/**
* \brief Private constructor with ucam::util::RegistryPO object and index to a particular language model.
* This constructor is only used when several language models are loaded.
* If the user wants to load several language models (e.g. --lm.load=lm1,lm2,lm3,lm4 and --lm.scale=0.25,0.25,0.25 ).
* The second and following instances of LoadLanguageModelTask will be created using the private constructor.
* In the constructor itself the next language model loader task is appended.
* this constructor, which has an index to the actual language model that must be loaded.
* \param rg : ucam::util::RegistryPO object, containing user parameters.
* \param index : Index to the actual language model.
* \param lmload : key word to access the registry object for language models
* \param lmscale : key word to access the registry object for language model scales.
*/
LoadLanguageModelTask ( const ucam::util::RegistryPO& rg ,
uint index ,
const std::string& lmload = HifstConstants::kLmLoad,
const std::string& lmscale = HifstConstants::kLmFeatureweights ,
const std::string& lmwp = HifstConstants::kLmWordPenalty,
const std::string& wordmapkey = HifstConstants::kLmWordmap
) :
rg_ ( rg ),
lmkey_ ( lmload ),
previous_ ( "" ),
built_ ( false ),
index_ ( index ),
isintegermapped_ (!rg.exists (wordmapkey)
|| rg.get<std::string> (wordmapkey) == ""),
wordmapkey_ (wordmapkey),
lmfile_ ( rg.getVectorString ( lmload , index ) ) {
LDEBUG ( "LM loader using parameters " << lmload << "/" << lmscale <<
", and key " << lmkey_ << ",index=" << index_ << ",wordmapkey=" <<
wordmapkey_);
setLanguageModelScale ( lmscale );
setLanguageModelWordPenalty ( lmwp );
if ( rg.getVectorString ( lmload ).size() > index_ + 1 ) {
LDEBUG ( "Appending Language model..." );
this->appendTask ( new LoadLanguageModelTask ( rg, index_ + 1, lmload, lmscale ,
lmwp , wordmapkey ) );
}
LDEBUG ( "." );
};
///Constructor with registry object, offset and keys
LoadSparseWeightFlowerLatticeTask ( const ucam::util::RegistryPO& rg,
const unsigned offset =
1, //minimum offset considering only one language model...
const std::string& alignmentlattices =
HifstConstants::kRuleflowerlatticeFilterbyalilats,
const std::string& grammarloadkey = HifstConstants::kRuleflowerlatticeLoad,
const std::string& grammarstorekey = HifstConstants::kRuleflowerlatticeStore
) :
offset_ ( offset ),
rg_ ( rg ),
alilats_ ( rg.exists ( alignmentlattices ) ? rg.get<std::string>
( alignmentlattices ) : "" ),
grammar_ ( rg.get<std::string> ( grammarloadkey ) ),
fscales_ ( fst::TropicalSparseTupleWeight<float>::Params() ),
filterbyalilats_ ( rg.exists ( alignmentlattices ) ),
grammarstorekey_ ( grammarstorekey ) {
};
///Constructor with ucam::util::RegistryPO object
ApplyLanguageModelTask ( const ucam::util::RegistryPO& rg ,
const std::string& lmkey = HifstConstants::kLmLoad ,
const std::string& latticeloadkey = "lm.lattice.load",
const std::string& latticestorekey = "lm.lattice.store",
bool deletelmscores = false
) :
lmkey_ ( lmkey ),
latticeloadkey_ ( latticeloadkey ),
latticestorekey_ ( latticestorekey ),
natlog_ ( !rg.exists ( HifstConstants::kLmLogTen ) ),
deletelmscores_ (deletelmscores) {
};
/**
* \brief Public constructor. If the user wants to load several language models (e.g. --lm.load=lm1,lm2,lm3,lm4 and --lm.scale=0.25,0.25,0.25 ),
* the second and following instances of LoadLanguageModelTask will be created using the private constructor (see below), which has an index to the actual language model that must be loaded.
* For the public constructor, the index is set to 0.
* \param rg ucam::util::RegistryPO object, containing user parameters.
* \param lmload key word to access the registry object for language models
* \param lmscale key word to access the registry object for language model scales.
* \param forceone To force the loading of only one language model (i.e. lm1 with scale 0.25).
*/
LoadLanguageModelTask ( const ucam::util::RegistryPO& rg
, const std::string& lmload = HifstConstants::kLmLoad
, const std::string& lmscale =
HifstConstants::kLmFeatureweights //if rg.get(lmscale)=="" the scale will default to 1
, const std::string& lmwp =
HifstConstants::kLmWordPenalty //if rg.get(wps)=="" the scale will default to 0
, const std::string& wordmapkey = HifstConstants::kLmWordmap
, bool forceone = false
)
: rg_ ( rg )
, lmkey_ ( lmload )
, previous_ ( "" )
, built_ ( false )
, index_ ( 0 )
, isintegermapped_ (!rg.exists (wordmapkey)
|| rg.get<std::string> (wordmapkey) == "")
, wordmapkey_ (wordmapkey)
, lmfile_ ( rg.getVectorString ( lmload , 0 ) )
{
LDEBUG ( "LM loader using parameters " << lmload << "/" << lmscale << "/" << lmwp
<< ", and key " << lmkey_ << ",index=" << index_ << ",wordmap=" <<
wordmapkey_);
FORCELINFO("Language model loader for " << lmfile_() );
setLanguageModelScale ( lmscale );
setLanguageModelWordPenalty ( lmwp );
if ( rg_.getVectorString ( lmload ).size() > 1 ) {
if ( !forceone ) {
LINFO ( "Appending Language model..." );
this->appendTask ( new LoadLanguageModelTask ( rg_, 1, lmload, lmscale , lmwp ,
wordmapkey ) );
} else {
LWARN ( "Only one loaded for " << lmload <<
". Extra language models are being ignored" );
}
}
LINFO ( "Finished constructor!" );
};
int SampleWFSAs( ucam::util::RegistryPO const& rg) {
using ucam::util::oszfstream;
using ucam::util::PatternAddress;
PatternAddress<unsigned> input(rg.get<std::string>(HifstConstants::kInput.c_str()));
PatternAddress<unsigned> output(rg.get<std::string>(HifstConstants::kOutput.c_str()));
unsigned n = rg.get<unsigned>(HifstConstants::kNbest.c_str());
unsigned ns = rg.get<unsigned>(HifstConstants::kNSamples.c_str());
float alpha = rg.get<float>(HifstConstants::kAlpha.c_str());
bool negatives = rg.exists(HifstConstants::kNegativeExamples.c_str());
bool binarytarget = rg.exists(HifstConstants::kBinaryTarget.c_str());
bool negate = !rg.exists(HifstConstants::kDontNegate.c_str());
std::string extTok = rg.getString(HifstConstants::kExternalTokenizer.c_str());
std::string wMap = rg.getString(HifstConstants::kWordMap.c_str());
// std::string wMap = "";
//
bool printOutputLabels = rg.exists(HifstConstants::kPrintOutputLabels.c_str());
std::string refFiles;
bool intRefs;
if (rg.exists(HifstConstants::kWordRefs)) {
refFiles = rg.getString(HifstConstants::kWordRefs);
intRefs = false;
}
if (rg.exists(HifstConstants::kIntRefs)) {
refFiles = rg.getString(HifstConstants::kIntRefs);
intRefs = true;
}
std::cerr << refFiles << "**" <<std::endl;
ucam::fsttools::BleuScorer bleuScorer(refFiles, extTok, n, intRefs, wMap);
ucam::fsttools::TuneSet< Arc > tuneSet(rg);
ucam::fsttools::Bleu ibs = tuneSet.ComputeBleu(bleuScorer);
FORCELINFO("Set level Bleu: " << ibs);
unsigned seed = time(NULL);
if (rg.exists(HifstConstants::kRandomSeed.c_str()))
seed = rg.get<unsigned>(HifstConstants::kRandomSeed.c_str());
FORCELINFO("random seed: " << seed);
srand(seed);
boost::scoped_ptr<oszfstream> out;
std::string old;
for (unsigned i=0; i<tuneSet.cachedLats.size(); i++) {
fst::VectorFst<Arc> ifst(*tuneSet.cachedLats[i]);
fst::VectorFst<Arc> nfst;
if (old != output (i) ) {
out.reset(new oszfstream (output(i)));
old = output(i);
}
if (!ifst.NumStates() ) {
FORCELINFO("EMPTY: " << i);
continue;
}
// Projecting allows unique to work for all cases.
fst::Project(&ifst, (printOutputLabels?PROJECT_OUTPUT:PROJECT_INPUT));
ShortestPath (ifst, &nfst, n, true );
std::vector<HypT> hyps;
fst::printStrings<Arc> (nfst, &hyps);
std::vector< LabeledFeature< float, typename Arc::Weight> > fea =
ProSBLEUSample<typename Arc::Weight, HypT>(bleuScorer, hyps, i, n, ns, alpha, negatives, negate);
for (unsigned s=0; s<fea.size(); s++) {
*out << (binarytarget ? (fea[s].value > 0.0 ? 1 : 0) : fea[s].value);
*out << " " << fea[s].fea << std::endl;
}
}
FORCELINFO("Done Sample WFSAs");
};