void Texture::setUp(dx::IDirect3DTexturePtr tex)
{
ptr_ = tex;
org_width_ = size().x;
org_height_ = size().y;
org_format_ = format();
org_miplevel_ = maxlevel();
}
//creates the ngramtable on demand from the sublm tables
int mixture::get(ngram& ng,int n,int lev)
{
if (usefulltable)
{
return ngramtable::get(ng,n,lev);
}
//free current tree
resetngramtable();
//get 1-word prefix from ng
ngram ug(dict,1);
*ug.wordp(1)=*ng.wordp(ng.size);
//local ngram to upload entries
ngram locng(dict,maxlevel());
//allocate subtrees from sublm
for (int i=0; i<numslm; i++) {
ngram subug(sublm[i]->dict,1);
subug.trans(ug);
if (sublm[i]->get(subug,1,1)) {
ngram subng(sublm[i]->dict,maxlevel());
*subng.wordp(maxlevel())=*subug.wordp(1);
sublm[i]->scan(subug.link,subug.info,1,subng,INIT,maxlevel());
while(sublm[i]->scan(subug.link,subug.info,1,subng,CONT,maxlevel())) {
locng.trans(subng);
put(locng);
}
}
}
return ngramtable::get(ng,n,lev);
}
int mshiftbeta::discount(ngram ng_,int size,double& fstar,double& lambda, int cv)
{
ngram ng(dict);
ng.trans(ng_);
//cout << "size :" << size << " " << ng <<"\n";
if (size > 1) {
ngram history=ng;
//singleton pruning only on real counts!!
if (ng.ckhisto(size) && get(history,size,size-1) && (history.freq > cv) &&
((size < 3) || ((history.freq-cv) > prunethresh ))) { // no history pruning with corrected counts!
int suc[3];
suc[0]=succ1(history.link);
suc[1]=succ2(history.link);
suc[2]=history.succ-suc[0]-suc[1];
if (get(ng,size,size) &&
(!prunesingletons() || mfreq(ng,size)>1 || size<3) &&
(!prunetopsingletons() || mfreq(ng,size)>1 || size<maxlevel())) {
ng.freq=mfreq(ng,size);
cv=(cv>ng.freq)?ng.freq:cv;
if (ng.freq>cv) {
double b=(ng.freq-cv>=3?beta[2][size]:beta[ng.freq-cv-1][size]);
fstar=(double)((double)(ng.freq - cv) - b)/(double)(history.freq-cv);
lambda=(beta[0][size] * suc[0] + beta[1][size] * suc[1] + beta[2][size] * suc[2])
/
(double)(history.freq-cv);
if ((size>=3 && prunesingletons()) ||
(size==maxlevel() && prunetopsingletons()))
//correction
lambda+=(double)(suc[0] * (1-beta[0][size])) / (double)(history.freq-cv);
} else {
// ng.freq==cv
ng.freq>=3?suc[2]--:suc[ng.freq-1]--; //update successor stat
fstar=0.0;
lambda=(beta[0][size] * suc[0] + beta[1][size] * suc[1] + beta[2][size] * suc[2])
/
(double)(history.freq-cv);
if ((size>=3 && prunesingletons()) ||
(size==maxlevel() && prunetopsingletons())) //correction
lambda+=(double)(suc[0] * (1-beta[0][size])) / (double)(history.freq-cv);
ng.freq>=3?suc[2]++:suc[ng.freq-1]++; //resume successor stat
}
} else {
fstar=0.0;
lambda=(beta[0][size] * suc[0] + beta[1][size] * suc[1] + beta[2][size] * suc[2])
/
(double)(history.freq-cv);
if ((size>=3 && prunesingletons()) ||
(size==maxlevel() && prunetopsingletons())) //correction
lambda+=(double)(suc[0] * (1-beta[0][size])) / (double)(history.freq-cv);
}
//cerr << "ngram :" << ng << "\n";
if (*ng.wordp(1)==dict->oovcode()) {
lambda+=fstar;
fstar=0.0;
} else {
*ng.wordp(1)=dict->oovcode();
if (get(ng,size,size)) {
ng.freq=mfreq(ng,size);
if ((!prunesingletons() || mfreq(ng,size)>1 || size<3) &&
(!prunetopsingletons() || mfreq(ng,size)>1 || size<maxlevel())) {
double b=(ng.freq>=3?beta[2][size]:beta[ng.freq-1][size]);
lambda+=(double)(ng.freq - b)/(double)(history.freq-cv);
}
}
}
} else {
fstar=0;
lambda=1;
}
} else { // unigram case, no cross-validation
lambda=0.0;
int unigrtotfreq=(size<lmsize()?btotfreq():totfreq());
if (get(ng,size,size))
fstar=(double) mfreq(ng,size)/(double)unigrtotfreq;
else {
cerr << "Missing probability for word: " << dict->decode(*ng.wordp(1)) << "\n";
exit(1);
}
}
return 1;
}
