb != l_ci_cb) {
/* For each feature stream f, compute:
* sum_k(mixw[f][k] den[f][k])
* and store the results in p_ci_op */
partial_ci_op(p_ci_op,
now_den[l_ci_cb],
now_den_idx[l_ci_cb],
mixw[state_seq[j].ci_mixw],
n_feat,
n_top);
/* For each feature stream and density compute the terms:
* w[f][k] den[f][k] / sum_k(w[f][k] den[f][k]) * post_j
* and store results in d_term_ci */
den_terms_ci(d_term_ci,
1.0, /* post_j = 1.0 */
p_ci_op,
now_den[l_ci_cb],
now_den_idx[l_ci_cb],
mixw[state_seq[j].ci_mixw],
n_feat,
n_top);
}
/* accumulate the probability for each density in the mixing
* weight reestimation accumulators */
if (mixw_reest) {
accum_den_terms(wacc[state_seq[j].l_mixw], d_term,
now_den_idx[l_cb], n_feat, n_top);
/* check if mixw and ci_mixw are different to avoid
* doubling the EM counts in a CI run. */
if (state_seq[j].mixw != state_seq[j].ci_mixw) {
if (n_cb < inv->n_mixw) {
/* semi-continuous, tied mixture, and discrete case */
accum_den_terms(wacc[state_seq[j].l_ci_mixw], d_term,
now_den_idx[l_cb], n_feat, n_top);
}
else {
/* continuous case */
accum_den_terms(wacc[state_seq[j].l_ci_mixw], d_term_ci,
now_den_idx[l_ci_cb], n_feat, n_top);
}
}
}
/* accumulate the probability for each density in the
* density reestimation accumulators */
if (mean_reest || var_reest) {
accum_den_terms(denacc[l_cb], d_term,
now_den_idx[l_cb], n_feat, n_top);
if (l_cb != l_ci_cb) {
accum_den_terms(denacc[l_ci_cb], d_term_ci,
now_den_idx[l_ci_cb], n_feat, n_top);
}
}
if (rsts_timer)
timing_stop(rsts_timer);
/* Note that there is only one state/frame so this is kind of
redundant */
if (rstf_timer)
timing_start(rstf_timer);
if (mean_reest || var_reest) {
/* Update the mean and variance reestimation accumulators */
if (pdumpfh)
fprintf(pdumpfh, "time %d:\n", t);
accum_gauden(denacc,
cb_inv,
n_lcl_cb,
feature[t],
now_den_idx,
g,
mean_reest,
var_reest,
pass2var,
inv->l_mixw_acc,
var_is_full,
pdumpfh,
fcb);
memset(&denacc[0][0][0], 0, denacc_size);
}
if (rstf_timer)
timing_stop(rstf_timer);
if (t > 0) {
prev = active_astate[t-1][bp[t][q]];
#if VITERBI_DEBUG
printf("Backtrace at time %d, %u => %u\n",
t, j, prev);
#endif
/* Backtrace and accumulate transition counts. */
if (tmat_reest) {
assert(tacc != NULL);
tacc[prev][j-prev] += 1.0;
}
q = bp[t][q];
j = prev;
}
}
/* If no error was found, add the resulting utterance reestimation
* accumulators to the global reestimation accumulators */
if (rstu_timer)
timing_start(rstu_timer);
accum_global(inv, state_seq, n_state,
mixw_reest, tmat_reest, mean_reest, var_reest,
var_is_full);
if (rstu_timer)
timing_stop(rstu_timer);
/* Find the final state */
for (i = 0; i < n_active_astate[n_obs-1]; ++i) {
if (active_astate[n_obs-1][i] == n_state-1)
break;
}
/* Calculate log[ p( O | \lambda ) ] */
assert(active_alpha[n_obs-1][i] > 0);
log_fp = log(active_alpha[n_obs-1][i]);
for (t = 0; t < n_obs; t++) {
assert(scale[t] > 0);
log_fp -= log(scale[t]);
for (j = 0; j < inv->gauden->n_feat; j++) {
log_fp += dscale[t][j];
}
}
*log_forw_prob = log_fp;
all_done:
ckd_free((void *)scale);
for (i = 0; i < n_obs; i++) {
if (dscale[i])
ckd_free((void *)dscale[i]);
}
ckd_free((void **)dscale);
ckd_free(n_active_astate);
for (i = 0; i < n_obs; i++) {
ckd_free((void *)active_alpha[i]);
ckd_free((void *)active_astate[i]);
ckd_free((void *)bp[i]);
}
ckd_free((void *)active_alpha);
ckd_free((void *)active_astate);
ckd_free((void *)active_cb);
if (denacc)
ckd_free_3d((void ***)denacc);
if (now_den)
ckd_free_3d((void ***)now_den);
if (now_den_idx)
ckd_free_3d((void ***)now_den_idx);
if (ret != S3_SUCCESS)
E_ERROR("%s ignored\n", corpus_utt_brief_name());
return ret;
}
int
main(int argc, char *argv[])
{
int i, j, offset;
int32 noframe, vsize, dsize, column;
int32 frm_begin, frm_end;
int is_header, is_describe;
float *z, **cep;
char const *cepfile;
print_appl_info(argv[0]);
cmd_ln_appl_enter(argc, argv, "default.arg", arg);
vsize = cmd_ln_int32("-i");
dsize = cmd_ln_int32("-d");
frm_begin = cmd_ln_int32("-b");
frm_end = cmd_ln_int32("-e");
is_header = cmd_ln_int32("-header");
is_describe = cmd_ln_int32("-describe");
if (vsize < 0)
E_FATAL("-i : Input vector size should be larger than 0.\n");
if (dsize < 0)
E_FATAL("-d : Column size should be larger than 0\n");
if (frm_begin < 0)
E_FATAL("-b : Beginning frame should be larger than 0\n");
/* The following condition is redundant
* if (frm_end < 0) E_FATAL("-e : Ending frame should be larger than 0\n");
*/
if (frm_begin >= frm_end)
E_FATAL
("Ending frame (-e) should be larger than beginning frame (-b).\n");
if ((cepfile = cmd_ln_str("-f")) == NULL) {
E_FATAL("Input file was not specified with (-f)\n");
}
if (read_cep(cepfile, &cep, &noframe, vsize) == IO_ERR)
E_FATAL("ERROR opening %s for reading\n", cepfile);
z = cep[0];
offset = 0;
column = (vsize > dsize) ? dsize : vsize;
frm_end = (frm_end > noframe) ? noframe : frm_end;
E_INFO("Displaying %d out of %d columns per frame\n", column, vsize);
E_INFO("Total %d frames\n\n", noframe);
/* This part should be moved to a special library if this file is
longer than 300 lines. */
if (is_header) {
if (is_describe) {
printf("\n%6s", "frame#:");
}
for (j = 0; j < column; ++j) {
printf("%3s%3d%s ", "c[", j, "]");
}
printf("\n");
}
offset += frm_begin * vsize;
for (i = frm_begin; i < frm_end; ++i) {
if (is_describe) {
printf("%6d:", i);
}
for (j = 0; j < column; ++j)
printf("%7.3f ", z[offset + j]);
printf("\n");
offset += vsize;
}
fflush(stdout);
cmd_ln_appl_exit();
ckd_free_2d(cep);
return (IO_SUCCESS);
}
int
main(int argc, char *argv[])
{
model_def_t *mdef;
uint32 n_tmat;
uint32 n_tied_state;
uint32 n_state_pm;
uint32 n_stream;
uint32 n_density;
float32 ***tmat;
float32 **proto_tmat;
float32 ***mixw;
uint32 i, j, k;
float32 mixw_ini;
int retval = 0;
parse_cmd_ln(argc, argv);
printf("%s(%d): Reading model definition file %s\n",
__FILE__, __LINE__, cmd_ln_str("-moddeffn"));
if (model_def_read(&mdef, cmd_ln_str("-moddeffn")) !=
S3_SUCCESS) {
return 1;
}
printf("%s(%d): %d models defined\n",
__FILE__, __LINE__, mdef->n_defn);
if (!cmd_ln_str("-tmatfn") && ! cmd_ln_str("-mixwfn")){
E_FATAL("Both -tmatfn and -mixwfn were not specified, forced exit\n");
}
if (cmd_ln_str("-tmatfn")) {
if (topo_read(&proto_tmat, &n_state_pm, cmd_ln_str("-topo")) != S3_SUCCESS)
return 1;
/* proto_tmat is normalized */
n_tmat = mdef->n_tied_tmat;
tmat = (float32 ***)ckd_calloc_3d(n_tmat, n_state_pm-1, n_state_pm,
sizeof(float32));
for (k = 0; k < n_tmat; k++) {
for (i = 0; i < n_state_pm-1; i++) {
for (j = 0; j < n_state_pm; j++) {
/* perhaps this could be replaced
with a block copy per tmat */
tmat[k][i][j] = proto_tmat[i][j];
}
}
}
if (s3tmat_write(cmd_ln_str("-tmatfn"),
tmat,
n_tmat,
n_state_pm) != S3_SUCCESS) {
retval = 1;
}
ckd_free_3d((void ***)tmat);
}
else {
E_INFO("No tmat file given; none generated\n");
}
n_tied_state = mdef->n_tied_state;
n_stream = cmd_ln_int32("-nstream");
n_density = cmd_ln_int32("-ndensity");
mixw = (float32 ***)ckd_calloc_3d(n_tied_state, n_stream, n_density,
sizeof(float32));
/* weight each density uniformly */
mixw_ini = 1.0 / (float)n_density;
for (i = 0; i < n_tied_state; i++) {
for (j = 0; j < n_stream; j++) {
for (k = 0; k < n_density; k++) {
mixw[i][j][k] = mixw_ini;
}
}
}
if (cmd_ln_str("-mixwfn")) {
if (s3mixw_write(cmd_ln_str("-mixwfn"),
mixw,
n_tied_state,
n_stream,
n_density) != S3_SUCCESS) {
retval = 2;
}
}
else {
E_INFO("No mixw file given; none generated\n");
}
ckd_free_3d((void ***)mixw);
return retval;
}
