/**
* Constructor. Initializes on-the-fly composition with a language model.
* \param fst Machine you want to apply the language to. Pass a delayed machine if you can, as it will expand it in constructor.
* \param model A KenLM language model
* \param epsilons List of words to work as epsilons
* \param natlog Use or not natural logs
* \param lmscale Language model scale
*/
ApplyLanguageModelOnTheFly ( const Fst<Arc>& fst,
KenLMModelT& model,
#ifndef USE_GOOGLE_SPARSE_HASH
unordered_set<Label>& epsilons,
#else
google::dense_hash_set<Label>& epsilons,
#endif
bool natlog,
float lmscale ,
float lmwp,
const IdBridgeT& idbridge
) :
composed_ ( NULL ) ,
natlog10_ ( natlog ? -lmscale* ::log ( 10.0 ) : -lmscale ),
fst_ ( fst ),
lmmodel_ ( model ),
vocab_ ( model.GetVocabulary() ),
wp_ ( lmwp ) ,
epsilons_ ( epsilons ) ,
history ( model.Order(), 0),
idbridge_ (idbridge) {
#ifdef USE_GOOGLE_SPARSE_HASH
stateexistence_.set_empty_key ( numeric_limits<ull>::max() );
statemap_.set_empty_key ( numeric_limits<uint64_t>::max() );
basic_string<unsigned> aux (KENLM_MAX_ORDER, numeric_limits<unsigned>::max() );
seenlmstates_.set_empty_key ( aux );
#endif
buffersize = ( model.Order() - 1 ) * sizeof ( unsigned int );
buffer = const_cast<unsigned *> ( history.c_str() );
if (!fst_.NumStates() ) {
LWARN ("Empty lattice");
return;
}
composed_ = new VectorFst<Arc>;
typename KenLMModelT::State bs = model.NullContextState();
///Initialize with first state
pair<StateId, bool> nextp = add ( bs, fst_.Start(),
fst_.Final ( fst_.Start() ) );
qc_.push ( nextp.first );
composed_->SetStart ( nextp.first );
};
