float64
best_q(float32 ****mixw,
/* ADDITION FOR CONTINUOUS_TREES 21 May 98 */
float32 ****means,
float32 ****vars,
uint32 *veclen,
/* END ADDITION FOR CONTINUOUS_TREES */
uint32 n_model,
uint32 n_state,
uint32 n_stream,
uint32 n_density,
float32 *stwt,
uint32 **dfeat,
uint32 n_dfeat,
quest_t *all_q,
uint32 n_all_q,
pset_t *pset,
uint32 *id,
uint32 n_id,
float32 ***dist,
/* ADDITION FOR CONTINUOUS_TREES 21 May 98 */
float64 node_wt_ent, /* Weighted entropy of node */
/* END ADDITION FOR CONTINUOUS_TREES */
quest_t **out_best_q)
{
float32 ***yes_dist;
/* ADDITION FOR CONTINUOUS_TREES */
float32 ***yes_means=0;
float32 ***yes_vars=0;
float32 varfloor=0;
float64 y_ent;
/* END ADDITION FOR CONTINUOUS_TREES */
float64 yes_dnom, yes_norm;
uint32 *yes_id;
float32 ***no_dist;
/* ADDITION FOR CONTINUOUS_TREES */
float32 ***no_means=0;
float32 ***no_vars=0;
float64 n_ent;
/* END ADDITION FOR CONTINUOUS_TREES */
float64 no_dnom, no_norm;
uint32 *no_id;
uint32 n_yes, n_b_yes = 0;
uint32 n_no, n_b_no = 0;
uint32 i, j, k, q, b_q=0, s;
uint32 ii;
float64 einc, b_einc = -1.0e+50;
/* ADDITION FOR CONTINUOUS_TREES; 20 May 98 */
char* type;
uint32 continuous, sumveclen=0;
type = (char *)cmd_ln_access("-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 = *(float32 *)cmd_ln_access("-varfloor");
/* Allocating for sumveclen is overallocation, but it eases coding */
for (ii=0,sumveclen=0; ii<n_stream; ii++) sumveclen += veclen[ii];
yes_means = (float32 ***)ckd_calloc_3d(n_state,n_stream,sumveclen,sizeof(float32));
yes_vars = (float32 ***)ckd_calloc_3d(n_state,n_stream,sumveclen,sizeof(float32));
no_means = (float32 ***)ckd_calloc_3d(n_state,n_stream,sumveclen,sizeof(float32));
no_vars = (float32 ***)ckd_calloc_3d(n_state,n_stream,sumveclen,sizeof(float32));
}
/* END ADDITIONS FOR CONTINUOUS_TREES */
n_yes = n_no = 0;
yes_dist = (float32 ***)ckd_calloc_3d(n_state, n_stream, n_density, sizeof(float32));
no_dist = (float32 ***)ckd_calloc_3d(n_state, n_stream, n_density, sizeof(float32));
for (q = 0; q < n_all_q; q++) {
memset(&yes_dist[0][0][0], 0, sizeof(float32) * n_state * n_stream * n_density);
memset(&no_dist[0][0][0], 0, sizeof(float32) * n_state * n_stream * n_density);
/* ADDITION FOR CONTINUOUS_TREES; If continuous hmm initialize means and vars to zero */
if (continuous == 1) {
memset(&yes_means[0][0][0], 0, sizeof(float32) * n_state * n_stream * sumveclen);
memset(&yes_vars[0][0][0], 0, sizeof(float32) * n_state * n_stream * sumveclen);
memset(&no_means[0][0][0], 0, sizeof(float32) * n_state * n_stream * sumveclen);
memset(&no_vars[0][0][0], 0, sizeof(float32) * n_state * n_stream * sumveclen);
}
/* END ADDITION FOR CONTINUOUS_TREES */
n_yes = n_no = 0;
for (ii = 0; ii < n_id; ii++) {
i = id[ii];
if (eval_quest(&all_q[q], dfeat[i], n_dfeat)) {
for (s = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
for (k = 0; k < n_density; k++) {
yes_dist[s][j][k] += mixw[i][s][j][k];
}
}
}
/* MODIFICATION FOR CONTINUOUS_TREES: ADDITIONS FOR CONTINUOUS CASE */
if (continuous == 1) {
for (s = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
for (k = 0; k < veclen[j]; k++) {
yes_means[s][j][k] += mixw[i][s][j][0] * means[i][s][j][k];
yes_vars[s][j][k] += mixw[i][s][j][0] * (vars[i][s][j][k] + means[i][s][j][k]*means[i][s][j][k]);
}
}
}
}
/* END MODIFICATION FOR CONTINUOUS_TREES */
++n_yes;
}
else {
for (s = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
for (k = 0; k < n_density; k++) {
no_dist[s][j][k] += mixw[i][s][j][k];
}
}
}
/* MODIFICATION FOR CONTINUOUS_TREES: ADDITIONS FOR CONTINUOUS CASE */
if (continuous == 1) {
for (s = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
for (k = 0; k < veclen[j]; k++) {
no_means[s][j][k] += mixw[i][s][j][0] * means[i][s][j][k];
no_vars[s][j][k] += mixw[i][s][j][0] * (vars[i][s][j][k] + means[i][s][j][k]*means[i][s][j][k]);
}
}
}
}
/* END MODIFICATION FOR CONTINUOUS_TREES */
++n_no;
}
}
if ((n_yes == 0) || (n_no == 0)) {
/* no split. All satisfy or all don't satisfy */
continue;
}
for (s = 0, einc = 0; s < n_state; s++) {
for (k = 0, yes_dnom = 0; k < n_density; k++) {
yes_dnom += yes_dist[s][0][k];
}
if (yes_dnom == 0)
break;
yes_norm = 1.0 / yes_dnom;
for (j = 0; j < n_stream; j++) {
for (k = 0; k < n_density; k++) {
yes_dist[s][j][k] *= yes_norm;
}
}
for (k = 0, no_dnom = 0; k < n_density; k++) {
no_dnom += no_dist[s][0][k];
}
if (no_dnom == 0)
break;
no_norm = 1.0 / no_dnom;
for (j = 0; j < n_stream; j++) {
for (k = 0; k < n_density; k++) {
no_dist[s][j][k] *= no_norm;
}
}
/* MODIFICATION FOR CONTINUOUS_TREES: Do appropriate operations for discrete and
continuous */
if (continuous == 1) {
y_ent = 0;
n_ent = 0;
for (j = 0; j < n_stream; j++) {
if (yes_dnom != 0) {
for (k = 0; k < veclen[j]; k++) {
yes_means[s][j][k] *= yes_norm;
yes_vars[s][j][k] = yes_vars[s][j][k]*yes_norm -
yes_means[s][j][k]*yes_means[s][j][k];
if (yes_vars[s][j][k] < varfloor) yes_vars[s][j][k] = varfloor;
}
}
if (no_dnom != 0) {
for (k = 0; k < veclen[j]; k++) {
no_means[s][j][k] *= no_norm;
no_vars[s][j][k] = no_vars[s][j][k]*no_norm -
no_means[s][j][k]*no_means[s][j][k];
if (no_vars[s][j][k] < varfloor) no_vars[s][j][k] = varfloor;
}
}
y_ent += yes_dnom * ent_cont(yes_means[s][j],yes_vars[s][j],veclen[j]);
n_ent += no_dnom * ent_cont(no_means[s][j],no_vars[s][j],veclen[j]);
}
einc += (float64)stwt[s] * (y_ent + n_ent);
}
else {
einc += (float64)stwt[s] * wt_ent_inc(yes_dist[s], yes_dnom,
no_dist[s], no_dnom,
dist[s], n_stream, n_density);
}
}
/* END MODIFICATION FOR CONTINUOUS_TREES */
/* ADDITION FOR CONTINUOUS_TREES; In current code this is true only for continous HMM */
if (continuous == 1) {
einc -= node_wt_ent;
}
/* END ADDITION FOR CONTINUOUS_TREES */
if (s < n_state) {
/* Ended iteration over states prematurely; assume 'bad' question */
continue;
}
if (einc > b_einc) {
b_einc = einc;
b_q = q;
n_b_yes = n_yes;
n_b_no = n_no;
}
}
if ((n_b_yes == 0) || (n_b_no == 0)) {
/* No best question */
*out_best_q = NULL;
return 0;
}
yes_id = (uint32 *)ckd_calloc(n_b_yes, sizeof(uint32));
no_id = (uint32 *)ckd_calloc(n_b_no, sizeof(uint32));
memset(&yes_dist[0][0][0], 0, sizeof(float32) * n_state * n_stream * n_density);
memset(&no_dist[0][0][0], 0, sizeof(float32) * n_state * n_stream * n_density);
n_yes = n_no = 0;
for (ii = 0; ii < n_id; ii++) {
i = id[ii];
if (eval_quest(&all_q[b_q], dfeat[i], n_dfeat)) {
for (s = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
for (k = 0; k < n_density; k++) {
yes_dist[s][j][k] += mixw[i][s][j][k];
}
}
}
yes_id[n_yes] = i;
++n_yes;
}
else {
for (s = 0; s < n_state; s++) {
for (j = 0; j < n_stream; j++) {
for (k = 0; k < n_density; k++) {
no_dist[s][j][k] += mixw[i][s][j][k];
}
}
}
no_id[n_no] = i;
++n_no;
}
}
ckd_free_3d((void ***)yes_dist);
ckd_free((void *)yes_id);
ckd_free_3d((void ***)no_dist);
ckd_free((void *)no_id);
/* ADDITION FOR CONTINUOUS_TREES */
if (continuous == 1) {
ckd_free_3d((void ***)yes_means);
ckd_free_3d((void ***)yes_vars);
ckd_free_3d((void ***)no_means);
ckd_free_3d((void ***)no_vars);
}
/* END ADDITION FOR CONTINUOUS_TREES */
*out_best_q = &all_q[b_q];
return b_einc;
}
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;
}
