int
s2_write_dhmm(float32 ***tmat,
float32 ***mixw,
model_def_t *mdef,
const char *dir)
{
acmod_set_t *acmod_set;
model_def_entry_t *defn;
uint32 b, n_state, i, j, err;
uint32 n_acmod;
float32 ***omixw = NULL;
err = 0;
acmod_set = mdef->acmod_set;
n_acmod = acmod_set_n_acmod(acmod_set);
for (i = 0; i < n_acmod; i++) {
defn = &mdef->defn[i];
b = (uint32)acmod_set_base_phone(acmod_set, (acmod_id_t)i);
if (omixw) ckd_free((void *)omixw);
n_state = defn->n_state;
omixw = (float32 ***)ckd_calloc(n_state, sizeof(float32 **));
for (j = 0; j < n_state-1; j++) {
omixw[j] = mixw[defn->state[j]];
}
if (put_dhmm(tmat[b],
omixw,
dir,
acmod_set_id2s2name(acmod_set, (acmod_id_t)i)) != S3_SUCCESS)
err = 1;
}
if (omixw) ckd_free((void *)omixw);
if (!err) {
return S3_SUCCESS;
}
else {
return S3_ERROR;
}
}
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;
}
/* Converts the state map, smap, into a global
one-to-one mapping from model states onto
consecutive integers from 0 to some max */
int
s2_convert_smap_to_global(acmod_set_t *acmod_set,
uint32 **smap,
uint32 **out_state_of,
uint32 *cluster_size)
{
uint32 i;
uint32 t;
uint32 offset;
uint32 n_ci = acmod_set_n_ci(acmod_set);
uint32 n = acmod_set_n_multi(acmod_set) + n_ci;
acmod_id_t b;
acmod_id_t id;
uint32 state;
uint32 *state_of;
uint32 ci_seno;
offset = n_ci * (S2_N_STATE-1);
E_INFO("|CI states| == %d\n", offset);
/* convert the cluster_size matrix into an offset table */
for (i = 0; i < n_ci; i++) {
t = cluster_size[i];
cluster_size[i] = offset;
offset += t;
E_INFO("|%s| == %d (%d)\n",
acmod_set_id2name(acmod_set, i),
t, offset);
}
cluster_size[i] = offset; /* total # of states */
state_of = ckd_calloc(offset, sizeof(uint32));
/* map the ci phones to unshared distribution */
for (id = 0; id < n_ci; id++) {
for (state = 0; state < S2_N_STATE-1; state++) {
ci_seno = id * (S2_N_STATE-1) + state;
smap[id][state] = ci_seno;
state_of[ci_seno] = state;
}
}
/* use the ci phone offsets to convert ci phone relative
mappings to a global one-to-one mapping onto
consecutive integers from 0 to some max */
for (; id < n; id++) {
for (state = 0; state < S2_N_STATE-1; state++) {
if (smap[id][state] == TYING_NO_ID) {
uint32 base_id;
base_id = acmod_set_base_phone(acmod_set, id);
E_WARN("%s<%d> is unmapped, approximating with CI state <%d>.\n",
acmod_set_id2name(acmod_set, id),
state, state);
smap[id][state] = smap[base_id][state];
/* no state_of[] assignment need bee done since it was done
above */
}
else {
b = acmod_set_base_phone(acmod_set, id);
smap[id][state] += cluster_size[b];
state_of[ smap[id][state] ] = state;
}
}
}
*out_state_of = state_of;
return S3_SUCCESS;
}
int
main(int argc, char *argv[])
{
model_def_t *mdef;
model_def_entry_t *defn;
uint32 n_defn;
uint32 *cluster_offset;
uint32 max_int;
uint32 *state_of;
uint32 max_state;
uint32 sstate;
int32 i;
uint32 j;
uint32 n_base_phone;
acmod_id_t base;
acmod_id_t p;
float32 ***out;
uint32 **smap;
char comment[4192];
time_t t;
parse_cmd_ln(argc, argv);
printf("%s(%d): Reading model definition file %s\n",
__FILE__, __LINE__, (const char *)cmd_ln_access("-moddeffn"));
if (model_def_read(&mdef, cmd_ln_access("-moddeffn")) != S3_SUCCESS) {
exit(1);
}
defn = mdef->defn;
n_defn = mdef->n_defn;
printf("%s(%d): %d models defined\n",
__FILE__, __LINE__, n_defn);
smap = ckd_calloc(n_defn, sizeof(uint32 *));
n_base_phone = acmod_set_n_ci(mdef->acmod_set);
cluster_offset = ckd_calloc(n_base_phone+1, sizeof(uint32));
max_int = 0;
--max_int; /* underflow offset values to max value */
for (i = 0; i < n_base_phone; i++) {
cluster_offset[i] = max_int;
}
for (i = 0, max_state = 0; i < n_defn; i++) {
for (j = 0; j < defn[i].n_state; j++) {
sstate = defn[i].state[j];
if ((sstate != TYING_NON_EMITTING) &&
(defn[i].state[j] > max_state)) max_state = defn[i].state[j];
}
}
/* record the total # of senones */
cluster_offset[n_base_phone] = max_state+1;
state_of = ckd_calloc(max_state+1, sizeof(uint32));
for (i = 0; i <= max_state; i++)
state_of[i] = NO_STATE;
for (i = 0; i < n_defn; i++) {
p = defn[i].p;
base = acmod_set_base_phone(mdef->acmod_set, defn[i].p);
smap[i] = defn[i].state;
for (j = 0; j < defn[i].n_state; j++) {
sstate = defn[i].state[j];
if (sstate != TYING_NON_EMITTING) {
if (state_of[sstate] == NO_STATE)
state_of[sstate] = j;
else if (state_of[sstate] != j) {
printf("%s %d appears as %d%s and %d%s model states\n",
acmod_set_id2name(mdef->acmod_set, acmod_set_base_phone(mdef->acmod_set, defn[i].p)),
sstate,
state_of[sstate],
ord_suff(state_of[sstate]),
j,
ord_suff(j));
}
if ((p != base) && (cluster_offset[base] > sstate)) {
cluster_offset[base] = sstate;
}
}
}
}
/* any untouched CLUSTER_OFFSET's implies a base phone
without any CD states. So offset is same as next
one */
for (i = (n_base_phone - 1); i >= 0 ; i--) {
if (cluster_offset[i] == max_int)
cluster_offset[i] = cluster_offset[i+1];
}
fflush(stdout);
for (i = 0; i < n_base_phone; i++) {
if (cluster_offset[i] != max_int) {
fprintf(stderr, "%s(%d): %s offset %d\n",
__FILE__, __LINE__,
acmod_set_id2name(mdef->acmod_set, i), cluster_offset[i]);
}
else {
fprintf(stderr, "%s(%d): %s <no CD states>\n",
__FILE__, __LINE__,
acmod_set_id2name(mdef->acmod_set, i));
}
}
fflush(stderr);
printf("%s(%d): Reading senone weights in %s with floor %e\n",
__FILE__, __LINE__, (const char *)cmd_ln_access("-hmmdir"),
*(float32 *)cmd_ln_access("-floor"));
out = s2_read_seno_3(mdef->acmod_set, cluster_offset,
cmd_ln_access("-hmmdir"),
(*(int32 *)cmd_ln_access("-ci2cd") ? NULL : smap),
*(float32 *)cmd_ln_access("-floor"),
state_of);
t = time(NULL);
sprintf(comment,
"Generated on %s\n\tmoddeffn: %s\n\tfloor: %e\n\thmmdir: %s\n\n\n\n\n\n\n\n\n",
ctime(&t),
(const char *)cmd_ln_access("-moddeffn"),
*(float32 *)cmd_ln_access("-floor"),
(const char *)cmd_ln_access("-hmmdir"));
fflush(stdout);
fprintf(stderr, "%s(%d): writing %s\n",
__FILE__, __LINE__,
(const char *)cmd_ln_access("-mixwfn"));
fflush(stderr);
if (s3mixw_write(cmd_ln_access("-mixwfn"),
out,
cluster_offset[n_base_phone], /* total # states */
S2_N_FEATURE,
S2_N_CODEWORD) != S3_SUCCESS) {
fflush(stdout);
fprintf(stderr, "%s(%d): couldn't write mixture weight file\n",
__FILE__, __LINE__);
perror(cmd_ln_access("-mixwfn"));
fflush(stderr);
}
ckd_free(state_of);
ckd_free(cluster_offset);
return 0;
}
static int
init_mixw()
{
model_def_t *src_mdef;
float32 ***src_mixw;
vector_t ***src_mean;
vector_t ***src_var = NULL;
vector_t ****src_fullvar = NULL;
float32 ***src_tmat;
model_def_t *dest_mdef;
float32 ***dest_mixw;
vector_t ***dest_mean;
vector_t ***dest_var = NULL;
vector_t ****dest_fullvar = NULL;
float32 ***dest_tmat;
uint32 n_mixw_src;
uint32 n_mixw_dest;
uint32 n_feat;
uint32 tmp_n_feat;
uint32 n_gau;
uint32 tmp_n_gau;
uint32 n_cb_src;
uint32 n_cb_dest;
uint32 n_state_pm;
uint32 n_tmat_src;
uint32 n_tmat_dest;
uint32 *veclen;
uint32 *tmp_veclen;
uint32 m;
uint32 dest_m;
uint32 dest_m_base;
uint32 src_m;
acmod_id_t src_m_base;
const char *dest_m_name;
const char *dest_m_base_name;
uint32 i;
uint32 n_ts;
uint32 n_cb;
const char *ts2cbfn;
E_INFO("Reading src %s\n", cmd_ln_str("-src_moddeffn"));
/* read in the source model definition file */
if (model_def_read(&src_mdef,
cmd_ln_str("-src_moddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
ts2cbfn = cmd_ln_str("-src_ts2cbfn");
if (strcmp(SEMI_LABEL, ts2cbfn) == 0) {
E_INFO("Generating semi-continous ts2cb mapping\n");
src_mdef->cb = semi_ts2cb(src_mdef->n_tied_state);
n_ts = src_mdef->n_tied_state;
n_cb = 1;
}
else if (strcmp(CONT_LABEL, ts2cbfn) == 0) {
E_INFO("Generating continous ts2cb mapping\n");
src_mdef->cb = cont_ts2cb(src_mdef->n_tied_state);
n_ts = src_mdef->n_tied_state;
n_cb = src_mdef->n_tied_state;
}
else if (strcmp(PTM_LABEL, ts2cbfn) == 0) {
E_INFO("Generating phonetically tied ts2cb mapping\n");
src_mdef->cb = ptm_ts2cb(src_mdef);
n_ts = src_mdef->n_tied_state;
n_cb = src_mdef->acmod_set->n_ci;
}
else {
E_INFO("Reading src %s\n", cmd_ln_str("-src_ts2cbfn"));
if (s3ts2cb_read(ts2cbfn,
&src_mdef->cb,
&n_ts,
&n_cb) != S3_SUCCESS) {
return S3_ERROR;
}
}
E_INFO("Reading src %s\n", cmd_ln_str("-src_mixwfn"));
/* read in the source mixing weight parameter file */
if (s3mixw_read(cmd_ln_str("-src_mixwfn"),
&src_mixw, &n_mixw_src, &n_feat, &n_gau) != S3_SUCCESS) {
return S3_ERROR;
}
E_INFO("Reading src %s\n",
cmd_ln_str("-src_tmatfn"));
if (s3tmat_read(cmd_ln_str("-src_tmatfn"),
&src_tmat,
&n_tmat_src,
&n_state_pm) != S3_SUCCESS) {
return S3_ERROR;
}
E_INFO("Reading src %s\n", cmd_ln_str("-src_meanfn"));
if (s3gau_read(cmd_ln_str("-src_meanfn"),
&src_mean,
&n_cb_src,
&tmp_n_feat,
&tmp_n_gau,
&veclen) != S3_SUCCESS) {
return S3_ERROR;
}
if (tmp_n_feat != n_feat) {
E_FATAL("src mean n_feat (== %u) != prior value (== %u)\n",
tmp_n_feat, n_feat);
}
if (tmp_n_gau != n_gau) {
E_FATAL("src mean n_gau (== %u) != prior value (== %u)\n",
tmp_n_gau, n_gau);
}
if (n_cb_src != n_cb) {
E_FATAL("src mean n_cb (== %u) is inconsistent with ts2cb mapping %u. Most probably phoneset has duplicated phones\n",
n_cb_src, n_cb);
}
E_INFO("Reading src %s\n", cmd_ln_str("-src_varfn"));
if (cmd_ln_int32("-fullvar")) {
if (s3gau_read_full(cmd_ln_str("-src_varfn"),
&src_fullvar,
&n_cb_src,
&tmp_n_feat,
&tmp_n_gau,
&tmp_veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
if (s3gau_read(cmd_ln_str("-src_varfn"),
&src_var,
&n_cb_src,
&tmp_n_feat,
&tmp_n_gau,
&tmp_veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
if (tmp_n_feat != n_feat) {
E_FATAL("src var n_feat (== %u) != prior value (== %u)\n",
tmp_n_feat, n_feat);
}
if (tmp_n_gau != n_gau) {
E_FATAL("src var n_gau (== %u) != prior value (== %u)\n",
tmp_n_gau, n_gau);
}
if (n_cb_src != n_cb) {
E_FATAL("src var n_cb (== %u) inconsistent w/ ts2cb mapping %u\n",
n_cb_src, n_cb);
}
if (n_mixw_src < src_mdef->n_tied_state) {
E_FATAL("Too few source mixing weights, %u, for the # of tied states, %u\n",
n_mixw_src, src_mdef->n_tied_state);
}
for (i = 0; i < n_feat; i++) {
if (veclen[i] != tmp_veclen[i]) {
E_FATAL("src var veclen[%u] (== %u) != prior value (== %u)\n",
i, tmp_veclen[i], veclen[i]);
}
}
ckd_free(tmp_veclen);
E_INFO("Reading dest %s\n",
cmd_ln_str("-dest_moddeffn"));
/* read in the destination model definition file */
if (model_def_read(&dest_mdef,
cmd_ln_str("-dest_moddeffn")) < S3_SUCCESS) {
return S3_ERROR;
}
ts2cbfn = cmd_ln_str("-dest_ts2cbfn");
if (strcmp(SEMI_LABEL, ts2cbfn) == 0) {
E_INFO("Generating semi-continous ts2cb mapping\n");
dest_mdef->cb = semi_ts2cb(dest_mdef->n_tied_state);
n_ts = dest_mdef->n_tied_state;
n_cb = 1;
}
else if (strcmp(CONT_LABEL, ts2cbfn) == 0) {
E_INFO("Generating continous ts2cb mapping\n");
dest_mdef->cb = cont_ts2cb(dest_mdef->n_tied_state);
n_ts = dest_mdef->n_tied_state;
n_cb = dest_mdef->n_tied_state;
}
else if (strcmp(PTM_LABEL, ts2cbfn) == 0) {
E_INFO("Generating phonetically tied ts2cb mapping\n");
dest_mdef->cb = ptm_ts2cb(dest_mdef);
n_ts = dest_mdef->n_tied_state;
n_cb = dest_mdef->acmod_set->n_ci;
}
else {
E_INFO("Reading dest %s\n",
cmd_ln_str("-dest_ts2cbfn"));
if (s3ts2cb_read(ts2cbfn,
&dest_mdef->cb,
&n_ts,
&n_cb) != S3_SUCCESS) {
return S3_ERROR;
}
}
E_INFO("Calculating initial model parameters\n");
n_tmat_dest = dest_mdef->n_tied_tmat;
tmat_dest_list = init_was_added(n_tmat_dest);
E_INFO("Alloc %ux%ux%u dest tmat\n",
n_tmat_dest,
n_state_pm-1,
n_state_pm);
dest_tmat = (float32 ***)ckd_calloc_3d(n_tmat_dest,
n_state_pm-1,
n_state_pm,
sizeof(float32));
n_mixw_dest = dest_mdef->n_tied_state;
mixw_dest_list = init_was_added(n_mixw_dest);
E_INFO("Alloc %ux%ux%u dest mixw\n",
n_mixw_dest, n_feat, n_gau);
dest_mixw = (float32 ***)ckd_calloc_3d(n_mixw_dest, n_feat, n_gau, sizeof(float32));
for (i = 0, n_cb_dest = 0; i < n_mixw_dest; i++) {
if (dest_mdef->cb[i] > n_cb_dest) {
n_cb_dest = dest_mdef->cb[i];
}
}
++n_cb_dest;
cb_dest_list = init_was_added(n_cb_dest);
E_INFO("Alloc %ux%ux%u dest mean and var\n",
n_cb_dest, n_feat, n_gau);
dest_mean = gauden_alloc_param(n_cb_dest, n_feat, n_gau, veclen);
if (src_var)
dest_var = gauden_alloc_param(n_cb_dest, n_feat, n_gau, veclen);
else if (src_fullvar)
dest_fullvar = gauden_alloc_param_full(n_cb_dest, n_feat, n_gau, veclen);
for (dest_m = 0; dest_m < dest_mdef->n_defn; dest_m++) {
dest_m_name = acmod_set_id2name(dest_mdef->acmod_set, dest_m);
src_m = acmod_set_name2id(src_mdef->acmod_set, dest_m_name);
if (src_m == NO_ACMOD) {
/* No corresponding phone model in the source set */
/* See if there is a source base phone corresponding to this destination model
base phone */
dest_m_base = acmod_set_base_phone(dest_mdef->acmod_set, dest_m);
dest_m_base_name = acmod_set_id2name(dest_mdef->acmod_set, dest_m_base);
src_m_base = acmod_set_name2id(src_mdef->acmod_set, dest_m_base_name);
if (src_m_base == NO_ACMOD) {
/* No corresponding model or base model found. Use uniform distribution */
E_INFO("No source base phone %s found. Initializing %s using uniform distribution\n",
dest_m_base_name, dest_m_name);
if (src_tmat) {
E_INFO("Uniform initialization of tmat not supported\n");
}
init_uniform(dest_mixw, &dest_mdef->defn[dest_m], n_feat, n_gau);
}
else {
/* No corresponding model, but a base model was found. Use base distribution. */
init_model(dest_mixw, dest_mean, dest_var, dest_fullvar, dest_tmat,
&dest_mdef->defn[dest_m], dest_mdef->cb, dest_mdef->acmod_set,
src_mixw, src_mean, src_var, src_fullvar, src_tmat,
&src_mdef->defn[src_m_base], src_mdef->cb, src_mdef->acmod_set,
n_feat, n_gau, n_state_pm, veclen);
}
}
else {
/* Found a corresponding model in the source set, so use source distributions to init
the destination */
init_model(dest_mixw, dest_mean, dest_var, dest_fullvar, dest_tmat,
&dest_mdef->defn[dest_m], dest_mdef->cb, dest_mdef->acmod_set,
src_mixw, src_mean, src_var, src_fullvar, src_tmat,
&src_mdef->defn[src_m], src_mdef->cb, src_mdef->acmod_set,
n_feat, n_gau, n_state_pm, veclen);
}
}
for (m = 0; m < n_mixw_dest; m++) {
if (mixw_dest_list[m] == NULL) {
E_WARN("Destination state %u has not been initialized!\n", m);
}
}
for (m = 0; m < n_cb_dest; m++) {
if (cb_dest_list[m] == NULL) {
E_WARN("Destination cb %u has not been initialized!\n", m);
}
else if (cb_dest_list[m]->next != NULL) {
E_WARN("dest cb %u has > 1 corresponding source cb\n", m);
}
}
E_INFO("Writing dest %s\n",
cmd_ln_str("-dest_tmatfn"));
if (s3tmat_write(cmd_ln_str("-dest_tmatfn"),
dest_tmat,
n_tmat_dest,
n_state_pm) != S3_SUCCESS) {
return S3_ERROR;
}
E_INFO("Writing dest %s\n",
cmd_ln_str("-dest_mixwfn"));
if (s3mixw_write(cmd_ln_str("-dest_mixwfn"),
dest_mixw,
dest_mdef->n_tied_state, n_feat, n_gau) < S3_SUCCESS) {
return S3_ERROR;
}
E_INFO("Writing dest %s\n",
cmd_ln_str("-dest_meanfn"));
if (s3gau_write(cmd_ln_str("-dest_meanfn"),
(const vector_t ***)dest_mean,
n_cb_dest,
n_feat,
n_gau,
veclen) != S3_SUCCESS) {
return S3_ERROR;
}
E_INFO("Writing dest %s\n",
cmd_ln_str("-dest_varfn"));
if (cmd_ln_int32("-fullvar")) {
if (s3gau_write_full(cmd_ln_str("-dest_varfn"),
(const vector_t ****)dest_fullvar,
n_cb_dest,
n_feat,
n_gau,
veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
if (s3gau_write(cmd_ln_str("-dest_varfn"),
(const vector_t ***)dest_var,
n_cb_dest,
n_feat,
n_gau,
veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
ckd_free(veclen);
return S3_SUCCESS;
}