// $Id$ // vim:tabstop=2 #include "moses/FeatureVector.h" #include "moses/TranslationModel/PhraseDictionaryTree.h" #include #include "util/check.hh" #include #include #include #include #include namespace Moses { template std::ostream& operator<<(std::ostream& out,const std::vector& x) { out<::const_iterator iend=x.end(); for(typename std::vector::const_iterator i=x.begin(); i!=iend; ++i) out<<*i<<' '; return out; } class TgtCand { IPhrase e; Scores sc; std::string m_alignment; IPhrase fnames; std::vector fvalues; static const float SPARSE_FLAG; public: TgtCand() {} TgtCand(const IPhrase& a, const Scores& b , const std::string& alignment) : e(a) , sc(b) , m_alignment(alignment) {} TgtCand(const IPhrase& a,const Scores& b) : e(a),sc(b) {} TgtCand(FILE* f) { readBin(f); } void writeBin(FILE* f) const { fWriteVector(f,e); //This is a bit ugly, but if there is a sparse vector, add //an extra score with value 100. Can the last score be 100? //Probably not, since scores are probabilities and phrase penalty. if (fnames.size()) { Scores sc_copy(sc); sc_copy.push_back(SPARSE_FLAG); fWriteVector(f,sc_copy); fWriteVector(f,fnames); fWriteVector(f,fvalues); } else { fWriteVector(f,sc); } } void readBin(FILE* f) { fReadVector(f,e); fReadVector(f,sc); if (sc.back() == 100) { sc.pop_back(); fReadVector(f,fnames); fReadVector(f,fvalues); } } void writeBinWithAlignment(FILE* f) const { writeBin(f); fWriteString(f, m_alignment.c_str(), m_alignment.size()); } void readBinWithAlignment(FILE* f) { readBin(f); fReadString(f, m_alignment); } const IPhrase& GetPhrase() const { return e; } const Scores& GetScores() const { return sc; } const std::string& GetAlignment() const { return m_alignment; } const IPhrase& GetFeatureNames() const { return fnames; } const std::vector GetFeatureValues() const { return fvalues; } void SetFeatures(const IPhrase& names, const std::vector& values) { CHECK(names.size() == values.size()); fnames = names; fvalues = values; } }; const float TgtCand::SPARSE_FLAG = 100; class TgtCands : public std::vector { typedef std::vector MyBase; public: TgtCands() : MyBase() {} void writeBin(FILE* f) const { unsigned s=size(); fWrite(f,s); for(size_t i=0; iisValid(); } typedef LVoc WordVoc; class PDTimp { public: typedef PrefixTreeF PTF; typedef FilePtr CPT; typedef std::vector Data; Data data; std::vector srcOffsets; FILE *os,*ot; WordVoc sv; WordVoc tv; ObjectPool pPool; // a comparison with the Boost MemPools might be useful bool needwordalign, haswordAlign; bool printwordalign; PDTimp() : os(0),ot(0), printwordalign(false) { PTF::setDefault(InvalidOffT); } ~PDTimp() { if(os) fClose(os); if(ot) fClose(ot); FreeMemory(); } inline void NeedAlignmentInfo(bool a) { needwordalign=a; } inline bool NeedAlignmentInfo() { return needwordalign; }; inline void HasAlignmentInfo(bool a) { haswordAlign=a; } inline bool HasAlignmentInfo() { return haswordAlign; }; inline void PrintWordAlignment(bool a) { printwordalign=a; }; inline bool PrintWordAlignment() { return printwordalign; }; void FreeMemory() { for(Data::iterator i=data.begin(); i!=data.end(); ++i) (*i).free(); pPool.reset(); } int Read(const std::string& fn); void GetTargetCandidates(const IPhrase& f,TgtCands& tgtCands) { if(f.empty()) return; if(f[0]>=data.size()) return; if(!data[f[0]]) return; CHECK(data[f[0]]->findKey(f[0])size()); OFF_T tCandOffset=data[f[0]]->find(f); if(tCandOffset==InvalidOffT) return; fSeek(ot,tCandOffset); if (HasAlignmentInfo()) tgtCands.readBinWithAlignment(ot); else tgtCands.readBin(ot); } typedef PhraseDictionaryTree::PrefixPtr PPtr; void GetTargetCandidates(PPtr p,TgtCands& tgtCands) { CHECK(p); if(p.imp->isRoot()) return; OFF_T tCandOffset=p.imp->ptr()->getData(p.imp->idx); if(tCandOffset==InvalidOffT) return; fSeek(ot,tCandOffset); if (HasAlignmentInfo()) tgtCands.readBinWithAlignment(ot); else tgtCands.readBin(ot); } void PrintTgtCand(const TgtCands& tcands,std::ostream& out) const; // convert target candidates from internal data structure to the external one void ConvertTgtCand(const TgtCands& tcands,std::vector& rv, std::vector* wa) const { for(TgtCands::const_iterator i=tcands.begin(); i!=tcands.end(); ++i) { rv.push_back(StringTgtCand()); const IPhrase& iphrase=i->GetPhrase(); rv.back().tokens.reserve(iphrase.size()); for(size_t j=0; jGetScores(); const IPhrase& fnames = i->GetFeatureNames(); for (size_t j = 0; j < fnames.size(); ++j) { rv.back().fnames.push_back(&tv.symbol(fnames[j])); } rv.back().fvalues = i->GetFeatureValues(); if (wa) wa->push_back(i->GetAlignment()); } } PPtr GetRoot() { return PPtr(pPool.get(PPimp(0,0,1))); } PPtr Extend(PPtr p,const std::string& w) { CHECK(p); if(w.empty() || w==EPSILON) return p; LabelId wi=sv.index(w); if(wi==InvalidLabelId) return PPtr(); // unknown word else if(p.imp->isRoot()) { if(wifindKeyPtr(wi); CHECK(ptr); return PPtr(pPool.get(PPimp(data[wi],data[wi]->findKey(wi),0))); } } else if(PTF const* nextP=p.imp->ptr()->getPtr(p.imp->idx)) { return PPtr(pPool.get(PPimp(nextP,nextP->findKey(wi),0))); } return PPtr(); } WordVoc* ReadVoc(const std::string& filename); }; //////////////////////////////////////////////////////////// // // member functions of PDTimp // //////////////////////////////////////////////////////////// int PDTimp::Read(const std::string& fn) { std::string ifs, ift, ifi, ifsv, iftv; HasAlignmentInfo(FileExists(fn+".binphr.srctree.wa")); if (NeedAlignmentInfo() && !HasAlignmentInfo()) { // ERROR std::stringstream strme; strme << "You are asking for word alignment but the binary phrase table does not contain any alignment info. Please check if you had generated the correct phrase table with word alignment (.wa)\n"; UserMessage::Add(strme.str()); return false; } if (HasAlignmentInfo()) { ifs=fn+".binphr.srctree.wa"; ift=fn+".binphr.tgtdata.wa"; } else { ifs=fn+".binphr.srctree"; ift=fn+".binphr.tgtdata"; } ifi=fn+".binphr.idx"; ifsv=fn+".binphr.srcvoc"; iftv=fn+".binphr.tgtvoc"; FILE *ii=fOpen(ifi.c_str(),"rb"); fReadVector(ii,srcOffsets); fClose(ii); os=fOpen(ifs.c_str(),"rb"); ot=fOpen(ift.c_str(),"rb"); data.resize(srcOffsets.size()); for(size_t i=0; i abort \n\n"); abort(); } } PhraseDictionaryTree::~PhraseDictionaryTree() { delete imp; } void PhraseDictionaryTree::NeedAlignmentInfo(bool a) { imp->NeedAlignmentInfo(a); }; void PhraseDictionaryTree::PrintWordAlignment(bool a) { imp->PrintWordAlignment(a); }; bool PhraseDictionaryTree::PrintWordAlignment() { return imp->PrintWordAlignment(); }; void PhraseDictionaryTree::FreeMemory() const { imp->FreeMemory(); } void PhraseDictionaryTree:: GetTargetCandidates(const std::vector& src, std::vector& rv) const { IPhrase f(src.size()); for(size_t i=0; isv.index(src[i]); if(f[i]==InvalidLabelId) return; } TgtCands tgtCands; imp->GetTargetCandidates(f,tgtCands); imp->ConvertTgtCand(tgtCands,rv,NULL); } void PhraseDictionaryTree:: GetTargetCandidates(const std::vector& src, std::vector& rv, std::vector& wa) const { IPhrase f(src.size()); for(size_t i=0; isv.index(src[i]); if(f[i]==InvalidLabelId) return; } TgtCands tgtCands; imp->GetTargetCandidates(f,tgtCands); imp->ConvertTgtCand(tgtCands,rv,&wa); } void PhraseDictionaryTree:: PrintTargetCandidates(const std::vector& src, std::ostream& out) const { IPhrase f(src.size()); for(size_t i=0; isv.index(src[i]); if(f[i]==InvalidLabelId) { TRACE_ERR("the source phrase '"<GetTargetCandidates(f,tcand); imp->PrintTgtCand(tcand,out); } int PhraseDictionaryTree::Create(std::istream& inFile,const std::string& out) { std::string line; size_t count = 0; std::string ofn(out+".binphr.srctree"), oft(out+".binphr.tgtdata"), ofi(out+".binphr.idx"), ofsv(out+".binphr.srcvoc"), oftv(out+".binphr.tgtvoc"); if (PrintWordAlignment()) { ofn+=".wa"; oft+=".wa"; } FILE *os=fOpen(ofn.c_str(),"wb"), *ot=fOpen(oft.c_str(),"wb"); typedef PrefixTreeSA PSA; PSA *psa=new PSA; PSA::setDefault(InvalidOffT); LabelId currFirstWord=InvalidLabelId; IPhrase currF; TgtCands tgtCands; std::vector vo; size_t lnc=0; size_t numElement = NOT_FOUND; // 3=old format, 5=async format which include word alignment info size_t missingAlignmentCount = 0; while(getline(inFile, line)) { ++lnc; std::vector tokens = TokenizeMultiCharSeparator( line , "|||" ); if (numElement == NOT_FOUND) { // init numElement numElement = tokens.size(); CHECK(numElement >= 3); } if (tokens.size() != numElement) { std::stringstream strme; strme << "Syntax error at line " << lnc << " : " << line; UserMessage::Add(strme.str()); abort(); } const std::string &sourcePhraseString =tokens[0] ,&targetPhraseString=tokens[1] ,&scoreString = tokens[2]; const std::string empty; const std::string &alignmentString = PrintWordAlignment() ? tokens[3] : empty; const std::string sparseFeatureString = tokens.size() > 5 ? tokens[5] : empty; IPhrase f,e; Scores sc; if (PrintWordAlignment() && alignmentString == " ") ++missingAlignmentCount; std::vector wordVec = Tokenize(sourcePhraseString); for (size_t i = 0 ; i < wordVec.size() ; ++i) f.push_back(imp->sv.add(wordVec[i])); wordVec = Tokenize(targetPhraseString); for (size_t i = 0 ; i < wordVec.size() ; ++i) e.push_back(imp->tv.add(wordVec[i])); // while(is>>w && w!="|||") sc.push_back(atof(w.c_str())); // Mauro: to handle 0 probs in phrase tables std::vector scoreVector = Tokenize(scoreString); for (size_t i = 0 ; i < scoreVector.size() ; ++i) { float tmp = scoreVector[i]; sc.push_back(((tmp>0.0)?tmp:(float)1.0e-38)); } if(f.empty()) { TRACE_ERR("WARNING: empty source phrase in line '"<insert(f); if(d==InvalidOffT) d=fTell(ot); else { TRACE_ERR("ERROR: source phrase already inserted (A)!\nline(" << lnc << "): '" < fvalues; if (!sparseFeatureString.empty()) { std::vector sparseTokens = Tokenize(sparseFeatureString); if (sparseTokens.size() % 2 != 0) { TRACE_ERR("ERROR: incorrectly formatted sparse feature string: " << sparseFeatureString << std::endl); abort(); } for (size_t i = 0; i < sparseTokens.size(); i+=2) { fnames.push_back(imp->tv.add(sparseTokens[i])); fvalues.push_back(Scan(sparseTokens[i+1])); } } if(currF!=f) { // new src phrase currF=f; if (PrintWordAlignment()) tgtCands.writeBinWithAlignment(ot); else tgtCands.writeBin(ot); tgtCands.clear(); if(++count%10000==0) { TRACE_ERR("."); if(count%500000==0) TRACE_ERR("[phrase:"<=vo.size()) vo.resize(currFirstWord+1,InvalidOffT); vo[currFirstWord]=fTell(os); pf.create(*psa,os); // clear delete psa; psa=new PSA; currFirstWord=f[0]; } // insert src phrase in prefix tree CHECK(psa); PSA::Data& d=psa->insert(f); if(d==InvalidOffT) d=fTell(ot); else { TRACE_ERR("ERROR: xsource phrase already inserted (B)!\nline(" << lnc << "): '" <=vo.size()) vo.resize(currFirstWord+1,InvalidOffT); vo[currFirstWord]=fTell(os); pf.create(*psa,os); delete psa; psa=0; TRACE_ERR("distinct source phrases: "< inv; for(size_t i=0; isv.Write(ofsv); imp->tv.Write(oftv); return 1; } int PhraseDictionaryTree::Read(const std::string& fn) { TRACE_ERR("size of OFF_T "<Read(fn); } PhraseDictionaryTree::PrefixPtr PhraseDictionaryTree::GetRoot() const { return imp->GetRoot(); } PhraseDictionaryTree::PrefixPtr PhraseDictionaryTree::Extend(PrefixPtr p, const std::string& w) const { return imp->Extend(p,w); } void PhraseDictionaryTree::PrintTargetCandidates(PrefixPtr p,std::ostream& out) const { TgtCands tcand; imp->GetTargetCandidates(p,tcand); out<<"there are "<PrintTgtCand(tcand,out); } void PhraseDictionaryTree:: GetTargetCandidates(PrefixPtr p, std::vector& rv) const { TgtCands tcands; imp->GetTargetCandidates(p,tcands); imp->ConvertTgtCand(tcands,rv,NULL); } void PhraseDictionaryTree:: GetTargetCandidates(PrefixPtr p, std::vector& rv, std::vector& wa) const { TgtCands tcands; imp->GetTargetCandidates(p,tcands); imp->ConvertTgtCand(tcands,rv,&wa); } }