hmm_t *hmm_init(int nstates, double *tprob, int ntprob)
{
hmm_t *hmm = (hmm_t*) calloc(1,sizeof(hmm_t));
hmm->nstates = nstates;
hmm->curr_tprob = (double*) malloc(sizeof(double)*nstates*nstates);
hmm->tmp = (double*) malloc(sizeof(double)*nstates*nstates);
hmm_set_tprob(hmm, tprob, ntprob);
return hmm;
}
static void flush_viterbi(args_t *args)
{
int i,j;
if ( !args->nsites ) return;
if ( !args->vi_training )
{
// single viterbi pass, one chromsome
hmm_run_viterbi(args->hmm, args->nsites, args->eprob, args->sites);
hmm_run_fwd_bwd(args->hmm, args->nsites, args->eprob, args->sites);
double *fwd = hmm_get_fwd_bwd_prob(args->hmm);
const char *chr = bcf_hdr_id2name(args->hdr,args->prev_rid);
uint8_t *vpath = hmm_get_viterbi_path(args->hmm);
for (i=0; i<args->nsites; i++)
{
int state = vpath[i*2]==STATE_AZ ? 1 : 0;
double *pval = fwd + i*2;
printf("%s\t%d\t%d\t%.1f\n", chr,args->sites[i]+1, state, phred_score(1.0-pval[state]));
}
return;
}
// viterbi training, multiple chromosomes
double t2az_prev, t2hw_prev;
double deltaz, delthw;
int niter = 0;
do
{
double *tprob_arr = hmm_get_tprob(args->hmm);
t2az_prev = MAT(tprob_arr,2,1,0); //args->t2AZ;
t2hw_prev = MAT(tprob_arr,2,0,1); //args->t2HW;
double tcounts[] = { 0,0,0,0 };
for (i=0; i<args->nrids; i++)
{
// run viterbi for each chromosomes. eprob and sites contain
// multiple chromosomes, rid_offs mark the boundaries
int ioff = args->rid_offs[i];
int nsites = (i+1==args->nrids ? args->nsites : args->rid_offs[i+1]) - ioff;
hmm_run_viterbi(args->hmm, nsites, args->eprob+ioff*2, args->sites+ioff);
// what transitions were observed: add to the total counts
uint8_t *vpath = hmm_get_viterbi_path(args->hmm);
for (j=1; j<nsites; j++)
{
// count the number of transitions
int prev_state = vpath[2*(j-1)];
int curr_state = vpath[2*j];
MAT(tcounts,2,curr_state,prev_state) += 1;
}
}
// update the transition matrix tprob
for (i=0; i<2; i++)
{
int n = 0;
for (j=0; j<2; j++) n += MAT(tcounts,2,i,j);
if ( !n) error("fixme: state %d not observed\n", i+1);
for (j=0; j<2; j++) MAT(tcounts,2,i,j) /= n;
}
if ( args->genmap_fname || args->rec_rate > 0 )
hmm_set_tprob(args->hmm, tcounts, 0);
else
hmm_set_tprob(args->hmm, tcounts, 10000);
tprob_arr = hmm_get_tprob(args->hmm);
deltaz = fabs(MAT(tprob_arr,2,1,0)-t2az_prev);
delthw = fabs(MAT(tprob_arr,2,0,1)-t2hw_prev);
niter++;
fprintf(pysamerr,"%d: %f %f\n", niter,deltaz,delthw);
}
while ( deltaz > 0.0 || delthw > 0.0 );
fprintf(pysamerr, "Viterbi training converged in %d iterations to", niter);
double *tprob_arr = hmm_get_tprob(args->hmm);
for (i=0; i<2; i++) for (j=0; j<2; j++) fprintf(pysamerr, " %f", MAT(tprob_arr,2,i,j));
fprintf(pysamerr, "\n");
// output the results
for (i=0; i<args->nrids; i++)
{
int ioff = args->rid_offs[i];
int nsites = (i+1==args->nrids ? args->nsites : args->rid_offs[i+1]) - ioff;
hmm_run_viterbi(args->hmm, nsites, args->eprob+ioff*2, args->sites+ioff);
uint8_t *vpath = hmm_get_viterbi_path(args->hmm);
const char *chr = bcf_hdr_id2name(args->hdr,args->rids[i]);
for (j=0; j<nsites; j++)
{
printf("%s\t%d\t%d\t..\n", chr,args->sites[ioff+j]+1,vpath[j*2]==STATE_AZ ? 1 : 0);
}
}
}
