Model* TableCategorical::ReadModel(ByteReader* byte_reader, const Schema& schema, size_t index) {
size_t predictor_size = byte_reader->ReadByte();
size_t cell_size = byte_reader->ReadByte();
std::vector<size_t> predictor_list;
for (size_t i = 0; i < predictor_size; ++i ) {
size_t pred = byte_reader->Read16Bit();
predictor_list.push_back(pred);
}
// set err to 0 because err is only used in training
TableCategorical* model = new TableCategorical(schema, predictor_list, index, 0);
size_t target_range = byte_reader->Read16Bit();
model->target_range_ = target_range;
// Read Model Parameters
size_t table_size = model->dynamic_list_.size();
for (size_t i = 0; i < table_size; ++i ) {
std::vector<Prob>& prob_segs = model->dynamic_list_[i].prob;
prob_segs.resize(target_range - 1);
for (size_t j = 0; j < prob_segs.size(); ++j )
if (cell_size == 16) {
prob_segs[j] = GetProb(byte_reader->Read16Bit(), 16);
} else {
prob_segs[j] = GetProb(byte_reader->ReadByte(), 8);
}
}
return model;
}
/* GetProb: return nSize-gram probability for ngram in wlab */
static double GetProb(LabId *wlab, int nSize)
{
/*
this routine will return the interpolated nSize-gram probability for
the words in wlab. Note that the context maybe shortened in the
case of multiple LMs and words which do not occur in some of them.
*/
int i,j;
LMInfo *li;
Boolean inThisLM,inAnyLM;
double x,prob,psum;
NameId nGram[LM_NSIZE];
if (nLModel==1) {
inThisLM = TRUE;
for (j=0; j<nSize; j++) {
if ((nGram[j] = l2nId[0][(int) (wlab[j]->aux)])==NULL)
inThisLM = FALSE;
}
if (inThisLM) {
prob = GetNGramProb(lmInfo[0].lm, nGram, nSize);
}
else if (nSize > 1)
prob = GetProb(wlab+1,nSize-1);
else {
prob = LZERO;
HError(-16690,"GetProb: assigning zero probability");
}
} else {
psum = 0.0;
inAnyLM = FALSE;
for (li=lmInfo, i=0; i<nLModel; i++, li++) {
for (inThisLM=TRUE, j=0; j<nSize; j++) {
if ((nGram[j] = l2nId[i][(int) (wlab[j]->aux)])==NULL)
inThisLM = FALSE;
}
if (!inThisLM)
continue;
x = GetNGramProb(li->lm, nGram, nSize);
#ifdef INTERPOLATE_MAX
if ((x = exp(x)) > psum)
psum = x;
#else
psum += li->weight*exp(x);
#endif
inAnyLM = TRUE;
}
if (inAnyLM)
prob = log(psum);
else if (nSize > 1)
prob = GetProb(wlab+1,nSize-1);
else {
prob = LZERO;
HError(-16690,"GetProb: assigning zero probability");
}
}
return prob;
}
ostream& Print( ostream& o ) const {
return o
<< GetSeqId() << "\t"
<< GetPos() << "\t"
<< GetProb(0) << "\t"
<< GetProb(1) << "\t"
<< GetProb(2) << "\t"
<< GetProb(3);
}
Model* StringModel::ReadModel(ByteReader* byte_reader, size_t index) {
StringModel* model = new StringModel(index);
model->char_count_.clear();
model->length_count_.clear();
model->char_prob_.resize(255);
model->length_prob_.resize(63);
for (int i = 0; i < 255; ++i )
model->char_prob_[i] = GetProb(byte_reader->Read16Bit(), 16);
for (int i = 0; i < 63; ++i )
model->length_prob_[i] = GetProb(byte_reader->ReadByte(), 8);
return model;
}
/* CalcPerplexity: compute perplexity and other statistics */
static void CalcPerplexity(PStats *sent, LabId *pLab, int numPLabs, int nSize)
{
int i,j;
LabId *p;
float prob;
Boolean hasOOV;
for (p=pLab, i=nSize-1; i<numPLabs; i++, p++)
{
if (pLab[i]==unkId)
continue; /* cannot predict OOVs */
if (skipOOV)
{
hasOOV = FALSE;
for (j=1; j<nSize; j++)
{
if (pLab[i-j]==unkId)
{
hasOOV=TRUE;
break;
}
}
if (hasOOV) continue; /* skip to next label since context contains OOV */
}
prob = GetProb(p, nSize);
sent->nWrd++; sent->logpp += prob; sent->logpp2 += prob*prob;
if (outStreamFN != NULL)
fprintf(outStream,"%e\n",exp(prob));
if (trace&T_PROB)
{
printf("logP(%s |", pLab[i]->name);
for (j=1; j<nSize; j++)
{
printf(" %s%s", (j==1)?"":",", pLab[i-j]->name);
}
printf(") = %.4f\n", prob);
/* if (trace&T_INST_INFO) PrintInstStats(nSize); */
fflush(stdout);
}
}
if (trace&T_SENT)
PrintInfo(sent,FALSE);
}
