ps_search_t *
state_align_search_init(const char *name,
cmd_ln_t *config,
acmod_t *acmod,
ps_alignment_t *al)
{
state_align_search_t *sas;
ps_alignment_iter_t *itor;
hmm_t *hmm;
sas = ckd_calloc(1, sizeof(*sas));
ps_search_init(ps_search_base(sas), &state_align_search_funcs,
PS_SEARCH_TYPE_STATE_ALIGN, name,
config, acmod, al->d2p->dict, al->d2p);
sas->hmmctx = hmm_context_init(bin_mdef_n_emit_state(acmod->mdef),
acmod->tmat->tp, NULL, acmod->mdef->sseq);
if (sas->hmmctx == NULL) {
ckd_free(sas);
return NULL;
}
sas->al = al;
/* Generate HMM vector from phone level of alignment. */
sas->n_phones = ps_alignment_n_phones(al);
sas->n_emit_state = ps_alignment_n_states(al);
sas->hmms = ckd_calloc(sas->n_phones, sizeof(*sas->hmms));
for (hmm = sas->hmms, itor = ps_alignment_phones(al); itor;
++hmm, itor = ps_alignment_iter_next(itor)) {
ps_alignment_entry_t *ent = ps_alignment_iter_get(itor);
hmm_init(sas->hmmctx, hmm, FALSE,
ent->id.pid.ssid, ent->id.pid.tmatid);
}
return ps_search_base(sas);
}
static int
phone_loop_search_reinit(ps_search_t *search, dict_t *dict, dict2pid_t *d2p)
{
phone_loop_search_t *pls = (phone_loop_search_t *)search;
cmd_ln_t *config = ps_search_config(search);
acmod_t *acmod = ps_search_acmod(search);
int i;
/* Free old dict2pid, dict, if necessary. */
ps_search_base_reinit(search, dict, d2p);
/* Initialize HMM context. */
if (pls->hmmctx)
hmm_context_free(pls->hmmctx);
pls->hmmctx = hmm_context_init(bin_mdef_n_emit_state(acmod->mdef),
acmod->tmat->tp, NULL, acmod->mdef->sseq);
if (pls->hmmctx == NULL)
return -1;
/* Initialize penalty storage */
pls->n_phones = bin_mdef_n_ciphone(acmod->mdef);
pls->window = cmd_ln_int32_r(config, "-pl_window");
if (pls->penalties)
ckd_free(pls->penalties);
pls->penalties = (int32 *)ckd_calloc(pls->n_phones, sizeof(*pls->penalties));
if (pls->pen_buf)
ckd_free_2d(pls->pen_buf);
pls->pen_buf = (int32 **)ckd_calloc_2d(pls->window, pls->n_phones, sizeof(**pls->pen_buf));
/* Initialize phone HMMs. */
if (pls->hmms) {
for (i = 0; i < pls->n_phones; ++i)
hmm_deinit((hmm_t *)&pls->hmms[i]);
ckd_free(pls->hmms);
}
pls->hmms = (hmm_t *)ckd_calloc(pls->n_phones, sizeof(*pls->hmms));
for (i = 0; i < pls->n_phones; ++i) {
hmm_init(pls->hmmctx, (hmm_t *)&pls->hmms[i],
FALSE,
bin_mdef_pid2ssid(acmod->mdef, i),
bin_mdef_pid2tmatid(acmod->mdef, i));
}
pls->penalty_weight = cmd_ln_float64_r(config, "-pl_weight");
pls->beam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_beam")) >> SENSCR_SHIFT;
pls->pbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_pbeam")) >> SENSCR_SHIFT;
pls->pip = logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-pl_pip")) >> SENSCR_SHIFT;
E_INFO("State beam %d Phone exit beam %d Insertion penalty %d\n",
pls->beam, pls->pbeam, pls->pip);
return 0;
}
static void init_hmm_unrelated(args_t *args)
{
int i,j;
args->nstates = 7;
args->tprob = (double*) malloc(sizeof(double)*args->nstates*args->nstates);
for (i=0; i<args->nstates; i++)
{
for (j=0; j<args->nstates; j++)
MAT(args->tprob,args->nstates,i,j) = args->pij;
}
MAT(args->tprob,args->nstates,UNRL_0101,UNRL_xxxx) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_0110,UNRL_xxxx) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_x0x0,UNRL_0x0x) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_0110,UNRL_0x0x) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_x00x,UNRL_0xx0) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_0101,UNRL_0xx0) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_0101,UNRL_x00x) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_0110,UNRL_x0x0) = args->pij*args->pij;
MAT(args->tprob,args->nstates,UNRL_0110,UNRL_0101) = args->pij*args->pij;
for (i=0; i<args->nstates; i++)
{
for (j=i+1; j<args->nstates; j++)
MAT(args->tprob,args->nstates,i,j) = MAT(args->tprob,args->nstates,j,i);
}
for (i=0; i<args->nstates; i++)
{
double sum = 0;
for (j=0; j<args->nstates; j++)
if ( i!=j ) sum += MAT(args->tprob,args->nstates,i,j);
MAT(args->tprob,args->nstates,i,i) = 1 - sum;
}
#if 0
for (i=0; i<args->nstates; i++)
{
for (j=0; j<args->nstates; j++)
fprintf(stderr,"\t%e",MAT(args->tprob,args->nstates,j,i));
fprintf(stderr,"\n");
}
#endif
args->hmm = hmm_init(args->nstates, args->tprob, 10000);
}
static void
ngram_fwdflat_allocate_1ph(ngram_search_t *ngs)
{
dict_t *dict = ps_search_dict(ngs);
int n_words = ps_search_n_words(ngs);
int i, w;
/* Allocate single-phone words, since they won't have
* been allocated for us by fwdtree initialization. */
ngs->n_1ph_words = 0;
for (w = 0; w < n_words; w++) {
if (dict_is_single_phone(dict, w))
++ngs->n_1ph_words;
}
ngs->single_phone_wid = ckd_calloc(ngs->n_1ph_words,
sizeof(*ngs->single_phone_wid));
ngs->rhmm_1ph = ckd_calloc(ngs->n_1ph_words, sizeof(*ngs->rhmm_1ph));
i = 0;
for (w = 0; w < n_words; w++) {
if (!dict_is_single_phone(dict, w))
continue;
/* DICT2PID location */
ngs->rhmm_1ph[i].ciphone = dict_first_phone(dict, w);
ngs->rhmm_1ph[i].ci2phone = bin_mdef_silphone(ps_search_acmod(ngs)->mdef);
hmm_init(ngs->hmmctx, &ngs->rhmm_1ph[i].hmm, TRUE,
/* ssid */ bin_mdef_pid2ssid(ps_search_acmod(ngs)->mdef,
ngs->rhmm_1ph[i].ciphone),
/* tmatid */ bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef,
ngs->rhmm_1ph[i].ciphone));
ngs->rhmm_1ph[i].next = NULL;
ngs->word_chan[w] = (chan_t *) &(ngs->rhmm_1ph[i]);
ngs->single_phone_wid[i] = w;
i++;
}
}
/*
* Add the word emitted by the given transition (fsglink) to the given lextree
* (rooted at root), and return the new lextree root. (There may actually be
* several root nodes, maintained in a linked list via fsg_pnode_t.sibling.
* "root" is the head of this list.)
* lclist, rclist: sets of left and right context phones for this link.
* alloc_head: head of a linear list of all allocated pnodes for the parent
* FSG state, kept elsewhere and updated by this routine.
*/
static fsg_pnode_t *
psubtree_add_trans(fsg_lextree_t *lextree,
fsg_pnode_t * root,
fsg_glist_linklist_t **curglist,
fsg_link_t * fsglink,
int16 *lclist, int16 *rclist,
fsg_pnode_t ** alloc_head)
{
int32 silcipid; /* Silence CI phone ID */
int32 pronlen; /* Pronunciation length */
int32 wid; /* FSG (not dictionary!!) word ID */
int32 dictwid; /* Dictionary (not FSG!!) word ID */
int32 ssid; /* Senone Sequence ID */
int32 tmatid;
gnode_t *gn;
fsg_pnode_t *pnode, *pred, *head;
int32 n_ci, p, lc, rc;
glist_t lc_pnodelist; /* Temp pnodes list for different left contexts */
glist_t rc_pnodelist; /* Temp pnodes list for different right contexts */
int32 i, j;
int n_lc_alloc = 0, n_int_alloc = 0, n_rc_alloc = 0;
silcipid = bin_mdef_silphone(lextree->mdef);
n_ci = bin_mdef_n_ciphone(lextree->mdef);
wid = fsg_link_wid(fsglink);
assert(wid >= 0); /* Cannot be a null transition */
dictwid = dict_wordid(lextree->dict,
fsg_model_word_str(lextree->fsg, wid));
pronlen = dict_pronlen(lextree->dict, dictwid);
assert(pronlen >= 1);
assert(lclist[0] >= 0); /* At least one phonetic context provided */
assert(rclist[0] >= 0);
head = *alloc_head;
pred = NULL;
if (pronlen == 1) { /* Single-phone word */
int ci = dict_first_phone(lextree->dict, dictwid);
/* Only non-filler words are mpx */
if (dict_filler_word(lextree->dict, dictwid)) {
/*
* Left diphone ID for single-phone words already assumes SIL is right
* context; only left contexts need to be handled.
*/
lc_pnodelist = NULL;
for (i = 0; lclist[i] >= 0; i++) {
lc = lclist[i];
ssid = dict2pid_lrdiph_rc(lextree->d2p, ci, lc, silcipid);
tmatid = bin_mdef_pid2tmatid(lextree->mdef, dict_first_phone(lextree->dict, dictwid));
/* Check if this ssid already allocated for some other context */
for (gn = lc_pnodelist; gn; gn = gnode_next(gn)) {
pnode = (fsg_pnode_t *) gnode_ptr(gn);
if (hmm_nonmpx_ssid(&pnode->hmm) == ssid) {
/* already allocated; share it for this context phone */
fsg_pnode_add_ctxt(pnode, lc);
break;
}
}
if (!gn) { /* ssid not already allocated */
pnode =
(fsg_pnode_t *) ckd_calloc(1, sizeof(fsg_pnode_t));
pnode->ctx = lextree->ctx;
pnode->next.fsglink = fsglink;
pnode->logs2prob =
(fsg_link_logs2prob(fsglink) >> SENSCR_SHIFT)
+ lextree->wip + lextree->pip;
pnode->ci_ext = dict_first_phone(lextree->dict, dictwid);
pnode->ppos = 0;
pnode->leaf = TRUE;
pnode->sibling = root; /* All root nodes linked together */
fsg_pnode_add_ctxt(pnode, lc); /* Initially zeroed by calloc above */
pnode->alloc_next = head;
head = pnode;
root = pnode;
++n_lc_alloc;
hmm_init(lextree->ctx, &pnode->hmm, FALSE, ssid, tmatid);
lc_pnodelist =
glist_add_ptr(lc_pnodelist, (void *) pnode);
}
}
/* segment using HMM and then histogram clustering */
void cluster_segment(int* q, double** features, int frames_read, int feature_length, int nHMM_states,
int histogram_length, int nclusters, int neighbour_limit)
{
int i, j;
/*****************************/
if (0) {
/* try just using the predominant bin number as a 'decoded state' */
nHMM_states = feature_length + 1; /* allow a 'zero' state */
double chroma_thresh = 0.05;
double maxval;
int maxbin;
for (i = 0; i < frames_read; i++)
{
maxval = 0;
for (j = 0; j < feature_length; j++)
{
if (features[i][j] > maxval)
{
maxval = features[i][j];
maxbin = j;
}
}
if (maxval > chroma_thresh)
q[i] = maxbin;
else
q[i] = feature_length;
}
}
if (1) {
/*****************************/
/* scale all the features to 'balance covariances' during HMM training */
double scale = 10;
for (i = 0; i < frames_read; i++)
for (j = 0; j < feature_length; j++)
features[i][j] *= scale;
/* train an HMM on the features */
/* create a model */
model_t* model = hmm_init(features, frames_read, feature_length, nHMM_states);
/* train the model */
hmm_train(features, frames_read, model);
/*
printf("\n\nafter training:\n");
hmm_print(model);
*/
/* decode the hidden state sequence */
viterbi_decode(features, frames_read, model, q);
hmm_close(model);
/*****************************/
}
/*****************************/
/*
fprintf(stderr, "HMM state sequence:\n");
for (i = 0; i < frames_read; i++)
fprintf(stderr, "%d ", q[i]);
fprintf(stderr, "\n\n");
*/
/* create histograms of states */
double* h = (double*) malloc(frames_read*nHMM_states*sizeof(double)); /* vector in row major order */
create_histograms(q, frames_read, nHMM_states, histogram_length, h);
/* cluster the histograms */
int nbsched = 20; /* length of inverse temperature schedule */
double* bsched = (double*) malloc(nbsched*sizeof(double)); /* inverse temperature schedule */
double b0 = 100;
double alpha = 0.7;
bsched[0] = b0;
for (i = 1; i < nbsched; i++)
bsched[i] = alpha * bsched[i-1];
cluster_melt(h, nHMM_states, frames_read, bsched, nbsched, nclusters, neighbour_limit, q);
/* now q holds a sequence of cluster assignments */
free(h);
free(bsched);
}
/**
* Build net from phone HMMs
*/
static int
phmm_build(allphone_search_t * allphs)
{
phmm_t *p, **pid2phmm;
bin_mdef_t *mdef;
int32 lrc_size;
uint32 *lc, *rc;
s3pid_t pid;
s3cipid_t ci;
s3cipid_t *filler;
int n_phmm, n_link;
int i, nphone;
mdef = ((ps_search_t *) allphs)->acmod->mdef;
allphs->ci_phmm =
(phmm_t **) ckd_calloc(bin_mdef_n_ciphone(mdef), sizeof(phmm_t *));
pid2phmm =
(phmm_t **) ckd_calloc(bin_mdef_n_phone(mdef), sizeof(phmm_t *));
/* For each unique ciphone/triphone entry in mdef, create a PHMM node */
n_phmm = 0;
nphone = allphs->ci_only ? bin_mdef_n_ciphone(mdef) : bin_mdef_n_phone(mdef);
E_INFO("Building PHMM net of %d phones\n", nphone);
for (pid = 0; pid < nphone; pid++) {
if ((p = phmm_lookup(allphs, pid)) == NULL) {
//not found, should be created
p = (phmm_t *) ckd_calloc(1, sizeof(*p));
hmm_init(allphs->hmmctx, &(p->hmm), FALSE,
mdef_pid2ssid(mdef, pid), mdef->phone[pid].tmat);
p->pid = pid;
p->ci = bin_mdef_pid2ci(mdef, pid);
p->succlist = NULL;
p->next = allphs->ci_phmm[p->ci];
allphs->ci_phmm[p->ci] = p;
n_phmm++;
}
pid2phmm[pid] = p;
}
/* Fill out bitvecs of each PHMM node, alloc continuous memory chunk for context bitvectors */
lrc_size = bitvec_size(bin_mdef_n_ciphone(mdef));
lc = ckd_calloc(n_phmm * 2 * lrc_size, sizeof(bitvec_t));
rc = lc + (n_phmm * lrc_size);
for (ci = 0; ci < mdef->n_ciphone; ci++) {
for (p = allphs->ci_phmm[ci]; p; p = p->next) {
p->lc = lc;
lc += lrc_size;
p->rc = rc;
rc += lrc_size;
}
}
/* Fill out lc and rc bitmaps (remember to map all fillers to each other!!) */
filler =
(s3cipid_t *) ckd_calloc(bin_mdef_n_ciphone(mdef) + 1,
sizeof(s3cipid_t));
/* Connect fillers */
i = 0;
for (ci = 0; ci < bin_mdef_n_ciphone(mdef); ci++) {
p = pid2phmm[ci];
bitvec_set_all(p->lc, bin_mdef_n_ciphone(mdef));
bitvec_set_all(p->rc, bin_mdef_n_ciphone(mdef));
if (mdef->phone[ci].info.ci.filler) {
filler[i++] = ci;
}
}
filler[i] = BAD_S3CIPID;
/* Loop over cdphones only if ci_only is not set */
for (pid = bin_mdef_n_ciphone(mdef); pid < nphone;
pid++) {
p = pid2phmm[pid];
if (mdef->phone[mdef->phone[pid].info.cd.ctx[1]].info.ci.filler) {
for (i = 0; IS_S3CIPID(filler[i]); i++)
bitvec_set(p->lc, filler[i]);
}
else
bitvec_set(p->lc, mdef->phone[pid].info.cd.ctx[1]);
if (mdef->phone[mdef->phone[pid].info.cd.ctx[2]].info.ci.filler) {
for (i = 0; IS_S3CIPID(filler[i]); i++)
bitvec_set(p->rc, filler[i]);
}
else
bitvec_set(p->rc, mdef->phone[pid].info.cd.ctx[2]);
}
ckd_free(pid2phmm);
ckd_free(filler);
/* Create links between PHMM nodes */
n_link = phmm_link(allphs);
E_INFO("%d nodes, %d links\n", n_phmm, n_link);
return 0;
}
/**
* Build HMM network for one utterance of fwdflat search.
*/
static void
build_fwdflat_chan(ngram_search_t *ngs)
{
int32 i, wid, p;
root_chan_t *rhmm;
chan_t *hmm, *prevhmm;
dict_t *dict;
dict2pid_t *d2p;
dict = ps_search_dict(ngs);
d2p = ps_search_dict2pid(ngs);
/* Build word HMMs for each word in the lattice. */
for (i = 0; ngs->fwdflat_wordlist[i] >= 0; i++) {
wid = ngs->fwdflat_wordlist[i];
/* Single-phone words are permanently allocated */
if (dict_is_single_phone(dict, wid))
continue;
assert(ngs->word_chan[wid] == NULL);
/* Multiplex root HMM for first phone (one root per word, flat
* lexicon). diphone is irrelevant here, for the time being,
* at least. */
rhmm = listelem_malloc(ngs->root_chan_alloc);
rhmm->ci2phone = dict_second_phone(dict, wid);
rhmm->ciphone = dict_first_phone(dict, wid);
rhmm->next = NULL;
hmm_init(ngs->hmmctx, &rhmm->hmm, TRUE,
bin_mdef_pid2ssid(ps_search_acmod(ngs)->mdef, rhmm->ciphone),
bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef, rhmm->ciphone));
/* HMMs for word-internal phones */
prevhmm = NULL;
for (p = 1; p < dict_pronlen(dict, wid) - 1; p++) {
hmm = listelem_malloc(ngs->chan_alloc);
hmm->ciphone = dict_pron(dict, wid, p);
hmm->info.rc_id = (p == dict_pronlen(dict, wid) - 1) ? 0 : -1;
hmm->next = NULL;
hmm_init(ngs->hmmctx, &hmm->hmm, FALSE,
dict2pid_internal(d2p,wid,p),
bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef, hmm->ciphone));
if (prevhmm)
prevhmm->next = hmm;
else
rhmm->next = hmm;
prevhmm = hmm;
}
/* Right-context phones */
ngram_search_alloc_all_rc(ngs, wid);
/* Link in just allocated right-context phones */
if (prevhmm)
prevhmm->next = ngs->word_chan[wid];
else
rhmm->next = ngs->word_chan[wid];
ngs->word_chan[wid] = (chan_t *) rhmm;
}
}
/*
* Add the word emitted by the given transition (fsglink) to the given lextree
* (rooted at root), and return the new lextree root. (There may actually be
* several root nodes, maintained in a linked list via fsg_pnode_t.sibling.
* "root" is the head of this list.)
* lclist, rclist: sets of left and right context phones for this link.
* alloc_head: head of a linear list of all allocated pnodes for the parent
* FSG state, kept elsewhere and updated by this routine.
*/
static fsg_pnode_t *
psubtree_add_trans(fsg_lextree_t *lextree,
fsg_pnode_t * root,
fsg_glist_linklist_t **curglist,
fsg_link_t * fsglink,
int16 *lclist, int16 *rclist,
fsg_pnode_t ** alloc_head)
{
int32 silcipid; /* Silence CI phone ID */
int32 pronlen; /* Pronunciation length */
int32 wid; /* FSG (not dictionary!!) word ID */
int32 dictwid; /* Dictionary (not FSG!!) word ID */
int32 ssid; /* Senone Sequence ID */
gnode_t *gn;
fsg_pnode_t *pnode, *pred, *head;
int32 n_ci, p, lc, rc;
glist_t lc_pnodelist; /* Temp pnodes list for different left contexts */
glist_t rc_pnodelist; /* Temp pnodes list for different right contexts */
int32 i, j;
silcipid = bin_mdef_silphone(lextree->mdef);
n_ci = bin_mdef_n_ciphone(lextree->mdef);
wid = fsg_link_wid(fsglink);
assert(wid >= 0); /* Cannot be a null transition */
dictwid = dict_wordid(lextree->dict,
fsg_model_word_str(lextree->fsg, wid));
pronlen = dict_pronlen(lextree->dict, dictwid);
assert(pronlen >= 1);
assert(lclist[0] >= 0); /* At least one phonetic context provided */
assert(rclist[0] >= 0);
head = *alloc_head;
pred = NULL;
if (pronlen == 1) { /* Single-phone word */
int ci = dict_first_phone(lextree->dict, dictwid);
/* Only non-filler words are mpx */
if (dict_filler_word(lextree->dict, dictwid)) {
/*
* Left diphone ID for single-phone words already assumes SIL is right
* context; only left contexts need to be handled.
*/
lc_pnodelist = NULL;
for (i = 0; lclist[i] >= 0; i++) {
lc = lclist[i];
ssid = dict2pid_lrdiph_rc(lextree->d2p, ci, lc, silcipid);
/* Check if this ssid already allocated for some other context */
for (gn = lc_pnodelist; gn; gn = gnode_next(gn)) {
pnode = (fsg_pnode_t *) gnode_ptr(gn);
if (hmm_nonmpx_ssid(&pnode->hmm) == ssid) {
/* already allocated; share it for this context phone */
fsg_pnode_add_ctxt(pnode, lc);
break;
}
}
if (!gn) { /* ssid not already allocated */
pnode =
(fsg_pnode_t *) ckd_calloc(1, sizeof(fsg_pnode_t));
pnode->ctx = lextree->ctx;
pnode->next.fsglink = fsglink;
pnode->logs2prob =
fsg_link_logs2prob(fsglink) + lextree->wip + lextree->pip;
pnode->ci_ext = dict_first_phone(lextree->dict, dictwid);
pnode->ppos = 0;
pnode->leaf = TRUE;
pnode->sibling = root; /* All root nodes linked together */
fsg_pnode_add_ctxt(pnode, lc); /* Initially zeroed by calloc above */
pnode->alloc_next = head;
head = pnode;
root = pnode;
hmm_init(lextree->ctx, &pnode->hmm, FALSE, ssid, pnode->ci_ext);
lc_pnodelist =
glist_add_ptr(lc_pnodelist, (void *) pnode);
}
}
glist_free(lc_pnodelist);
}
else { /* Filler word; no context modelled */
ssid = bin_mdef_pid2ssid(lextree->mdef, ci); /* probably the same... */
pnode = (fsg_pnode_t *) ckd_calloc(1, sizeof(fsg_pnode_t));
pnode->ctx = lextree->ctx;
pnode->next.fsglink = fsglink;
pnode->logs2prob = fsg_link_logs2prob(fsglink) + lextree->wip + lextree->pip;
pnode->ci_ext = silcipid; /* Presents SIL as context to neighbors */
pnode->ppos = 0;
pnode->leaf = TRUE;
pnode->sibling = root;
fsg_pnode_add_all_ctxt(&(pnode->ctxt));
pnode->alloc_next = head;
head = pnode;
root = pnode;
hmm_init(lextree->ctx, &pnode->hmm, FALSE, ssid, pnode->ci_ext);
}
}
else { /* Multi-phone word */
fsg_pnode_t **ssid_pnode_map; /* Temp array of ssid->pnode mapping */
ssid_pnode_map =
(fsg_pnode_t **) ckd_calloc(n_ci, sizeof(fsg_pnode_t *));
lc_pnodelist = NULL;
rc_pnodelist = NULL;
for (p = 0; p < pronlen; p++) {
int ci = dict_pron(lextree->dict, dictwid, p);
if (p == 0) { /* Root phone, handle required left contexts */
/* Find if we already have an lc_pnodelist for the first phone of this word */
fsg_glist_linklist_t *predglist=*curglist;
fsg_glist_linklist_t *glist=*curglist;
rc = dict_pron(lextree->dict, dictwid, 1);
while (glist && glist->glist && glist->ci != ci && glist->rc != rc){
glist = glist->next;
}
if (glist && glist->ci == ci && glist->rc == rc && glist->glist) {
/* We've found a valid glist. Hook to it and move to next phoneme */
lc_pnodelist = glist->glist;
/* Set the predecessor node for the future tree first */
pred = (fsg_pnode_t *) gnode_ptr(lc_pnodelist);
continue;
}
else {
/* Two cases that can bring us here
* a. glist == NULL, i.e. end of current list. Create new entry.
* b. glist->glist == NULL, i.e. first entry into list.
*/
if (!glist) { /* Case a; reduce it to case b by allocing glist */
glist = (fsg_glist_linklist_t*) ckd_calloc(1, sizeof(fsg_glist_linklist_t));
glist->next = predglist;
*curglist = glist;
}
glist->ci = ci;
glist->rc = rc;
glist->lc = -1;
lc_pnodelist = glist->glist = NULL; /* Gets created below */
}
for (i = 0; lclist[i] >= 0; i++) {
lc = lclist[i];
ssid = dict2pid_ldiph_lc(lextree->d2p, ci, rc, lc);
/* Compression is not done by d2p, so we do it
* here. This might be slow, but it might not
* be... we'll see. */
pnode = ssid_pnode_map[0];
for (j = 0; j < n_ci && ssid_pnode_map[j] != NULL; ++j) {
pnode = ssid_pnode_map[j];
if (hmm_nonmpx_ssid(&pnode->hmm) == ssid)
break;
}
assert(j < n_ci);
if (!pnode) { /* Allocate pnode for this new ssid */
pnode =
(fsg_pnode_t *) ckd_calloc(1,
sizeof
(fsg_pnode_t));
pnode->ctx = lextree->ctx;
/* This bit is tricky! For now we'll put the prob in the final link only */
/* pnode->logs2prob = fsg_link_logs2prob(fsglink)
+ lextree->wip + lextree->pip; */
pnode->logs2prob = lextree->wip + lextree->pip;
pnode->ci_ext = dict_first_phone(lextree->dict, dictwid);
pnode->ppos = 0;
pnode->leaf = FALSE;
pnode->sibling = root; /* All root nodes linked together */
pnode->alloc_next = head;
head = pnode;
root = pnode;
hmm_init(lextree->ctx, &pnode->hmm, FALSE, ssid, pnode->ci_ext);
lc_pnodelist =
glist_add_ptr(lc_pnodelist, (void *) pnode);
ssid_pnode_map[j] = pnode;
}
fsg_pnode_add_ctxt(pnode, lc);
}
/* Put the lc_pnodelist back into glist */
glist->glist = lc_pnodelist;
/* The predecessor node for the future tree is the root */
pred = root;
}
else if (p != pronlen - 1) { /* Word internal phone */
fsg_pnode_t *pnodeyoungest;
ssid = dict2pid_internal(lextree->d2p, dictwid, p);
/* First check if we already have this ssid in our tree */
pnode = pred->next.succ;
pnodeyoungest = pnode; /* The youngest sibling */
while (pnode && (hmm_nonmpx_ssid(&pnode->hmm) != ssid || pnode->leaf)) {
pnode = pnode->sibling;
}
if (pnode && (hmm_nonmpx_ssid(&pnode->hmm) == ssid && !pnode->leaf)) {
/* Found the ssid; go to next phoneme */
pred = pnode;
continue;
}
/* pnode not found, allocate it */
pnode = (fsg_pnode_t *) ckd_calloc(1, sizeof(fsg_pnode_t));
pnode->ctx = lextree->ctx;
pnode->logs2prob = lextree->pip;
pnode->ci_ext = dict_pron(lextree->dict, dictwid, p);
pnode->ppos = p;
pnode->leaf = FALSE;
pnode->sibling = pnodeyoungest; /* May be NULL */
if (p == 1) { /* Predecessor = set of root nodes for left ctxts */
for (gn = lc_pnodelist; gn; gn = gnode_next(gn)) {
pred = (fsg_pnode_t *) gnode_ptr(gn);
pred->next.succ = pnode;
}
}
else { /* Predecessor = word internal node */
pred->next.succ = pnode;
}
pnode->alloc_next = head;
head = pnode;
hmm_init(lextree->ctx, &pnode->hmm, FALSE, ssid, pnode->ci_ext);
pred = pnode;
}
else { /* Leaf phone, handle required right contexts */
/* Note, leaf phones are not part of the tree */
xwdssid_t *rssid;
memset((void *) ssid_pnode_map, 0,
n_ci * sizeof(fsg_pnode_t *));
lc = dict_pron(lextree->dict, dictwid, p-1);
rssid = dict2pid_rssid(lextree->d2p, ci, lc);
for (i = 0; rclist[i] >= 0; i++) {
rc = rclist[i];
j = rssid->cimap[rc];
ssid = rssid->ssid[j];
pnode = ssid_pnode_map[j];
if (!pnode) { /* Allocate pnode for this new ssid */
pnode =
(fsg_pnode_t *) ckd_calloc(1,
sizeof
(fsg_pnode_t));
pnode->ctx = lextree->ctx;
/* We are plugging the word prob here. Ugly */
/* pnode->logs2prob = lextree->pip; */
pnode->logs2prob = fsg_link_logs2prob(fsglink) + lextree->pip;
pnode->ci_ext = dict_pron(lextree->dict, dictwid, p);
pnode->ppos = p;
pnode->leaf = TRUE;
pnode->sibling = rc_pnodelist ?
(fsg_pnode_t *) gnode_ptr(rc_pnodelist) : NULL;
pnode->next.fsglink = fsglink;
pnode->alloc_next = head;
head = pnode;
hmm_init(lextree->ctx, &pnode->hmm, FALSE, ssid, pnode->ci_ext);
rc_pnodelist =
glist_add_ptr(rc_pnodelist, (void *) pnode);
ssid_pnode_map[j] = pnode;
}
else {
assert(hmm_nonmpx_ssid(&pnode->hmm) == ssid);
}
fsg_pnode_add_ctxt(pnode, rc);
}
if (p == 1) { /* Predecessor = set of root nodes for left ctxts */
for (gn = lc_pnodelist; gn; gn = gnode_next(gn)) {
pred = (fsg_pnode_t *) gnode_ptr(gn);
if (!pred->next.succ)
pred->next.succ = (fsg_pnode_t *) gnode_ptr(rc_pnodelist);
else {
/* Link to the end of the sibling chain */
fsg_pnode_t *succ = pred->next.succ;
while (succ->sibling) succ = succ->sibling;
succ->sibling = (fsg_pnode_t*) gnode_ptr(rc_pnodelist);
/* Since all entries of lc_pnodelist point
to the same array, sufficient to update it once */
break;
}
}
}
else { /* Predecessor = word internal node */
if (!pred->next.succ)
pred->next.succ = (fsg_pnode_t *) gnode_ptr(rc_pnodelist);
else {
/* Link to the end of the sibling chain */
fsg_pnode_t *succ = pred->next.succ;
while (succ->sibling) succ = succ->sibling;
succ->sibling = (fsg_pnode_t *) gnode_ptr(rc_pnodelist);
}
}
}
}
ckd_free((void *) ssid_pnode_map);
/* glist_free(lc_pnodelist); Nope; this gets freed outside */
glist_free(rc_pnodelist);
}
*alloc_head = head;
return root;
}
static void init_data(args_t *args)
{
args->prev_rid = args->skip_rid = -1;
args->hdr = args->files->readers[0].header;
if ( !args->sample )
{
if ( bcf_hdr_nsamples(args->hdr)>1 ) error("Missing the option -s, --sample\n");
args->sample = strdup(args->hdr->samples[0]);
}
if ( !bcf_hdr_nsamples(args->hdr) ) error("No samples in the VCF?\n");
// Set samples
kstring_t str = {0,0,0};
if ( args->estimate_AF && strcmp("-",args->estimate_AF) )
{
int i, n;
char **smpls = hts_readlist(args->estimate_AF, 1, &n);
// Make sure the query sample is included
for (i=0; i<n; i++)
if ( !strcmp(args->sample,smpls[i]) ) break;
// Add the query sample if not present
if ( i!=n ) kputs(args->sample, &str);
for (i=0; i<n; i++)
{
if ( str.l ) kputc(',', &str);
kputs(smpls[i], &str);
free(smpls[i]);
}
free(smpls);
}
else if ( !args->estimate_AF )
kputs(args->sample, &str);
if ( str.l )
{
int ret = bcf_hdr_set_samples(args->hdr, str.s, 0);
if ( ret<0 ) error("Error parsing the list of samples: %s\n", str.s);
else if ( ret>0 ) error("The %d-th sample not found in the VCF\n", ret);
}
if ( args->af_tag )
if ( !bcf_hdr_idinfo_exists(args->hdr,BCF_HL_INFO,bcf_hdr_id2int(args->hdr,BCF_DT_ID,args->af_tag)) )
error("No such INFO tag in the VCF: %s\n", args->af_tag);
args->nsmpl = bcf_hdr_nsamples(args->hdr);
args->ismpl = bcf_hdr_id2int(args->hdr, BCF_DT_SAMPLE, args->sample);
free(str.s);
int i;
for (i=0; i<256; i++) args->pl2p[i] = pow(10., -i/10.);
// Init transition matrix and HMM
double tprob[4];
MAT(tprob,2,STATE_HW,STATE_HW) = 1 - args->t2AZ;
MAT(tprob,2,STATE_HW,STATE_AZ) = args->t2HW;
MAT(tprob,2,STATE_AZ,STATE_HW) = args->t2AZ;
MAT(tprob,2,STATE_AZ,STATE_AZ) = 1 - args->t2HW;
if ( args->genmap_fname )
{
args->hmm = hmm_init(2, tprob, 0);
hmm_set_tprob_func(args->hmm, set_tprob_genmap, args);
}
else if ( args->rec_rate > 0 )
{
args->hmm = hmm_init(2, tprob, 0);
hmm_set_tprob_func(args->hmm, set_tprob_recrate, args);
}
else
args->hmm = hmm_init(2, tprob, 10000);
// print header
printf("# This file was produced by: bcftools roh(%s+htslib-%s)\n", bcftools_version(),hts_version());
printf("# The command line was:\tbcftools %s", args->argv[0]);
for (i=1; i<args->argc; i++)
printf(" %s",args->argv[i]);
printf("\n#\n");
printf("# [1]Chromosome\t[2]Position\t[3]State (0:HW, 1:AZ)\t[4]Quality\n");
}
static void init_hmm_trio(args_t *args)
{
int i,j;
args->nstates = 8;
args->tprob = (double*) malloc(sizeof(double)*args->nstates*args->nstates);
for (i=0; i<args->nstates; i++)
for (j=0; j<args->nstates; j++) hap_switch[i][j] = 0;
hap_switch[TRIO_AD][TRIO_AC] = SW_FATHER;
hap_switch[TRIO_BC][TRIO_AC] = SW_MOTHER;
hap_switch[TRIO_BD][TRIO_AC] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_AC][TRIO_AD] = SW_FATHER;
hap_switch[TRIO_BC][TRIO_AD] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_BD][TRIO_AD] = SW_MOTHER;
hap_switch[TRIO_AC][TRIO_BC] = SW_MOTHER;
hap_switch[TRIO_AD][TRIO_BC] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_BD][TRIO_BC] = SW_FATHER;
hap_switch[TRIO_AC][TRIO_BD] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_AD][TRIO_BD] = SW_MOTHER;
hap_switch[TRIO_BC][TRIO_BD] = SW_FATHER;
hap_switch[TRIO_DA][TRIO_CA] = SW_FATHER;
hap_switch[TRIO_CB][TRIO_CA] = SW_MOTHER;
hap_switch[TRIO_DB][TRIO_CA] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_CA][TRIO_DA] = SW_FATHER;
hap_switch[TRIO_CB][TRIO_DA] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_DB][TRIO_DA] = SW_MOTHER;
hap_switch[TRIO_CA][TRIO_CB] = SW_MOTHER;
hap_switch[TRIO_DA][TRIO_CB] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_DB][TRIO_CB] = SW_FATHER;
hap_switch[TRIO_CA][TRIO_DB] = SW_MOTHER | SW_FATHER;
hap_switch[TRIO_DA][TRIO_DB] = SW_MOTHER;
hap_switch[TRIO_CB][TRIO_DB] = SW_FATHER;
for (i=0; i<args->nstates; i++)
{
for (j=0; j<args->nstates; j++)
{
if ( !hap_switch[i][j] ) MAT(args->tprob,args->nstates,i,j) = 0;
else
{
MAT(args->tprob,args->nstates,i,j) = 1;
if ( hap_switch[i][j] & SW_MOTHER ) MAT(args->tprob,args->nstates,i,j) *= args->pij;
if ( hap_switch[i][j] & SW_FATHER ) MAT(args->tprob,args->nstates,i,j) *= args->pij;
}
}
}
for (i=0; i<args->nstates; i++)
{
double sum = 0;
for (j=0; j<args->nstates; j++)
{
if ( i!=j ) sum += MAT(args->tprob,args->nstates,i,j);
}
MAT(args->tprob,args->nstates,i,i) = 1 - sum;
}
#if 0
for (i=0; i<args->nstates; i++)
{
for (j=0; j<args->nstates; j++)
fprintf(stderr,"\t%d",hap_switch[j][i]);
fprintf(stderr,"\n");
}
for (i=0; i<args->nstates; i++)
{
for (j=0; j<args->nstates; j++)
fprintf(stderr,"\t%e",MAT(args->tprob,args->nstates,j,i));
fprintf(stderr,"\n");
}
#endif
args->hmm = hmm_init(args->nstates, args->tprob, 10000);
}