diff options
author | Nicola Bertoldi <bertoldi@fbk.eu> | 2014-11-18 17:16:16 +0300 |
---|---|---|
committer | Nicola Bertoldi <bertoldi@fbk.eu> | 2014-11-18 17:16:16 +0300 |
commit | 46596ed49a63c5b76f822b3dfff6f10c4da5ef22 (patch) | |
tree | 7a581634bcfe3166f819631449d8f1e5d810f352 /moses-cmd | |
parent | 92c2d3e2a4386cee14bba7be9d5ae1f732b4b9a0 (diff) | |
parent | 070afadbcd47c83a0cfce785efe6af065492cfbd (diff) |
Merge branch 'master' of https://github.com/moses-smt/mosesdecoder
Diffstat (limited to 'moses-cmd')
-rw-r--r-- | moses-cmd/IOWrapper.cpp | 661 | ||||
-rw-r--r-- | moses-cmd/IOWrapper.h | 164 | ||||
-rw-r--r-- | moses-cmd/Jamfile | 2 | ||||
-rw-r--r-- | moses-cmd/LatticeMBR.cpp | 669 | ||||
-rw-r--r-- | moses-cmd/LatticeMBR.h | 153 | ||||
-rw-r--r-- | moses-cmd/LatticeMBRGrid.cpp | 18 | ||||
-rw-r--r-- | moses-cmd/Main.cpp | 494 | ||||
-rw-r--r-- | moses-cmd/Main.h | 5 | ||||
-rw-r--r-- | moses-cmd/TranslationAnalysis.cpp | 137 | ||||
-rw-r--r-- | moses-cmd/TranslationAnalysis.h | 24 | ||||
-rw-r--r-- | moses-cmd/mbr.cpp | 178 | ||||
-rw-r--r-- | moses-cmd/mbr.h | 28 | ||||
-rw-r--r-- | moses-cmd/simulate-pe.cc | 777 |
13 files changed, 144 insertions, 3166 deletions
diff --git a/moses-cmd/IOWrapper.cpp b/moses-cmd/IOWrapper.cpp deleted file mode 100644 index 1483e0472..000000000 --- a/moses-cmd/IOWrapper.cpp +++ /dev/null @@ -1,661 +0,0 @@ -// $Id$ - -/*********************************************************************** -Moses - factored phrase-based language decoder -Copyright (c) 2006 University of Edinburgh -All rights reserved. - -Redistribution and use in source and binary forms, with or without modification, -are permitted provided that the following conditions are met: - - * Redistributions of source code must retain the above copyright notice, - this list of conditions and the following disclaimer. - * Redistributions in binary form must reproduce the above copyright notice, - this list of conditions and the following disclaimer in the documentation - and/or other materials provided with the distribution. - * Neither the name of the University of Edinburgh nor the names of its contributors - may be used to endorse or promote products derived from this software - without specific prior written permission. - -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, -THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR -PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS -BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER -IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. - ***********************************************************************/ - -// example file on how to use moses library - -#include <iostream> -#include <stack> -#include <boost/algorithm/string.hpp> - -#include "moses/TypeDef.h" -#include "moses/Util.h" -#include "moses/Hypothesis.h" -#include "moses/WordsRange.h" -#include "moses/TrellisPathList.h" -#include "moses/StaticData.h" -#include "moses/FeatureVector.h" -#include "moses/InputFileStream.h" -#include "moses/FF/StatefulFeatureFunction.h" -#include "moses/FF/StatelessFeatureFunction.h" -#include "util/exception.hh" - -#include "IOWrapper.h" - -using namespace std; -using namespace Moses; - -namespace MosesCmd -{ - -IOWrapper::IOWrapper( - const vector<FactorType> &inputFactorOrder - , const vector<FactorType> &outputFactorOrder - , const FactorMask &inputFactorUsed - , size_t nBestSize - , const string &nBestFilePath) - :m_inputFactorOrder(inputFactorOrder) - ,m_outputFactorOrder(outputFactorOrder) - ,m_inputFactorUsed(inputFactorUsed) - ,m_inputFile(NULL) - ,m_inputStream(&std::cin) - ,m_nBestStream(NULL) - ,m_outputWordGraphStream(NULL) - ,m_outputSearchGraphStream(NULL) - ,m_detailedTranslationReportingStream(NULL) - ,m_alignmentOutputStream(NULL) -{ - Initialization(inputFactorOrder, outputFactorOrder - , inputFactorUsed - , nBestSize, nBestFilePath); -} - -IOWrapper::IOWrapper(const std::vector<FactorType> &inputFactorOrder - , const std::vector<FactorType> &outputFactorOrder - , const FactorMask &inputFactorUsed - , size_t nBestSize - , const std::string &nBestFilePath - , const std::string &inputFilePath) - :m_inputFactorOrder(inputFactorOrder) - ,m_outputFactorOrder(outputFactorOrder) - ,m_inputFactorUsed(inputFactorUsed) - ,m_inputFilePath(inputFilePath) - ,m_inputFile(new InputFileStream(inputFilePath)) - ,m_nBestStream(NULL) - ,m_outputWordGraphStream(NULL) - ,m_outputSearchGraphStream(NULL) - ,m_detailedTranslationReportingStream(NULL) - ,m_alignmentOutputStream(NULL) -{ - Initialization(inputFactorOrder, outputFactorOrder - , inputFactorUsed - , nBestSize, nBestFilePath); - - m_inputStream = m_inputFile; -} - -IOWrapper::~IOWrapper() -{ - if (m_inputFile != NULL) - delete m_inputFile; - if (m_nBestStream != NULL && !m_surpressSingleBestOutput) { - // outputting n-best to file, rather than stdout. need to close file and delete obj - delete m_nBestStream; - } - if (m_outputWordGraphStream != NULL) { - delete m_outputWordGraphStream; - } - if (m_outputSearchGraphStream != NULL) { - delete m_outputSearchGraphStream; - } - delete m_detailedTranslationReportingStream; - delete m_alignmentOutputStream; -} - -void IOWrapper::Initialization(const std::vector<FactorType> &/*inputFactorOrder*/ - , const std::vector<FactorType> &/*outputFactorOrder*/ - , const FactorMask &/*inputFactorUsed*/ - , size_t nBestSize - , const std::string &nBestFilePath) -{ - const StaticData &staticData = StaticData::Instance(); - - // n-best - m_surpressSingleBestOutput = false; - - if (nBestSize > 0) { - if (nBestFilePath == "-" || nBestFilePath == "/dev/stdout") { - m_nBestStream = &std::cout; - m_surpressSingleBestOutput = true; - } else { - std::ofstream *file = new std::ofstream; - m_nBestStream = file; - file->open(nBestFilePath.c_str()); - } - } - - // wordgraph output - if (staticData.GetOutputWordGraph()) { - string fileName = staticData.GetParam("output-word-graph")[0]; - std::ofstream *file = new std::ofstream; - m_outputWordGraphStream = file; - file->open(fileName.c_str()); - } - - - // search graph output - if (staticData.GetOutputSearchGraph()) { - string fileName; - if (staticData.GetOutputSearchGraphExtended()) - fileName = staticData.GetParam("output-search-graph-extended")[0]; - else - fileName = staticData.GetParam("output-search-graph")[0]; - std::ofstream *file = new std::ofstream; - m_outputSearchGraphStream = file; - file->open(fileName.c_str()); - } - - // detailed translation reporting - if (staticData.IsDetailedTranslationReportingEnabled()) { - const std::string &path = staticData.GetDetailedTranslationReportingFilePath(); - m_detailedTranslationReportingStream = new std::ofstream(path.c_str()); - UTIL_THROW_IF(!m_detailedTranslationReportingStream->good(), - util::FileOpenException, - "File for output of detailed translation report could not be open"); - } - - // sentence alignment output - if (! staticData.GetAlignmentOutputFile().empty()) { - m_alignmentOutputStream = new ofstream(staticData.GetAlignmentOutputFile().c_str()); - UTIL_THROW_IF(!m_alignmentOutputStream->good(), - util::FileOpenException, - "File for output of word alignment could not be open"); - } - -} - -InputType* -IOWrapper:: -GetInput(InputType* inputType) -{ - if(inputType->Read(*m_inputStream, m_inputFactorOrder)) { - if (long x = inputType->GetTranslationId()) { - if (x>=m_translationId) m_translationId = x+1; - } else inputType->SetTranslationId(m_translationId++); - - return inputType; - } else { - delete inputType; - return NULL; - } -} - -std::map<size_t, const Factor*> GetPlaceholders(const Hypothesis &hypo, FactorType placeholderFactor) -{ - const InputPath &inputPath = hypo.GetTranslationOption().GetInputPath(); - const Phrase &inputPhrase = inputPath.GetPhrase(); - - std::map<size_t, const Factor*> ret; - - for (size_t sourcePos = 0; sourcePos < inputPhrase.GetSize(); ++sourcePos) { - const Factor *factor = inputPhrase.GetFactor(sourcePos, placeholderFactor); - if (factor) { - std::set<size_t> targetPos = hypo.GetTranslationOption().GetTargetPhrase().GetAlignTerm().GetAlignmentsForSource(sourcePos); - UTIL_THROW_IF2(targetPos.size() != 1, - "Placeholder should be aligned to 1, and only 1, word"); - ret[*targetPos.begin()] = factor; - } - } - - return ret; -} - -/*** - * print surface factor only for the given phrase - */ -void OutputSurface(std::ostream &out, const Hypothesis &edge, const std::vector<FactorType> &outputFactorOrder, - char reportSegmentation, bool reportAllFactors) -{ - UTIL_THROW_IF2(outputFactorOrder.size() == 0, - "Must specific at least 1 output factor"); - const TargetPhrase& phrase = edge.GetCurrTargetPhrase(); - bool markUnknown = StaticData::Instance().GetMarkUnknown(); - if (reportAllFactors == true) { - out << phrase; - } else { - FactorType placeholderFactor = StaticData::Instance().GetPlaceholderFactor(); - - std::map<size_t, const Factor*> placeholders; - if (placeholderFactor != NOT_FOUND) { - // creates map of target position -> factor for placeholders - placeholders = GetPlaceholders(edge, placeholderFactor); - } - - size_t size = phrase.GetSize(); - for (size_t pos = 0 ; pos < size ; pos++) { - const Factor *factor = phrase.GetFactor(pos, outputFactorOrder[0]); - - if (placeholders.size()) { - // do placeholders - std::map<size_t, const Factor*>::const_iterator iter = placeholders.find(pos); - if (iter != placeholders.end()) { - factor = iter->second; - } - } - - UTIL_THROW_IF2(factor == NULL, - "No factor 0 at position " << pos); - - //preface surface form with UNK if marking unknowns - const Word &word = phrase.GetWord(pos); - if(markUnknown && word.IsOOV()) { - out << "UNK" << *factor; - } else { - out << *factor; - } - - for (size_t i = 1 ; i < outputFactorOrder.size() ; i++) { - const Factor *factor = phrase.GetFactor(pos, outputFactorOrder[i]); - UTIL_THROW_IF2(factor == NULL, - "No factor " << i << " at position " << pos); - - out << "|" << *factor; - } - out << " "; - } - } - - // trace ("report segmentation") option "-t" / "-tt" - if (reportSegmentation > 0 && phrase.GetSize() > 0) { - const WordsRange &sourceRange = edge.GetCurrSourceWordsRange(); - const int sourceStart = sourceRange.GetStartPos(); - const int sourceEnd = sourceRange.GetEndPos(); - out << "|" << sourceStart << "-" << sourceEnd; // enriched "-tt" - if (reportSegmentation == 2) { - out << ",wa="; - const AlignmentInfo &ai = edge.GetCurrTargetPhrase().GetAlignTerm(); - OutputAlignment(out, ai, 0, 0); - out << ",total="; - out << edge.GetScore() - edge.GetPrevHypo()->GetScore(); - out << ","; - ScoreComponentCollection scoreBreakdown(edge.GetScoreBreakdown()); - scoreBreakdown.MinusEquals(edge.GetPrevHypo()->GetScoreBreakdown()); - OutputAllFeatureScores(scoreBreakdown, out); - } - out << "| "; - } -} - -void OutputBestSurface(std::ostream &out, const Hypothesis *hypo, const std::vector<FactorType> &outputFactorOrder, - char reportSegmentation, bool reportAllFactors) -{ - if (hypo != NULL) { - // recursively retrace this best path through the lattice, starting from the end of the hypothesis sentence - OutputBestSurface(out, hypo->GetPrevHypo(), outputFactorOrder, reportSegmentation, reportAllFactors); - OutputSurface(out, *hypo, outputFactorOrder, reportSegmentation, reportAllFactors); - } -} - -void OutputAlignment(ostream &out, const AlignmentInfo &ai, size_t sourceOffset, size_t targetOffset) -{ - typedef std::vector< const std::pair<size_t,size_t>* > AlignVec; - AlignVec alignments = ai.GetSortedAlignments(); - - AlignVec::const_iterator it; - for (it = alignments.begin(); it != alignments.end(); ++it) { - const std::pair<size_t,size_t> &alignment = **it; - out << alignment.first + sourceOffset << "-" << alignment.second + targetOffset << " "; - } - -} - -void OutputAlignment(ostream &out, const vector<const Hypothesis *> &edges) -{ - size_t targetOffset = 0; - - for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - const TargetPhrase &tp = edge.GetCurrTargetPhrase(); - size_t sourceOffset = edge.GetCurrSourceWordsRange().GetStartPos(); - - OutputAlignment(out, tp.GetAlignTerm(), sourceOffset, targetOffset); - - targetOffset += tp.GetSize(); - } - out << std::endl; -} - -void OutputAlignment(std::ostream &out, const Moses::Hypothesis *hypo) -{ - std::vector<const Hypothesis *> edges; - const Hypothesis *currentHypo = hypo; - while (currentHypo) { - edges.push_back(currentHypo); - currentHypo = currentHypo->GetPrevHypo(); - } - - OutputAlignment(out, edges); - -} - -void OutputAlignment(OutputCollector* collector, size_t lineNo , const vector<const Hypothesis *> &edges) -{ - ostringstream out; - OutputAlignment(out, edges); - - collector->Write(lineNo,out.str()); -} - -void OutputAlignment(OutputCollector* collector, size_t lineNo , const Hypothesis *hypo) -{ - if (collector) { - std::vector<const Hypothesis *> edges; - const Hypothesis *currentHypo = hypo; - while (currentHypo) { - edges.push_back(currentHypo); - currentHypo = currentHypo->GetPrevHypo(); - } - - OutputAlignment(collector,lineNo, edges); - } -} - -void OutputAlignment(OutputCollector* collector, size_t lineNo , const TrellisPath &path) -{ - if (collector) { - OutputAlignment(collector,lineNo, path.GetEdges()); - } -} - -void OutputBestHypo(const Moses::TrellisPath &path, long /*translationId*/, char reportSegmentation, bool reportAllFactors, std::ostream &out) -{ - const std::vector<const Hypothesis *> &edges = path.GetEdges(); - - for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - OutputSurface(out, edge, StaticData::Instance().GetOutputFactorOrder(), reportSegmentation, reportAllFactors); - } - out << endl; -} - -void IOWrapper::Backtrack(const Hypothesis *hypo) -{ - - if (hypo->GetPrevHypo() != NULL) { - VERBOSE(3,hypo->GetId() << " <= "); - Backtrack(hypo->GetPrevHypo()); - } -} - -void OutputBestHypo(const std::vector<Word>& mbrBestHypo, long /*translationId*/, char /*reportSegmentation*/, bool /*reportAllFactors*/, ostream& out) -{ - - for (size_t i = 0 ; i < mbrBestHypo.size() ; i++) { - const Factor *factor = mbrBestHypo[i].GetFactor(StaticData::Instance().GetOutputFactorOrder()[0]); - UTIL_THROW_IF2(factor == NULL, - "No factor 0 at position " << i); - if (i>0) out << " " << *factor; - else out << *factor; - } - out << endl; -} - - -void OutputInput(std::vector<const Phrase*>& map, const Hypothesis* hypo) -{ - if (hypo->GetPrevHypo()) { - OutputInput(map, hypo->GetPrevHypo()); - map[hypo->GetCurrSourceWordsRange().GetStartPos()] = &hypo->GetTranslationOption().GetInputPath().GetPhrase(); - } -} - -void OutputInput(std::ostream& os, const Hypothesis* hypo) -{ - size_t len = hypo->GetInput().GetSize(); - std::vector<const Phrase*> inp_phrases(len, 0); - OutputInput(inp_phrases, hypo); - for (size_t i=0; i<len; ++i) - if (inp_phrases[i]) os << *inp_phrases[i]; -} - -void IOWrapper::OutputBestHypo(const Hypothesis *hypo, long /*translationId*/, char reportSegmentation, bool reportAllFactors) -{ - if (hypo != NULL) { - VERBOSE(1,"BEST TRANSLATION: " << *hypo << endl); - VERBOSE(3,"Best path: "); - Backtrack(hypo); - VERBOSE(3,"0" << std::endl); - if (!m_surpressSingleBestOutput) { - if (StaticData::Instance().GetOutputHypoScore()) { - cout << hypo->GetTotalScore() << " "; - } - - if (StaticData::Instance().IsPathRecoveryEnabled()) { - OutputInput(cout, hypo); - cout << "||| "; - } - OutputBestSurface(cout, hypo, m_outputFactorOrder, reportSegmentation, reportAllFactors); - cout << endl; - } - } else { - VERBOSE(1, "NO BEST TRANSLATION" << endl); - if (!m_surpressSingleBestOutput) { - cout << endl; - } - } -} - -void OutputNBest(std::ostream& out - , const Moses::TrellisPathList &nBestList - , const std::vector<Moses::FactorType>& outputFactorOrder - , long translationId - , char reportSegmentation) -{ - const StaticData &staticData = StaticData::Instance(); - bool reportAllFactors = staticData.GetReportAllFactorsNBest(); - bool includeSegmentation = staticData.NBestIncludesSegmentation(); - bool includeWordAlignment = staticData.PrintAlignmentInfoInNbest(); - - TrellisPathList::const_iterator iter; - for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) { - const TrellisPath &path = **iter; - const std::vector<const Hypothesis *> &edges = path.GetEdges(); - - // print the surface factor of the translation - out << translationId << " ||| "; - for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - OutputSurface(out, edge, outputFactorOrder, reportSegmentation, reportAllFactors); - } - out << " |||"; - - // print scores with feature names - OutputAllFeatureScores(path.GetScoreBreakdown(), out ); - - // total - out << " ||| " << path.GetTotalScore(); - - //phrase-to-phrase segmentation - if (includeSegmentation) { - out << " |||"; - for (int currEdge = (int)edges.size() - 2 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - const WordsRange &sourceRange = edge.GetCurrSourceWordsRange(); - WordsRange targetRange = path.GetTargetWordsRange(edge); - out << " " << sourceRange.GetStartPos(); - if (sourceRange.GetStartPos() < sourceRange.GetEndPos()) { - out << "-" << sourceRange.GetEndPos(); - } - out<< "=" << targetRange.GetStartPos(); - if (targetRange.GetStartPos() < targetRange.GetEndPos()) { - out<< "-" << targetRange.GetEndPos(); - } - } - } - - if (includeWordAlignment) { - out << " ||| "; - for (int currEdge = (int)edges.size() - 2 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - const WordsRange &sourceRange = edge.GetCurrSourceWordsRange(); - WordsRange targetRange = path.GetTargetWordsRange(edge); - const int sourceOffset = sourceRange.GetStartPos(); - const int targetOffset = targetRange.GetStartPos(); - const AlignmentInfo &ai = edge.GetCurrTargetPhrase().GetAlignTerm(); - - OutputAlignment(out, ai, sourceOffset, targetOffset); - - } - } - - if (StaticData::Instance().IsPathRecoveryEnabled()) { - out << " ||| "; - OutputInput(out, edges[0]); - } - - out << endl; - } - - out << std::flush; -} - -void OutputAllFeatureScores(const Moses::ScoreComponentCollection &features - , std::ostream &out) -{ - std::string lastName = ""; - const vector<const StatefulFeatureFunction*>& sff = StatefulFeatureFunction::GetStatefulFeatureFunctions(); - for( size_t i=0; i<sff.size(); i++ ) { - const StatefulFeatureFunction *ff = sff[i]; - if (ff->GetScoreProducerDescription() != "BleuScoreFeature" - && ff->IsTuneable()) { - OutputFeatureScores( out, features, ff, lastName ); - } - } - const vector<const StatelessFeatureFunction*>& slf = StatelessFeatureFunction::GetStatelessFeatureFunctions(); - for( size_t i=0; i<slf.size(); i++ ) { - const StatelessFeatureFunction *ff = slf[i]; - if (ff->IsTuneable()) { - OutputFeatureScores( out, features, ff, lastName ); - } - } -} - -void OutputFeatureScores( std::ostream& out - , const ScoreComponentCollection &features - , const FeatureFunction *ff - , std::string &lastName ) -{ - const StaticData &staticData = StaticData::Instance(); - bool labeledOutput = staticData.IsLabeledNBestList(); - - // regular features (not sparse) - if (ff->GetNumScoreComponents() != 0) { - if( labeledOutput && lastName != ff->GetScoreProducerDescription() ) { - lastName = ff->GetScoreProducerDescription(); - out << " " << lastName << "="; - } - vector<float> scores = features.GetScoresForProducer( ff ); - for (size_t j = 0; j<scores.size(); ++j) { - out << " " << scores[j]; - } - } - - // sparse features - const FVector scores = features.GetVectorForProducer( ff ); - for(FVector::FNVmap::const_iterator i = scores.cbegin(); i != scores.cend(); i++) { - out << " " << i->first << "= " << i->second; - } -} - -void OutputLatticeMBRNBest(std::ostream& out, const vector<LatticeMBRSolution>& solutions,long translationId) -{ - for (vector<LatticeMBRSolution>::const_iterator si = solutions.begin(); si != solutions.end(); ++si) { - out << translationId; - out << " |||"; - const vector<Word> mbrHypo = si->GetWords(); - for (size_t i = 0 ; i < mbrHypo.size() ; i++) { - const Factor *factor = mbrHypo[i].GetFactor(StaticData::Instance().GetOutputFactorOrder()[0]); - if (i>0) out << " " << *factor; - else out << *factor; - } - out << " |||"; - out << " map: " << si->GetMapScore(); - out << " w: " << mbrHypo.size(); - const vector<float>& ngramScores = si->GetNgramScores(); - for (size_t i = 0; i < ngramScores.size(); ++i) { - out << " " << ngramScores[i]; - } - out << " ||| " << si->GetScore(); - - out << endl; - } -} - - -void IOWrapper::OutputLatticeMBRNBestList(const vector<LatticeMBRSolution>& solutions,long translationId) -{ - OutputLatticeMBRNBest(*m_nBestStream, solutions,translationId); -} - -bool ReadInput(IOWrapper &ioWrapper, InputTypeEnum inputType, InputType*& source) -{ - if (source) delete source; - switch(inputType) { - case SentenceInput: - source = ioWrapper.GetInput(new Sentence); - break; - case ConfusionNetworkInput: - source = ioWrapper.GetInput(new ConfusionNet); - break; - case WordLatticeInput: - source = ioWrapper.GetInput(new WordLattice); - break; - default: - TRACE_ERR("Unknown input type: " << inputType << "\n"); - source = NULL; - } - return (source ? true : false); -} - - - -IOWrapper *GetIOWrapper(const StaticData &staticData) -{ - IOWrapper *ioWrapper; - const std::vector<FactorType> &inputFactorOrder = staticData.GetInputFactorOrder() - ,&outputFactorOrder = staticData.GetOutputFactorOrder(); - FactorMask inputFactorUsed(inputFactorOrder); - - // io - if (staticData.GetParam("input-file").size() == 1) { - VERBOSE(2,"IO from File" << endl); - string filePath = staticData.GetParam("input-file")[0]; - - ioWrapper = new IOWrapper(inputFactorOrder, outputFactorOrder, inputFactorUsed - , staticData.GetNBestSize() - , staticData.GetNBestFilePath() - , filePath); - } else { - VERBOSE(1,"IO from STDOUT/STDIN" << endl); - ioWrapper = new IOWrapper(inputFactorOrder, outputFactorOrder, inputFactorUsed - , staticData.GetNBestSize() - , staticData.GetNBestFilePath()); - } - ioWrapper->ResetTranslationId(); - - IFVERBOSE(1) - PrintUserTime("Created input-output object"); - - return ioWrapper; -} - -} - diff --git a/moses-cmd/IOWrapper.h b/moses-cmd/IOWrapper.h deleted file mode 100644 index 93539ce0b..000000000 --- a/moses-cmd/IOWrapper.h +++ /dev/null @@ -1,164 +0,0 @@ -// $Id$ - -/*********************************************************************** -Moses - factored phrase-based language decoder -Copyright (c) 2006 University of Edinburgh -All rights reserved. - -Redistribution and use in source and binary forms, with or without modification, -are permitted provided that the following conditions are met: - - * Redistributions of source code must retain the above copyright notice, - this list of conditions and the following disclaimer. - * Redistributions in binary form must reproduce the above copyright notice, - this list of conditions and the following disclaimer in the documentation - and/or other materials provided with the distribution. - * Neither the name of the University of Edinburgh nor the names of its contributors - may be used to endorse or promote products derived from this software - without specific prior written permission. - -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, -THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR -PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS -BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER -IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -// example file on how to use moses library - -#ifndef moses_cmd_IOWrapper_h -#define moses_cmd_IOWrapper_h - -#include <cassert> -#include <fstream> -#include <ostream> -#include <vector> - -#include "moses/TypeDef.h" -#include "moses/Sentence.h" -#include "moses/FactorTypeSet.h" -#include "moses/FactorCollection.h" -#include "moses/Hypothesis.h" -#include "moses/OutputCollector.h" -#include "moses/TrellisPathList.h" -#include "moses/InputFileStream.h" -#include "moses/InputType.h" -#include "moses/WordLattice.h" -#include "LatticeMBR.h" - -namespace Moses -{ -class ScoreComponentCollection; -class Hypothesis; -class Factor; -} - -namespace MosesCmd -{ - -/** Helper class that holds misc variables to write data out to command line. - */ -class IOWrapper -{ -protected: - long m_translationId; - - const std::vector<Moses::FactorType> &m_inputFactorOrder; - const std::vector<Moses::FactorType> &m_outputFactorOrder; - const Moses::FactorMask &m_inputFactorUsed; - std::string m_inputFilePath; - Moses::InputFileStream *m_inputFile; - std::istream *m_inputStream; - std::ostream *m_nBestStream - ,*m_outputWordGraphStream,*m_outputSearchGraphStream; - std::ostream *m_detailedTranslationReportingStream; - std::ofstream *m_alignmentOutputStream; - bool m_surpressSingleBestOutput; - - void Initialization(const std::vector<Moses::FactorType> &inputFactorOrder - , const std::vector<Moses::FactorType> &outputFactorOrder - , const Moses::FactorMask &inputFactorUsed - , size_t nBestSize - , const std::string &nBestFilePath); - - -public: - IOWrapper(const std::vector<Moses::FactorType> &inputFactorOrder - , const std::vector<Moses::FactorType> &outputFactorOrder - , const Moses::FactorMask &inputFactorUsed - , size_t nBestSize - , const std::string &nBestFilePath); - - IOWrapper(const std::vector<Moses::FactorType> &inputFactorOrder - , const std::vector<Moses::FactorType> &outputFactorOrder - , const Moses::FactorMask &inputFactorUsed - , size_t nBestSize - , const std::string &nBestFilePath - , const std::string &infilePath); - ~IOWrapper(); - - Moses::InputType* GetInput(Moses::InputType *inputType); - - void OutputBestHypo(const Moses::Hypothesis *hypo, long translationId, char reportSegmentation, bool reportAllFactors); - void OutputLatticeMBRNBestList(const std::vector<LatticeMBRSolution>& solutions,long translationId); - void Backtrack(const Moses::Hypothesis *hypo); - - void ResetTranslationId() { - m_translationId = 0; - } - - std::ofstream *GetAlignmentOutputStream() { - return m_alignmentOutputStream; - } - - std::ostream &GetOutputWordGraphStream() { - return *m_outputWordGraphStream; - } - std::ostream &GetOutputSearchGraphStream() { - return *m_outputSearchGraphStream; - } - - std::ostream &GetDetailedTranslationReportingStream() { - assert (m_detailedTranslationReportingStream); - return *m_detailedTranslationReportingStream; - } -}; - -IOWrapper *GetIOWrapper(const Moses::StaticData &staticData); -bool ReadInput(IOWrapper &ioWrapper, Moses::InputTypeEnum inputType, Moses::InputType*& source); -void OutputLanguageModelOrder(std::ostream &out, const Moses::Hypothesis *hypo, Moses::Manager &manager); -void OutputBestSurface(std::ostream &out, const Moses::Hypothesis *hypo, const std::vector<Moses::FactorType> &outputFactorOrder, char reportSegmentation, bool reportAllFactors); -void OutputLatticeMBRNBest(std::ostream& out, const std::vector<LatticeMBRSolution>& solutions,long translationId); -void OutputBestHypo(const std::vector<Moses::Word>& mbrBestHypo, long /*translationId*/, - char reportSegmentation, bool reportAllFactors, std::ostream& out); -void OutputBestHypo(const Moses::TrellisPath &path, long /*translationId*/,char reportSegmentation, bool reportAllFactors, std::ostream &out); -void OutputInput(std::ostream& os, const Moses::Hypothesis* hypo); -void OutputAlignment(Moses::OutputCollector* collector, size_t lineNo, const Moses::Hypothesis *hypo); -void OutputAlignment(Moses::OutputCollector* collector, size_t lineNo, const Moses::TrellisPath &path); -void OutputAlignment(std::ostream &out, const Moses::Hypothesis *hypo); -void OutputAlignment(std::ostream &out, const Moses::AlignmentInfo &ai, size_t sourceOffset, size_t targetOffset); - -void OutputNBest(std::ostream& out - , const Moses::TrellisPathList &nBestList - , const std::vector<Moses::FactorType>& outputFactorOrder - , long translationId - , char reportSegmentation); -void OutputAllFeatureScores(const Moses::ScoreComponentCollection &features - , std::ostream &out); -void OutputFeatureScores( std::ostream& out - , const Moses::ScoreComponentCollection &features - , const Moses::FeatureFunction *ff - , std::string &lastName ); - -// creates a map of TARGET positions which should be replaced by word using placeholder -std::map<size_t, const Moses::Factor*> GetPlaceholders(const Moses::Hypothesis &hypo, Moses::FactorType placeholderFactor); - -} - -#endif diff --git a/moses-cmd/Jamfile b/moses-cmd/Jamfile index 79af196e9..7ee90850c 100644 --- a/moses-cmd/Jamfile +++ b/moses-cmd/Jamfile @@ -1,4 +1,4 @@ -alias deps : IOWrapper.cpp mbr.cpp LatticeMBR.cpp TranslationAnalysis.cpp ..//z ..//boost_iostreams ..//boost_filesystem ../moses//moses ; +alias deps : ..//z ..//boost_iostreams ..//boost_filesystem ../moses//moses ; exe moses : Main.cpp deps ; exe lmbrgrid : LatticeMBRGrid.cpp deps ; diff --git a/moses-cmd/LatticeMBR.cpp b/moses-cmd/LatticeMBR.cpp deleted file mode 100644 index 148b44743..000000000 --- a/moses-cmd/LatticeMBR.cpp +++ /dev/null @@ -1,669 +0,0 @@ -/* - * LatticeMBR.cpp - * moses-cmd - * - * Created by Abhishek Arun on 26/01/2010. - * Copyright 2010 __MyCompanyName__. All rights reserved. - * - */ - -#include "LatticeMBR.h" -#include "moses/StaticData.h" -#include <algorithm> -#include <set> - -using namespace std; -using namespace Moses; - -namespace MosesCmd -{ - -size_t bleu_order = 4; -float UNKNGRAMLOGPROB = -20; -void GetOutputWords(const TrellisPath &path, vector <Word> &translation) -{ - const std::vector<const Hypothesis *> &edges = path.GetEdges(); - - // print the surface factor of the translation - for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - const Phrase &phrase = edge.GetCurrTargetPhrase(); - size_t size = phrase.GetSize(); - for (size_t pos = 0 ; pos < size ; pos++) { - translation.push_back(phrase.GetWord(pos)); - } - } -} - - -void extract_ngrams(const vector<Word >& sentence, map < Phrase, int > & allngrams) -{ - for (int k = 0; k < (int)bleu_order; k++) { - for(int i =0; i < max((int)sentence.size()-k,0); i++) { - Phrase ngram( k+1); - for ( int j = i; j<= i+k; j++) { - ngram.AddWord(sentence[j]); - } - ++allngrams[ngram]; - } - } -} - - - -void NgramScores::addScore(const Hypothesis* node, const Phrase& ngram, float score) -{ - set<Phrase>::const_iterator ngramIter = m_ngrams.find(ngram); - if (ngramIter == m_ngrams.end()) { - ngramIter = m_ngrams.insert(ngram).first; - } - map<const Phrase*,float>& ngramScores = m_scores[node]; - map<const Phrase*,float>::iterator scoreIter = ngramScores.find(&(*ngramIter)); - if (scoreIter == ngramScores.end()) { - ngramScores[&(*ngramIter)] = score; - } else { - ngramScores[&(*ngramIter)] = log_sum(score,scoreIter->second); - } -} - -NgramScores::NodeScoreIterator NgramScores::nodeBegin(const Hypothesis* node) -{ - return m_scores[node].begin(); -} - - -NgramScores::NodeScoreIterator NgramScores::nodeEnd(const Hypothesis* node) -{ - return m_scores[node].end(); -} - -LatticeMBRSolution::LatticeMBRSolution(const TrellisPath& path, bool isMap) : - m_score(0.0f) -{ - const std::vector<const Hypothesis *> &edges = path.GetEdges(); - - for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - const Phrase &phrase = edge.GetCurrTargetPhrase(); - size_t size = phrase.GetSize(); - for (size_t pos = 0 ; pos < size ; pos++) { - m_words.push_back(phrase.GetWord(pos)); - } - } - if (isMap) { - m_mapScore = path.GetTotalScore(); - } else { - m_mapScore = 0; - } -} - - -void LatticeMBRSolution::CalcScore(map<Phrase, float>& finalNgramScores, const vector<float>& thetas, float mapWeight) -{ - m_ngramScores.assign(thetas.size()-1, -10000); - - map < Phrase, int > counts; - extract_ngrams(m_words,counts); - - //Now score this translation - m_score = thetas[0] * m_words.size(); - - //Calculate the ngramScores, working in log space at first - for (map < Phrase, int >::iterator ngrams = counts.begin(); ngrams != counts.end(); ++ngrams) { - float ngramPosterior = UNKNGRAMLOGPROB; - map<Phrase,float>::const_iterator ngramPosteriorIt = finalNgramScores.find(ngrams->first); - if (ngramPosteriorIt != finalNgramScores.end()) { - ngramPosterior = ngramPosteriorIt->second; - } - size_t ngramSize = ngrams->first.GetSize(); - m_ngramScores[ngramSize-1] = log_sum(log((float)ngrams->second) + ngramPosterior,m_ngramScores[ngramSize-1]); - } - - //convert from log to probability and create weighted sum - for (size_t i = 0; i < m_ngramScores.size(); ++i) { - m_ngramScores[i] = exp(m_ngramScores[i]); - m_score += thetas[i+1] * m_ngramScores[i]; - } - - - //The map score - m_score += m_mapScore*mapWeight; -} - - -void pruneLatticeFB(Lattice & connectedHyp, map < const Hypothesis*, set <const Hypothesis* > > & outgoingHyps, map<const Hypothesis*, vector<Edge> >& incomingEdges, - const vector< float> & estimatedScores, const Hypothesis* bestHypo, size_t edgeDensity, float scale) -{ - - //Need hyp 0 in connectedHyp - Find empty hypothesis - VERBOSE(2,"Pruning lattice to edge density " << edgeDensity << endl); - const Hypothesis* emptyHyp = connectedHyp.at(0); - while (emptyHyp->GetId() != 0) { - emptyHyp = emptyHyp->GetPrevHypo(); - } - connectedHyp.push_back(emptyHyp); //Add it to list of hyps - - //Need hyp 0's outgoing Hyps - for (size_t i = 0; i < connectedHyp.size(); ++i) { - if (connectedHyp[i]->GetId() > 0 && connectedHyp[i]->GetPrevHypo()->GetId() == 0) - outgoingHyps[emptyHyp].insert(connectedHyp[i]); - } - - //sort hyps based on estimated scores - do so by copying to multimap - multimap<float, const Hypothesis*> sortHypsByVal; - for (size_t i =0; i < estimatedScores.size(); ++i) { - sortHypsByVal.insert(make_pair(estimatedScores[i], connectedHyp[i])); - } - - multimap<float, const Hypothesis*>::const_iterator it = --sortHypsByVal.end(); - float bestScore = it->first; - //store best score as score of hyp 0 - sortHypsByVal.insert(make_pair(bestScore, emptyHyp)); - - - IFVERBOSE(3) { - for (multimap<float, const Hypothesis*>::const_iterator it = --sortHypsByVal.end(); it != --sortHypsByVal.begin(); --it) { - const Hypothesis* currHyp = it->second; - cerr << "Hyp " << currHyp->GetId() << ", estimated score: " << it->first << endl; - } - } - - - set <const Hypothesis*> survivingHyps; //store hyps that make the cut in this - - VERBOSE(2, "BEST HYPO TARGET LENGTH : " << bestHypo->GetSize() << endl) - size_t numEdgesTotal = edgeDensity * bestHypo->GetSize(); //as per Shankar, aim for (density * target length of MAP solution) arcs - size_t numEdgesCreated = 0; - VERBOSE(2, "Target edge count: " << numEdgesTotal << endl); - - float prevScore = -999999; - - //now iterate over multimap - for (multimap<float, const Hypothesis*>::const_iterator it = --sortHypsByVal.end(); it != --sortHypsByVal.begin(); --it) { - float currEstimatedScore = it->first; - const Hypothesis* currHyp = it->second; - - if (numEdgesCreated >= numEdgesTotal && prevScore > currEstimatedScore) //if this hyp has equal estimated score to previous, include its edges too - break; - - prevScore = currEstimatedScore; - VERBOSE(3, "Num edges created : "<< numEdgesCreated << ", numEdges wanted " << numEdgesTotal << endl) - VERBOSE(3, "Considering hyp " << currHyp->GetId() << ", estimated score: " << it->first << endl) - - survivingHyps.insert(currHyp); //CurrHyp made the cut - - // is its best predecessor already included ? - if (survivingHyps.find(currHyp->GetPrevHypo()) != survivingHyps.end()) { //yes, then add an edge - vector <Edge>& edges = incomingEdges[currHyp]; - Edge winningEdge(currHyp->GetPrevHypo(),currHyp,scale*(currHyp->GetScore() - currHyp->GetPrevHypo()->GetScore()),currHyp->GetCurrTargetPhrase()); - edges.push_back(winningEdge); - ++numEdgesCreated; - } - - //let's try the arcs too - const ArcList *arcList = currHyp->GetArcList(); - if (arcList != NULL) { - ArcList::const_iterator iterArcList; - for (iterArcList = arcList->begin() ; iterArcList != arcList->end() ; ++iterArcList) { - const Hypothesis *loserHypo = *iterArcList; - const Hypothesis* loserPrevHypo = loserHypo->GetPrevHypo(); - if (survivingHyps.find(loserPrevHypo) != survivingHyps.end()) { //found it, add edge - double arcScore = loserHypo->GetScore() - loserPrevHypo->GetScore(); - Edge losingEdge(loserPrevHypo, currHyp, arcScore*scale, loserHypo->GetCurrTargetPhrase()); - vector <Edge>& edges = incomingEdges[currHyp]; - edges.push_back(losingEdge); - ++numEdgesCreated; - } - } - } - - //Now if a successor node has already been visited, add an edge connecting the two - map < const Hypothesis*, set < const Hypothesis* > >::const_iterator outgoingIt = outgoingHyps.find(currHyp); - - if (outgoingIt != outgoingHyps.end()) {//currHyp does have successors - const set<const Hypothesis*> & outHyps = outgoingIt->second; //the successors - for (set<const Hypothesis*>::const_iterator outHypIts = outHyps.begin(); outHypIts != outHyps.end(); ++outHypIts) { - const Hypothesis* succHyp = *outHypIts; - - if (survivingHyps.find(succHyp) == survivingHyps.end()) //Have we encountered the successor yet? - continue; //No, move on to next - - //Curr Hyp can be : a) the best predecessor of succ b) or an arc attached to succ - if (succHyp->GetPrevHypo() == currHyp) { //best predecessor - vector <Edge>& succEdges = incomingEdges[succHyp]; - Edge succWinningEdge(currHyp, succHyp, scale*(succHyp->GetScore() - currHyp->GetScore()), succHyp->GetCurrTargetPhrase()); - succEdges.push_back(succWinningEdge); - survivingHyps.insert(succHyp); - ++numEdgesCreated; - } - - //now, let's find an arc - const ArcList *arcList = succHyp->GetArcList(); - if (arcList != NULL) { - ArcList::const_iterator iterArcList; - //QUESTION: What happens if there's more than one loserPrevHypo? - for (iterArcList = arcList->begin() ; iterArcList != arcList->end() ; ++iterArcList) { - const Hypothesis *loserHypo = *iterArcList; - const Hypothesis* loserPrevHypo = loserHypo->GetPrevHypo(); - if (loserPrevHypo == currHyp) { //found it - vector <Edge>& succEdges = incomingEdges[succHyp]; - double arcScore = loserHypo->GetScore() - currHyp->GetScore(); - Edge losingEdge(currHyp, succHyp,scale* arcScore, loserHypo->GetCurrTargetPhrase()); - succEdges.push_back(losingEdge); - ++numEdgesCreated; - } - } - } - } - } - } - - connectedHyp.clear(); - for (set <const Hypothesis*>::iterator it = survivingHyps.begin(); it != survivingHyps.end(); ++it) { - connectedHyp.push_back(*it); - } - - VERBOSE(2, "Done! Num edges created : "<< numEdgesCreated << ", numEdges wanted " << numEdgesTotal << endl) - - IFVERBOSE(3) { - cerr << "Surviving hyps: " ; - for (set <const Hypothesis*>::iterator it = survivingHyps.begin(); it != survivingHyps.end(); ++it) { - cerr << (*it)->GetId() << " "; - } - cerr << endl; - } - - -} - -void calcNgramExpectations(Lattice & connectedHyp, map<const Hypothesis*, vector<Edge> >& incomingEdges, - map<Phrase, float>& finalNgramScores, bool posteriors) -{ - - sort(connectedHyp.begin(),connectedHyp.end(),ascendingCoverageCmp); //sort by increasing source word cov - - /*cerr << "Lattice:" << endl; - for (Lattice::const_iterator i = connectedHyp.begin(); i != connectedHyp.end(); ++i) { - const Hypothesis* h = *i; - cerr << *h << endl; - const vector<Edge>& edges = incomingEdges[h]; - for (size_t e = 0; e < edges.size(); ++e) { - cerr << edges[e]; - } - }*/ - - map<const Hypothesis*, float> forwardScore; - forwardScore[connectedHyp[0]] = 0.0f; //forward score of hyp 0 is 1 (or 0 in logprob space) - set< const Hypothesis *> finalHyps; //store completed hyps - - NgramScores ngramScores;//ngram scores for each hyp - - for (size_t i = 1; i < connectedHyp.size(); ++i) { - const Hypothesis* currHyp = connectedHyp[i]; - if (currHyp->GetWordsBitmap().IsComplete()) { - finalHyps.insert(currHyp); - } - - VERBOSE(3, "Processing hyp: " << currHyp->GetId() << ", num words cov= " << currHyp->GetWordsBitmap().GetNumWordsCovered() << endl) - - vector <Edge> & edges = incomingEdges[currHyp]; - for (size_t e = 0; e < edges.size(); ++e) { - const Edge& edge = edges[e]; - if (forwardScore.find(currHyp) == forwardScore.end()) { - forwardScore[currHyp] = forwardScore[edge.GetTailNode()] + edge.GetScore(); - VERBOSE(3, "Fwd score["<<currHyp->GetId()<<"] = fwdScore["<<edge.GetTailNode()->GetId() << "] + edge Score: " << edge.GetScore() << endl) - } else { - forwardScore[currHyp] = log_sum(forwardScore[currHyp], forwardScore[edge.GetTailNode()] + edge.GetScore()); - VERBOSE(3, "Fwd score["<<currHyp->GetId()<<"] += fwdScore["<<edge.GetTailNode()->GetId() << "] + edge Score: " << edge.GetScore() << endl) - } - } - - //Process ngrams now - for (size_t j =0 ; j < edges.size(); ++j) { - Edge& edge = edges[j]; - const NgramHistory & incomingPhrases = edge.GetNgrams(incomingEdges); - - //let's first score ngrams introduced by this edge - for (NgramHistory::const_iterator it = incomingPhrases.begin(); it != incomingPhrases.end(); ++it) { - const Phrase& ngram = it->first; - const PathCounts& pathCounts = it->second; - VERBOSE(4, "Calculating score for: " << it->first << endl) - - for (PathCounts::const_iterator pathCountIt = pathCounts.begin(); pathCountIt != pathCounts.end(); ++pathCountIt) { - //Score of an n-gram is forward score of head node of leftmost edge + all edge scores - const Path& path = pathCountIt->first; - //cerr << "path count for " << ngram << " is " << pathCountIt->second << endl; - float score = forwardScore[path[0]->GetTailNode()]; - for (size_t i = 0; i < path.size(); ++i) { - score += path[i]->GetScore(); - } - //if we're doing expectations, then the number of times the ngram - //appears on the path is relevant. - size_t count = posteriors ? 1 : pathCountIt->second; - for (size_t k = 0; k < count; ++k) { - ngramScores.addScore(currHyp,ngram,score); - } - } - } - - //Now score ngrams that are just being propagated from the history - for (NgramScores::NodeScoreIterator it = ngramScores.nodeBegin(edge.GetTailNode()); - it != ngramScores.nodeEnd(edge.GetTailNode()); ++it) { - const Phrase & currNgram = *(it->first); - float currNgramScore = it->second; - VERBOSE(4, "Calculating score for: " << currNgram << endl) - - // For posteriors, don't double count ngrams - if (!posteriors || incomingPhrases.find(currNgram) == incomingPhrases.end()) { - float score = edge.GetScore() + currNgramScore; - ngramScores.addScore(currHyp,currNgram,score); - } - } - - } - } - - float Z = 9999999; //the total score of the lattice - - //Done - Print out ngram posteriors for final hyps - for (set< const Hypothesis *>::iterator finalHyp = finalHyps.begin(); finalHyp != finalHyps.end(); ++finalHyp) { - const Hypothesis* hyp = *finalHyp; - - for (NgramScores::NodeScoreIterator it = ngramScores.nodeBegin(hyp); it != ngramScores.nodeEnd(hyp); ++it) { - const Phrase& ngram = *(it->first); - if (finalNgramScores.find(ngram) == finalNgramScores.end()) { - finalNgramScores[ngram] = it->second; - } else { - finalNgramScores[ngram] = log_sum(it->second, finalNgramScores[ngram]); - } - } - - if (Z == 9999999) { - Z = forwardScore[hyp]; - } else { - Z = log_sum(Z, forwardScore[hyp]); - } - } - - //Z *= scale; //scale the score - - for (map<Phrase, float>::iterator finalScoresIt = finalNgramScores.begin(); finalScoresIt != finalNgramScores.end(); ++finalScoresIt) { - finalScoresIt->second = finalScoresIt->second - Z; - IFVERBOSE(2) { - VERBOSE(2,finalScoresIt->first << " [" << finalScoresIt->second << "]" << endl); - } - } - -} - -const NgramHistory& Edge::GetNgrams(map<const Hypothesis*, vector<Edge> > & incomingEdges) -{ - - if (m_ngrams.size() > 0) - return m_ngrams; - - const Phrase& currPhrase = GetWords(); - //Extract the n-grams local to this edge - for (size_t start = 0; start < currPhrase.GetSize(); ++start) { - for (size_t end = start; end < start + bleu_order; ++end) { - if (end < currPhrase.GetSize()) { - Phrase edgeNgram(end-start+1); - for (size_t index = start; index <= end; ++index) { - edgeNgram.AddWord(currPhrase.GetWord(index)); - } - //cout << "Inserting Phrase : " << edgeNgram << endl; - vector<const Edge*> edgeHistory; - edgeHistory.push_back(this); - storeNgramHistory(edgeNgram, edgeHistory); - } else { - break; - } - } - } - - map<const Hypothesis*, vector<Edge> >::iterator it = incomingEdges.find(m_tailNode); - if (it != incomingEdges.end()) { //node has incoming edges - vector<Edge> & inEdges = it->second; - - for (vector<Edge>::iterator edge = inEdges.begin(); edge != inEdges.end(); ++edge) {//add the ngrams straddling prev and curr edge - const NgramHistory & edgeIncomingNgrams = edge->GetNgrams(incomingEdges); - for (NgramHistory::const_iterator edgeInNgramHist = edgeIncomingNgrams.begin(); edgeInNgramHist != edgeIncomingNgrams.end(); ++edgeInNgramHist) { - const Phrase& edgeIncomingNgram = edgeInNgramHist->first; - const PathCounts & edgeIncomingNgramPaths = edgeInNgramHist->second; - size_t back = min(edgeIncomingNgram.GetSize(), edge->GetWordsSize()); - const Phrase& edgeWords = edge->GetWords(); - IFVERBOSE(3) { - cerr << "Edge: "<< *edge <<endl; - cerr << "edgeWords: " << edgeWords << endl; - cerr << "edgeInNgram: " << edgeIncomingNgram << endl; - } - - Phrase edgeSuffix(ARRAY_SIZE_INCR); - Phrase ngramSuffix(ARRAY_SIZE_INCR); - GetPhraseSuffix(edgeWords,back,edgeSuffix); - GetPhraseSuffix(edgeIncomingNgram,back,ngramSuffix); - - if (ngramSuffix == edgeSuffix) { //we've got the suffix of previous edge - size_t edgeInNgramSize = edgeIncomingNgram.GetSize(); - - for (size_t i = 0; i < GetWordsSize() && i + edgeInNgramSize < bleu_order ; ++i) { - Phrase newNgram(edgeIncomingNgram); - for (size_t j = 0; j <= i ; ++j) { - newNgram.AddWord(GetWords().GetWord(j)); - } - VERBOSE(3, "Inserting New Phrase : " << newNgram << endl) - - for (PathCounts::const_iterator pathIt = edgeIncomingNgramPaths.begin(); pathIt != edgeIncomingNgramPaths.end(); ++pathIt) { - Path newNgramPath = pathIt->first; - newNgramPath.push_back(this); - storeNgramHistory(newNgram, newNgramPath, pathIt->second); - } - } - } - } - } - } - return m_ngrams; -} - -//Add the last lastN words of origPhrase to targetPhrase -void Edge::GetPhraseSuffix(const Phrase& origPhrase, size_t lastN, Phrase& targetPhrase) const -{ - size_t origSize = origPhrase.GetSize(); - size_t startIndex = origSize - lastN; - for (size_t index = startIndex; index < origPhrase.GetSize(); ++index) { - targetPhrase.AddWord(origPhrase.GetWord(index)); - } -} - -bool Edge::operator< (const Edge& compare ) const -{ - if (m_headNode->GetId() < compare.m_headNode->GetId()) - return true; - if (compare.m_headNode->GetId() < m_headNode->GetId()) - return false; - if (m_tailNode->GetId() < compare.m_tailNode->GetId()) - return true; - if (compare.m_tailNode->GetId() < m_tailNode->GetId()) - return false; - return GetScore() < compare.GetScore(); -} - -ostream& operator<< (ostream& out, const Edge& edge) -{ - out << "Head: " << edge.m_headNode->GetId() << ", Tail: " << edge.m_tailNode->GetId() << ", Score: " << edge.m_score << ", Phrase: " << edge.m_targetPhrase << endl; - return out; -} - -bool ascendingCoverageCmp(const Hypothesis* a, const Hypothesis* b) -{ - return a->GetWordsBitmap().GetNumWordsCovered() < b->GetWordsBitmap().GetNumWordsCovered(); -} - -void getLatticeMBRNBest(Manager& manager, TrellisPathList& nBestList, - vector<LatticeMBRSolution>& solutions, size_t n) -{ - const StaticData& staticData = StaticData::Instance(); - std::map < int, bool > connected; - std::vector< const Hypothesis *> connectedList; - map<Phrase, float> ngramPosteriors; - std::map < const Hypothesis*, set <const Hypothesis*> > outgoingHyps; - map<const Hypothesis*, vector<Edge> > incomingEdges; - vector< float> estimatedScores; - manager.GetForwardBackwardSearchGraph(&connected, &connectedList, &outgoingHyps, &estimatedScores); - pruneLatticeFB(connectedList, outgoingHyps, incomingEdges, estimatedScores, manager.GetBestHypothesis(), staticData.GetLatticeMBRPruningFactor(),staticData.GetMBRScale()); - calcNgramExpectations(connectedList, incomingEdges, ngramPosteriors,true); - - vector<float> mbrThetas = staticData.GetLatticeMBRThetas(); - float p = staticData.GetLatticeMBRPrecision(); - float r = staticData.GetLatticeMBRPRatio(); - float mapWeight = staticData.GetLatticeMBRMapWeight(); - if (mbrThetas.size() == 0) { //thetas not specified on the command line, use p and r instead - mbrThetas.push_back(-1); //Theta 0 - mbrThetas.push_back(1/(bleu_order*p)); - for (size_t i = 2; i <= bleu_order; ++i) { - mbrThetas.push_back(mbrThetas[i-1] / r); - } - } - IFVERBOSE(2) { - VERBOSE(2,"Thetas: "); - for (size_t i = 0; i < mbrThetas.size(); ++i) { - VERBOSE(2,mbrThetas[i] << " "); - } - VERBOSE(2,endl); - } - TrellisPathList::const_iterator iter; - size_t ctr = 0; - LatticeMBRSolutionComparator comparator; - for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter, ++ctr) { - const TrellisPath &path = **iter; - solutions.push_back(LatticeMBRSolution(path,iter==nBestList.begin())); - solutions.back().CalcScore(ngramPosteriors,mbrThetas,mapWeight); - sort(solutions.begin(), solutions.end(), comparator); - while (solutions.size() > n) { - solutions.pop_back(); - } - } - VERBOSE(2,"LMBR Score: " << solutions[0].GetScore() << endl); -} - -vector<Word> doLatticeMBR(Manager& manager, TrellisPathList& nBestList) -{ - - vector<LatticeMBRSolution> solutions; - getLatticeMBRNBest(manager, nBestList, solutions,1); - return solutions.at(0).GetWords(); -} - -const TrellisPath doConsensusDecoding(Manager& manager, TrellisPathList& nBestList) -{ - static const int BLEU_ORDER = 4; - static const float SMOOTH = 1; - - //calculate the ngram expectations - const StaticData& staticData = StaticData::Instance(); - std::map < int, bool > connected; - std::vector< const Hypothesis *> connectedList; - map<Phrase, float> ngramExpectations; - std::map < const Hypothesis*, set <const Hypothesis*> > outgoingHyps; - map<const Hypothesis*, vector<Edge> > incomingEdges; - vector< float> estimatedScores; - manager.GetForwardBackwardSearchGraph(&connected, &connectedList, &outgoingHyps, &estimatedScores); - pruneLatticeFB(connectedList, outgoingHyps, incomingEdges, estimatedScores, manager.GetBestHypothesis(), staticData.GetLatticeMBRPruningFactor(),staticData.GetMBRScale()); - calcNgramExpectations(connectedList, incomingEdges, ngramExpectations,false); - - //expected length is sum of expected unigram counts - //cerr << "Thread " << pthread_self() << " Ngram expectations size: " << ngramExpectations.size() << endl; - float ref_length = 0.0f; - for (map<Phrase,float>::const_iterator ref_iter = ngramExpectations.begin(); - ref_iter != ngramExpectations.end(); ++ref_iter) { - //cerr << "Ngram: " << ref_iter->first << " score: " << - // ref_iter->second << endl; - if (ref_iter->first.GetSize() == 1) { - ref_length += exp(ref_iter->second); - // cerr << "Expected for " << ref_iter->first << " is " << exp(ref_iter->second) << endl; - } - } - - VERBOSE(2,"REF Length: " << ref_length << endl); - - //use the ngram expectations to rescore the nbest list. - TrellisPathList::const_iterator iter; - TrellisPathList::const_iterator best = nBestList.end(); - float bestScore = -100000; - //cerr << "nbest list size: " << nBestList.GetSize() << endl; - for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) { - const TrellisPath &path = **iter; - vector<Word> words; - map<Phrase,int> ngrams; - GetOutputWords(path,words); - /*for (size_t i = 0; i < words.size(); ++i) { - cerr << words[i].GetFactor(0)->GetString() << " "; - } - cerr << endl; - */ - extract_ngrams(words,ngrams); - - vector<float> comps(2*BLEU_ORDER+1); - float logbleu = 0.0; - float brevity = 0.0; - int hyp_length = words.size(); - for (int i = 0; i < BLEU_ORDER; ++i) { - comps[2*i] = 0.0; - comps[2*i+1] = max(hyp_length-i,0); - } - - for (map<Phrase,int>::const_iterator hyp_iter = ngrams.begin(); - hyp_iter != ngrams.end(); ++hyp_iter) { - map<Phrase,float>::const_iterator ref_iter = ngramExpectations.find(hyp_iter->first); - if (ref_iter != ngramExpectations.end()) { - comps[2*(hyp_iter->first.GetSize()-1)] += min(exp(ref_iter->second), (float)(hyp_iter->second)); - } - - } - comps[comps.size()-1] = ref_length; - /*for (size_t i = 0; i < comps.size(); ++i) { - cerr << comps[i] << " "; - } - cerr << endl; - */ - - float score = 0.0f; - if (comps[0] != 0) { - for (int i=0; i<BLEU_ORDER; i++) { - if ( i > 0 ) { - logbleu += log((float)comps[2*i]+SMOOTH)-log((float)comps[2*i+1]+SMOOTH); - } else { - logbleu += log((float)comps[2*i])-log((float)comps[2*i+1]); - } - } - logbleu /= BLEU_ORDER; - brevity = 1.0-(float)comps[comps.size()-1]/comps[1]; // comps[comps_n-1] is the ref length, comps[1] is the test length - if (brevity < 0.0) { - logbleu += brevity; - } - score = exp(logbleu); - } - - //cerr << "score: " << score << " bestScore: " << bestScore << endl; - if (score > bestScore) { - bestScore = score; - best = iter; - VERBOSE(2,"NEW BEST: " << score << endl); - //for (size_t i = 0; i < comps.size(); ++i) { - // cerr << comps[i] << " "; - //} - //cerr << endl; - } - } - - assert (best != nBestList.end()); - return **best; - //vector<Word> bestWords; - //GetOutputWords(**best,bestWords); - //return bestWords; -} - -} - - diff --git a/moses-cmd/LatticeMBR.h b/moses-cmd/LatticeMBR.h deleted file mode 100644 index ab8b3cb76..000000000 --- a/moses-cmd/LatticeMBR.h +++ /dev/null @@ -1,153 +0,0 @@ -/* - * LatticeMBR.h - * moses-cmd - * - * Created by Abhishek Arun on 26/01/2010. - * Copyright 2010 __MyCompanyName__. All rights reserved. - * - */ - -#ifndef moses_cmd_LatticeMBR_h -#define moses_cmd_LatticeMBR_h - -#include <map> -#include <vector> -#include <set> -#include "moses/Hypothesis.h" -#include "moses/Manager.h" -#include "moses/TrellisPathList.h" - - - -namespace MosesCmd -{ - -class Edge; - -typedef std::vector< const Moses::Hypothesis *> Lattice; -typedef std::vector<const Edge*> Path; -typedef std::map<Path, size_t> PathCounts; -typedef std::map<Moses::Phrase, PathCounts > NgramHistory; - -class Edge -{ - const Moses::Hypothesis* m_tailNode; - const Moses::Hypothesis* m_headNode; - float m_score; - Moses::TargetPhrase m_targetPhrase; - NgramHistory m_ngrams; - -public: - Edge(const Moses::Hypothesis* from, const Moses::Hypothesis* to, float score, const Moses::TargetPhrase& targetPhrase) : m_tailNode(from), m_headNode(to), m_score(score), m_targetPhrase(targetPhrase) { - //cout << "Creating new edge from Node " << from->GetId() << ", to Node : " << to->GetId() << ", score: " << score << " phrase: " << targetPhrase << endl; - } - - const Moses::Hypothesis* GetHeadNode() const { - return m_headNode; - } - - const Moses::Hypothesis* GetTailNode() const { - return m_tailNode; - } - - float GetScore() const { - return m_score; - } - - size_t GetWordsSize() const { - return m_targetPhrase.GetSize(); - } - - const Moses::Phrase& GetWords() const { - return m_targetPhrase; - } - - friend std::ostream& operator<< (std::ostream& out, const Edge& edge); - - const NgramHistory& GetNgrams( std::map<const Moses::Hypothesis*, std::vector<Edge> > & incomingEdges) ; - - bool operator < (const Edge & compare) const; - - void GetPhraseSuffix(const Moses::Phrase& origPhrase, size_t lastN, Moses::Phrase& targetPhrase) const; - - void storeNgramHistory(const Moses::Phrase& phrase, Path & path, size_t count = 1) { - m_ngrams[phrase][path]+= count; - } - -}; - -/** -* Data structure to hold the ngram scores as we traverse the lattice. Maps (hypo,ngram) to score -*/ -class NgramScores -{ -public: - NgramScores() {} - - /** logsum this score to the existing score */ - void addScore(const Moses::Hypothesis* node, const Moses::Phrase& ngram, float score); - - /** Iterate through ngrams for selected node */ - typedef std::map<const Moses::Phrase*, float>::const_iterator NodeScoreIterator; - NodeScoreIterator nodeBegin(const Moses::Hypothesis* node); - NodeScoreIterator nodeEnd(const Moses::Hypothesis* node); - -private: - std::set<Moses::Phrase> m_ngrams; - std::map<const Moses::Hypothesis*, std::map<const Moses::Phrase*, float> > m_scores; -}; - - -/** Holds a lattice mbr solution, and its scores */ -class LatticeMBRSolution -{ -public: - /** Read the words from the path */ - LatticeMBRSolution(const Moses::TrellisPath& path, bool isMap); - const std::vector<float>& GetNgramScores() const { - return m_ngramScores; - } - const std::vector<Moses::Word>& GetWords() const { - return m_words; - } - float GetMapScore() const { - return m_mapScore; - } - float GetScore() const { - return m_score; - } - - /** Initialise ngram scores */ - void CalcScore(std::map<Moses::Phrase, float>& finalNgramScores, const std::vector<float>& thetas, float mapWeight); - -private: - std::vector<Moses::Word> m_words; - float m_mapScore; - std::vector<float> m_ngramScores; - float m_score; -}; - -struct LatticeMBRSolutionComparator { - bool operator()(const LatticeMBRSolution& a, const LatticeMBRSolution& b) { - return a.GetScore() > b.GetScore(); - } -}; - -void pruneLatticeFB(Lattice & connectedHyp, std::map < const Moses::Hypothesis*, std::set <const Moses::Hypothesis* > > & outgoingHyps, std::map<const Moses::Hypothesis*, std::vector<Edge> >& incomingEdges, - const std::vector< float> & estimatedScores, const Moses::Hypothesis*, size_t edgeDensity,float scale); - -//Use the ngram scores to rerank the nbest list, return at most n solutions -void getLatticeMBRNBest(Moses::Manager& manager, Moses::TrellisPathList& nBestList, std::vector<LatticeMBRSolution>& solutions, size_t n); -//calculate expectated ngram counts, clipping at 1 (ie calculating posteriors) if posteriors==true. -void calcNgramExpectations(Lattice & connectedHyp, std::map<const Moses::Hypothesis*, std::vector<Edge> >& incomingEdges, std::map<Moses::Phrase, - float>& finalNgramScores, bool posteriors); -void GetOutputFactors(const Moses::TrellisPath &path, std::vector <Moses::Word> &translation); -void extract_ngrams(const std::vector<Moses::Word >& sentence, std::map < Moses::Phrase, int > & allngrams); -bool ascendingCoverageCmp(const Moses::Hypothesis* a, const Moses::Hypothesis* b); -std::vector<Moses::Word> doLatticeMBR(Moses::Manager& manager, Moses::TrellisPathList& nBestList); -const Moses::TrellisPath doConsensusDecoding(Moses::Manager& manager, Moses::TrellisPathList& nBestList); -//std::vector<Moses::Word> doConsensusDecoding(Moses::Manager& manager, Moses::TrellisPathList& nBestList); - -} - -#endif diff --git a/moses-cmd/LatticeMBRGrid.cpp b/moses-cmd/LatticeMBRGrid.cpp index 4c6b421fc..f00f40fd0 100644 --- a/moses-cmd/LatticeMBRGrid.cpp +++ b/moses-cmd/LatticeMBRGrid.cpp @@ -46,8 +46,8 @@ POSSIBILITY OF SUCH DAMAGE. #include <stdexcept> #include <set> -#include "IOWrapper.h" -#include "LatticeMBR.h" +#include "moses/IOWrapper.h" +#include "moses/LatticeMBR.h" #include "moses/Manager.h" #include "moses/StaticData.h" #include "util/exception.hh" @@ -55,12 +55,11 @@ POSSIBILITY OF SUCH DAMAGE. using namespace std; using namespace Moses; -using namespace MosesCmd; //keys enum gridkey {lmbr_p,lmbr_r,lmbr_prune,lmbr_scale}; -namespace MosesCmd +namespace Moses { class Grid @@ -159,7 +158,7 @@ int main(int argc, char* argv[]) StaticData& staticData = const_cast<StaticData&>(StaticData::Instance()); staticData.SetUseLatticeMBR(true); - IOWrapper* ioWrapper = GetIOWrapper(staticData); + IOWrapper* ioWrapper = IOWrapper::GetIOWrapper(staticData); if (!ioWrapper) { throw runtime_error("Failed to initialise IOWrapper"); @@ -178,10 +177,11 @@ int main(int argc, char* argv[]) const vector<float>& prune_grid = grid.getGrid(lmbr_prune); const vector<float>& scale_grid = grid.getGrid(lmbr_scale); - while(ReadInput(*ioWrapper,staticData.GetInputType(),source)) { + while(ioWrapper->ReadInput(staticData.GetInputType(),source)) { ++lineCount; - Sentence sentence; - Manager manager(lineCount, *source, staticData.GetSearchAlgorithm()); + source->SetTranslationId(lineCount); + + Manager manager(*source, staticData.GetSearchAlgorithm()); manager.ProcessSentence(); TrellisPathList nBestList; manager.CalcNBest(nBestSize, nBestList,true); @@ -200,7 +200,7 @@ int main(int argc, char* argv[]) staticData.SetMBRScale(scale); cout << lineCount << " ||| " << p << " " << r << " " << prune << " " << scale << " ||| "; vector<Word> mbrBestHypo = doLatticeMBR(manager,nBestList); - OutputBestHypo(mbrBestHypo, lineCount, staticData.GetReportSegmentation(), + ioWrapper->OutputBestHypo(mbrBestHypo, lineCount, staticData.GetReportSegmentation(), staticData.GetReportAllFactors(),cout); } } diff --git a/moses-cmd/Main.cpp b/moses-cmd/Main.cpp index a8b50db17..7e8a4a0e3 100644 --- a/moses-cmd/Main.cpp +++ b/moses-cmd/Main.cpp @@ -34,21 +34,18 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA //#include <vld.h> #endif -#include "TranslationAnalysis.h" -#include "IOWrapper.h" -#include "mbr.h" - +#include "moses/IOWrapper.h" #include "moses/Hypothesis.h" #include "moses/HypergraphOutput.h" #include "moses/Manager.h" #include "moses/StaticData.h" +#include "moses/TypeDef.h" #include "moses/Util.h" #include "moses/Timer.h" -#include "moses/ThreadPool.h" -#include "moses/OutputCollector.h" #include "moses/TranslationModel/PhraseDictionary.h" #include "moses/FF/StatefulFeatureFunction.h" #include "moses/FF/StatelessFeatureFunction.h" +#include "moses/TranslationTask.h" #ifdef HAVE_PROTOBUF #include "hypergraph.pb.h" @@ -56,368 +53,10 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA using namespace std; using namespace Moses; -using namespace MosesCmd; - -namespace MosesCmd -{ -// output floats with five significant digits -static const size_t PRECISION = 3; - -/** Enforce rounding */ -void fix(std::ostream& stream, size_t size) -{ - stream.setf(std::ios::fixed); - stream.precision(size); -} -/** Translates a sentence. - * - calls the search (Manager) - * - applies the decision rule - * - outputs best translation and additional reporting - **/ -class TranslationTask : public Task +namespace Moses { -public: - - TranslationTask(size_t lineNumber, - InputType* source, OutputCollector* outputCollector, OutputCollector* nbestCollector, - OutputCollector* latticeSamplesCollector, - OutputCollector* wordGraphCollector, OutputCollector* searchGraphCollector, - OutputCollector* detailedTranslationCollector, - OutputCollector* alignmentInfoCollector, - OutputCollector* unknownsCollector, - bool outputSearchGraphSLF, - boost::shared_ptr<HypergraphOutput<Manager> > hypergraphOutput) : - m_source(source), m_lineNumber(lineNumber), - m_outputCollector(outputCollector), m_nbestCollector(nbestCollector), - m_latticeSamplesCollector(latticeSamplesCollector), - m_wordGraphCollector(wordGraphCollector), m_searchGraphCollector(searchGraphCollector), - m_detailedTranslationCollector(detailedTranslationCollector), - m_alignmentInfoCollector(alignmentInfoCollector), - m_unknownsCollector(unknownsCollector), - m_outputSearchGraphSLF(outputSearchGraphSLF), - m_hypergraphOutput(hypergraphOutput) {} - - /** Translate one sentence - * gets called by main function implemented at end of this source file */ - void Run() { - // shorthand for "global data" - const StaticData &staticData = StaticData::Instance(); - - // input sentence - Sentence sentence; - - // report wall time spent on translation - Timer translationTime; - translationTime.start(); - - // report thread number -#if defined(WITH_THREADS) && defined(BOOST_HAS_PTHREADS) - TRACE_ERR("Translating line " << m_lineNumber << " in thread id " << pthread_self() << std::endl); -#endif - - - // execute the translation - // note: this executes the search, resulting in a search graph - // we still need to apply the decision rule (MAP, MBR, ...) - Timer initTime; - initTime.start(); - Manager manager(m_lineNumber, *m_source,staticData.GetSearchAlgorithm()); - VERBOSE(1, "Line " << m_lineNumber << ": Initialize search took " << initTime << " seconds total" << endl); - manager.ProcessSentence(); - - // we are done with search, let's look what we got - Timer additionalReportingTime; - additionalReportingTime.start(); - - // output word graph - if (m_wordGraphCollector) { - ostringstream out; - fix(out,PRECISION); - manager.GetWordGraph(m_lineNumber, out); - m_wordGraphCollector->Write(m_lineNumber, out.str()); - } - - // output search graph - if (m_searchGraphCollector) { - ostringstream out; - fix(out,PRECISION); - manager.OutputSearchGraph(m_lineNumber, out); - m_searchGraphCollector->Write(m_lineNumber, out.str()); - -#ifdef HAVE_PROTOBUF - if (staticData.GetOutputSearchGraphPB()) { - ostringstream sfn; - sfn << staticData.GetParam("output-search-graph-pb")[0] << '/' << m_lineNumber << ".pb" << ends; - string fn = sfn.str(); - VERBOSE(2, "Writing search graph to " << fn << endl); - fstream output(fn.c_str(), ios::trunc | ios::binary | ios::out); - manager.SerializeSearchGraphPB(m_lineNumber, output); - } -#endif - } - - // Output search graph in HTK standard lattice format (SLF) - if (m_outputSearchGraphSLF) { - stringstream fileName; - fileName << staticData.GetParam("output-search-graph-slf")[0] << "/" << m_lineNumber << ".slf"; - std::ofstream *file = new std::ofstream; - file->open(fileName.str().c_str()); - if (file->is_open() && file->good()) { - ostringstream out; - fix(out,PRECISION); - manager.OutputSearchGraphAsSLF(m_lineNumber, out); - *file << out.str(); - file -> flush(); - } else { - TRACE_ERR("Cannot output HTK standard lattice for line " << m_lineNumber << " because the output file is not open or not ready for writing" << std::endl); - } - delete file; - } - - // Output search graph in hypergraph format for Kenneth Heafield's lazy hypergraph decoder - if (m_hypergraphOutput.get()) { - m_hypergraphOutput->Write(manager); - } - - additionalReportingTime.stop(); - - // apply decision rule and output best translation(s) - if (m_outputCollector) { - ostringstream out; - ostringstream debug; - fix(debug,PRECISION); - - // all derivations - send them to debug stream - if (staticData.PrintAllDerivations()) { - additionalReportingTime.start(); - manager.PrintAllDerivations(m_lineNumber, debug); - additionalReportingTime.stop(); - } - - Timer decisionRuleTime; - decisionRuleTime.start(); - - // MAP decoding: best hypothesis - const Hypothesis* bestHypo = NULL; - if (!staticData.UseMBR()) { - bestHypo = manager.GetBestHypothesis(); - if (bestHypo) { - if (StaticData::Instance().GetOutputHypoScore()) { - out << bestHypo->GetTotalScore() << ' '; - } - if (staticData.IsPathRecoveryEnabled()) { - OutputInput(out, bestHypo); - out << "||| "; - } - if (staticData.GetParam("print-id").size() && Scan<bool>(staticData.GetParam("print-id")[0]) ) { - out << m_source->GetTranslationId() << " "; - } - - if (staticData.GetReportSegmentation() == 2) { - manager.GetOutputLanguageModelOrder(out, bestHypo); - } - OutputBestSurface( - out, - bestHypo, - staticData.GetOutputFactorOrder(), - staticData.GetReportSegmentation(), - staticData.GetReportAllFactors()); - if (staticData.PrintAlignmentInfo()) { - out << "||| "; - OutputAlignment(out, bestHypo); - } - - OutputAlignment(m_alignmentInfoCollector, m_lineNumber, bestHypo); - IFVERBOSE(1) { - debug << "BEST TRANSLATION: " << *bestHypo << endl; - } - } else { - VERBOSE(1, "NO BEST TRANSLATION" << endl); - } - - out << endl; - } - - // MBR decoding (n-best MBR, lattice MBR, consensus) - else { - // we first need the n-best translations - size_t nBestSize = staticData.GetMBRSize(); - if (nBestSize <= 0) { - cerr << "ERROR: negative size for number of MBR candidate translations not allowed (option mbr-size)" << endl; - exit(1); - } - TrellisPathList nBestList; - manager.CalcNBest(nBestSize, nBestList,true); - VERBOSE(2,"size of n-best: " << nBestList.GetSize() << " (" << nBestSize << ")" << endl); - IFVERBOSE(2) { - PrintUserTime("calculated n-best list for (L)MBR decoding"); - } - - // lattice MBR - if (staticData.UseLatticeMBR()) { - if (m_nbestCollector) { - //lattice mbr nbest - vector<LatticeMBRSolution> solutions; - size_t n = min(nBestSize, staticData.GetNBestSize()); - getLatticeMBRNBest(manager,nBestList,solutions,n); - ostringstream out; - OutputLatticeMBRNBest(out, solutions,m_lineNumber); - m_nbestCollector->Write(m_lineNumber, out.str()); - } else { - //Lattice MBR decoding - vector<Word> mbrBestHypo = doLatticeMBR(manager,nBestList); - OutputBestHypo(mbrBestHypo, m_lineNumber, staticData.GetReportSegmentation(), - staticData.GetReportAllFactors(),out); - IFVERBOSE(2) { - PrintUserTime("finished Lattice MBR decoding"); - } - } - } - - // consensus decoding - else if (staticData.UseConsensusDecoding()) { - const TrellisPath &conBestHypo = doConsensusDecoding(manager,nBestList); - OutputBestHypo(conBestHypo, m_lineNumber, - staticData.GetReportSegmentation(), - staticData.GetReportAllFactors(),out); - OutputAlignment(m_alignmentInfoCollector, m_lineNumber, conBestHypo); - IFVERBOSE(2) { - PrintUserTime("finished Consensus decoding"); - } - } - - // n-best MBR decoding - else { - const Moses::TrellisPath &mbrBestHypo = doMBR(nBestList); - OutputBestHypo(mbrBestHypo, m_lineNumber, - staticData.GetReportSegmentation(), - staticData.GetReportAllFactors(),out); - OutputAlignment(m_alignmentInfoCollector, m_lineNumber, mbrBestHypo); - IFVERBOSE(2) { - PrintUserTime("finished MBR decoding"); - } - } - } - - // report best translation to output collector - m_outputCollector->Write(m_lineNumber,out.str(),debug.str()); - - decisionRuleTime.stop(); - VERBOSE(1, "Line " << m_lineNumber << ": Decision rule took " << decisionRuleTime << " seconds total" << endl); - } - - additionalReportingTime.start(); - - // output n-best list - if (m_nbestCollector && !staticData.UseLatticeMBR()) { - TrellisPathList nBestList; - ostringstream out; - manager.CalcNBest(staticData.GetNBestSize(), nBestList,staticData.GetDistinctNBest()); - OutputNBest(out, nBestList, staticData.GetOutputFactorOrder(), m_lineNumber, - staticData.GetReportSegmentation()); - m_nbestCollector->Write(m_lineNumber, out.str()); - } - - //lattice samples - if (m_latticeSamplesCollector) { - TrellisPathList latticeSamples; - ostringstream out; - manager.CalcLatticeSamples(staticData.GetLatticeSamplesSize(), latticeSamples); - OutputNBest(out,latticeSamples, staticData.GetOutputFactorOrder(), m_lineNumber, - staticData.GetReportSegmentation()); - m_latticeSamplesCollector->Write(m_lineNumber, out.str()); - } - - // detailed translation reporting - if (m_detailedTranslationCollector) { - ostringstream out; - fix(out,PRECISION); - TranslationAnalysis::PrintTranslationAnalysis(out, manager.GetBestHypothesis()); - m_detailedTranslationCollector->Write(m_lineNumber,out.str()); - } - - //list of unknown words - if (m_unknownsCollector) { - const vector<const Phrase*>& unknowns = manager.getSntTranslationOptions()->GetUnknownSources(); - ostringstream out; - for (size_t i = 0; i < unknowns.size(); ++i) { - out << *(unknowns[i]); - } - out << endl; - m_unknownsCollector->Write(m_lineNumber, out.str()); - } - - // report additional statistics - manager.CalcDecoderStatistics(); - VERBOSE(1, "Line " << m_lineNumber << ": Additional reporting took " << additionalReportingTime << " seconds total" << endl); - VERBOSE(1, "Line " << m_lineNumber << ": Translation took " << translationTime << " seconds total" << endl); - IFVERBOSE(2) { - PrintUserTime("Sentence Decoding Time:"); - } - } - - ~TranslationTask() { - delete m_source; - } - -private: - InputType* m_source; - size_t m_lineNumber; - OutputCollector* m_outputCollector; - OutputCollector* m_nbestCollector; - OutputCollector* m_latticeSamplesCollector; - OutputCollector* m_wordGraphCollector; - OutputCollector* m_searchGraphCollector; - OutputCollector* m_detailedTranslationCollector; - OutputCollector* m_alignmentInfoCollector; - OutputCollector* m_unknownsCollector; - bool m_outputSearchGraphSLF; - boost::shared_ptr<HypergraphOutput<Manager> > m_hypergraphOutput; - std::ofstream *m_alignmentStream; - - -}; - -static void PrintFeatureWeight(const FeatureFunction* ff) -{ - cout << ff->GetScoreProducerDescription() << "="; - size_t numScoreComps = ff->GetNumScoreComponents(); - vector<float> values = StaticData::Instance().GetAllWeights().GetScoresForProducer(ff); - for (size_t i = 0; i < numScoreComps; ++i) { - cout << " " << values[i]; - } - cout << endl; -} - -static void ShowWeights() -{ - //TODO: Find a way of ensuring this order is synced with the nbest - fix(cout,6); - const vector<const StatelessFeatureFunction*>& slf = StatelessFeatureFunction::GetStatelessFeatureFunctions(); - const vector<const StatefulFeatureFunction*>& sff = StatefulFeatureFunction::GetStatefulFeatureFunctions(); - - for (size_t i = 0; i < sff.size(); ++i) { - const StatefulFeatureFunction *ff = sff[i]; - if (ff->IsTuneable()) { - PrintFeatureWeight(ff); - } - else { - cout << ff->GetScoreProducerDescription() << " UNTUNEABLE" << endl; - } - } - for (size_t i = 0; i < slf.size(); ++i) { - const StatelessFeatureFunction *ff = slf[i]; - if (ff->IsTuneable()) { - PrintFeatureWeight(ff); - } - else { - cout << ff->GetScoreProducerDescription() << " UNTUNEABLE" << endl; - } - } -} - void OutputFeatureWeightsForHypergraph(std::ostream &outputSearchGraphStream) { outputSearchGraphStream.setf(std::ios::fixed); @@ -445,8 +84,8 @@ int main(int argc, char** argv) } // set number of significant decimals in output - fix(cout,PRECISION); - fix(cerr,PRECISION); + IOWrapper::FixPrecision(cout); + IOWrapper::FixPrecision(cerr); // load all the settings into the Parameter class // (stores them as strings, or array of strings) @@ -476,7 +115,7 @@ int main(int argc, char** argv) srand(time(NULL)); // set up read/writing class - IOWrapper* ioWrapper = GetIOWrapper(staticData); + IOWrapper* ioWrapper = IOWrapper::GetIOWrapper(staticData); if (!ioWrapper) { cerr << "Error; Failed to create IO object" << endl; exit(1); @@ -489,92 +128,17 @@ int main(int argc, char** argv) TRACE_ERR(weights); TRACE_ERR("\n"); } - boost::shared_ptr<HypergraphOutput<Manager> > hypergraphOutput; - if (staticData.GetOutputSearchGraphHypergraph()) { - hypergraphOutput.reset(new HypergraphOutput<Manager>(PRECISION)); - } + boost::shared_ptr<HypergraphOutput<Manager> > hypergraphOutput; + boost::shared_ptr<HypergraphOutput<ChartManager> > hypergraphOutputChart; - // initialize output streams - // note: we can't just write to STDOUT or files - // because multithreading may return sentences in shuffled order - auto_ptr<OutputCollector> outputCollector; // for translations - auto_ptr<OutputCollector> nbestCollector; // for n-best lists - auto_ptr<OutputCollector> latticeSamplesCollector; //for lattice samples - auto_ptr<ofstream> nbestOut; - auto_ptr<ofstream> latticeSamplesOut; - size_t nbestSize = staticData.GetNBestSize(); - string nbestFile = staticData.GetNBestFilePath(); - bool output1best = true; - if (nbestSize) { - if (nbestFile == "-" || nbestFile == "/dev/stdout") { - // nbest to stdout, no 1-best - nbestCollector.reset(new OutputCollector()); - output1best = false; - } else { - // nbest to file, 1-best to stdout - nbestOut.reset(new ofstream(nbestFile.c_str())); - if (!nbestOut->good()) { - TRACE_ERR("ERROR: Failed to open " << nbestFile << " for nbest lists" << endl); - exit(1); - } - nbestCollector.reset(new OutputCollector(nbestOut.get())); - } - } - size_t latticeSamplesSize = staticData.GetLatticeSamplesSize(); - string latticeSamplesFile = staticData.GetLatticeSamplesFilePath(); - if (latticeSamplesSize) { - if (latticeSamplesFile == "-" || latticeSamplesFile == "/dev/stdout") { - latticeSamplesCollector.reset(new OutputCollector()); - output1best = false; - } else { - latticeSamplesOut.reset(new ofstream(latticeSamplesFile.c_str())); - if (!latticeSamplesOut->good()) { - TRACE_ERR("ERROR: Failed to open " << latticeSamplesFile << " for lattice samples" << endl); - exit(1); - } - latticeSamplesCollector.reset(new OutputCollector(latticeSamplesOut.get())); - } - } - if (output1best) { - outputCollector.reset(new OutputCollector()); - } - - // initialize stream for word graph (aka: output lattice) - auto_ptr<OutputCollector> wordGraphCollector; - if (staticData.GetOutputWordGraph()) { - wordGraphCollector.reset(new OutputCollector(&(ioWrapper->GetOutputWordGraphStream()))); - } - - // initialize stream for search graph - // note: this is essentially the same as above, but in a different format - auto_ptr<OutputCollector> searchGraphCollector; - if (staticData.GetOutputSearchGraph()) { - searchGraphCollector.reset(new OutputCollector(&(ioWrapper->GetOutputSearchGraphStream()))); - } - - // initialize stram for details about the decoder run - auto_ptr<OutputCollector> detailedTranslationCollector; - if (staticData.IsDetailedTranslationReportingEnabled()) { - detailedTranslationCollector.reset(new OutputCollector(&(ioWrapper->GetDetailedTranslationReportingStream()))); - } - - // initialize stram for word alignment between input and output - auto_ptr<OutputCollector> alignmentInfoCollector; - if (!staticData.GetAlignmentOutputFile().empty()) { - alignmentInfoCollector.reset(new OutputCollector(ioWrapper->GetAlignmentOutputStream())); - } - - //initialise stream for unknown (oov) words - auto_ptr<OutputCollector> unknownsCollector; - auto_ptr<ofstream> unknownsStream; - if (!staticData.GetOutputUnknownsFile().empty()) { - unknownsStream.reset(new ofstream(staticData.GetOutputUnknownsFile().c_str())); - if (!unknownsStream->good()) { - TRACE_ERR("Unable to open " << staticData.GetOutputUnknownsFile() << " for unknowns"); - exit(1); - } - unknownsCollector.reset(new OutputCollector(unknownsStream.get())); + if (staticData.GetOutputSearchGraphHypergraph()) { + if (staticData.IsChart()) { + hypergraphOutputChart.reset(new HypergraphOutput<ChartManager>(PRECISION)); + } + else { + hypergraphOutput.reset(new HypergraphOutput<Manager>(PRECISION)); + } } #ifdef WITH_THREADS @@ -584,7 +148,8 @@ int main(int argc, char** argv) // main loop over set of input sentences InputType* source = NULL; size_t lineCount = staticData.GetStartTranslationId(); - while(ReadInput(*ioWrapper,staticData.GetInputType(),source)) { + while(ioWrapper->ReadInput(staticData.GetInputType(),source)) { + source->SetTranslationId(lineCount); IFVERBOSE(1) { ResetUserTime(); } @@ -592,17 +157,18 @@ int main(int argc, char** argv) FeatureFunction::CallChangeSource(source); // set up task of translating one sentence - TranslationTask* task = - new TranslationTask(lineCount,source, outputCollector.get(), - nbestCollector.get(), - latticeSamplesCollector.get(), - wordGraphCollector.get(), - searchGraphCollector.get(), - detailedTranslationCollector.get(), - alignmentInfoCollector.get(), - unknownsCollector.get(), - staticData.GetOutputSearchGraphSLF(), - hypergraphOutput); + TranslationTask* task; + if (staticData.IsChart()) { + // scfg + task = new TranslationTask(source, *ioWrapper, hypergraphOutputChart); + } + else { + // pb + task = new TranslationTask(source, *ioWrapper, + staticData.GetOutputSearchGraphSLF(), + hypergraphOutput); + } + // execute task #ifdef WITH_THREADS pool.Submit(task); diff --git a/moses-cmd/Main.h b/moses-cmd/Main.h index 362c1f245..49fee0219 100644 --- a/moses-cmd/Main.h +++ b/moses-cmd/Main.h @@ -1,3 +1,4 @@ +#pragma once // $Id$ /*********************************************************************** @@ -32,12 +33,10 @@ POSSIBILITY OF SUCH DAMAGE. // example file on how to use moses library -#ifndef moses_cmd_Main_h -#define moses_cmd_Main_h #include "moses/StaticData.h" class IOWrapper; int main(int argc, char* argv[]); -#endif + diff --git a/moses-cmd/TranslationAnalysis.cpp b/moses-cmd/TranslationAnalysis.cpp deleted file mode 100644 index e77486162..000000000 --- a/moses-cmd/TranslationAnalysis.cpp +++ /dev/null @@ -1,137 +0,0 @@ -// $Id$ - -#include <iostream> -#include <sstream> -#include <algorithm> -#include "moses/StaticData.h" -#include "moses/Hypothesis.h" -#include "TranslationAnalysis.h" -#include "moses/FF/StatefulFeatureFunction.h" -#include "moses/FF/StatelessFeatureFunction.h" -#include "moses/LM/Base.h" - -using namespace Moses; - -namespace TranslationAnalysis -{ - -void PrintTranslationAnalysis(std::ostream &os, const Hypothesis* hypo) -{ - os << std::endl << "TRANSLATION HYPOTHESIS DETAILS:" << std::endl; - std::vector<const Hypothesis*> translationPath; - - while (hypo) { - translationPath.push_back(hypo); - hypo = hypo->GetPrevHypo(); - } - - std::reverse(translationPath.begin(), translationPath.end()); - std::vector<std::string> droppedWords; - std::vector<const Hypothesis*>::iterator tpi = translationPath.begin(); - if(tpi == translationPath.end()) - return; - ++tpi; // skip initial translation state - std::vector<std::string> sourceMap; - std::vector<std::string> targetMap; - std::vector<unsigned int> lmAcc(0); - size_t lmCalls = 0; - bool doLMStats = ((*tpi)->GetLMStats() != 0); - if (doLMStats) - lmAcc.resize((*tpi)->GetLMStats()->size(), 0); - for (; tpi != translationPath.end(); ++tpi) { - std::ostringstream sms; - std::ostringstream tms; - std::string target = (*tpi)->GetTargetPhraseStringRep(); - std::string source = (*tpi)->GetSourcePhraseStringRep(); - WordsRange twr = (*tpi)->GetCurrTargetWordsRange(); - WordsRange swr = (*tpi)->GetCurrSourceWordsRange(); - const AlignmentInfo &alignmentInfo = (*tpi)->GetCurrTargetPhrase().GetAlignTerm(); - // language model backoff stats, - if (doLMStats) { - std::vector<std::vector<unsigned int> >& lmstats = *(*tpi)->GetLMStats(); - std::vector<std::vector<unsigned int> >::iterator i = lmstats.begin(); - std::vector<unsigned int>::iterator acc = lmAcc.begin(); - - for (; i != lmstats.end(); ++i, ++acc) { - std::vector<unsigned int>::iterator j = i->begin(); - lmCalls += i->size(); - for (; j != i->end(); ++j) { - (*acc) += *j; - } - } - } - - bool epsilon = false; - if (target == "") { - target="<EPSILON>"; - epsilon = true; - droppedWords.push_back(source); - } - os << " SOURCE: " << swr << " " << source << std::endl - << " TRANSLATED AS: " << target << std::endl - << " WORD ALIGNED: " << alignmentInfo << std::endl; - size_t twr_i = twr.GetStartPos(); - size_t swr_i = swr.GetStartPos(); - if (!epsilon) { - sms << twr_i; - } - if (epsilon) { - tms << "del(" << swr_i << ")"; - } else { - tms << swr_i; - } - swr_i++; - twr_i++; - for (; twr_i <= twr.GetEndPos() && twr.GetEndPos() != NOT_FOUND; twr_i++) { - sms << '-' << twr_i; - } - for (; swr_i <= swr.GetEndPos() && swr.GetEndPos() != NOT_FOUND; swr_i++) { - tms << '-' << swr_i; - } - if (!epsilon) targetMap.push_back(sms.str()); - sourceMap.push_back(tms.str()); - } - std::vector<std::string>::iterator si = sourceMap.begin(); - std::vector<std::string>::iterator ti = targetMap.begin(); - os << std::endl << "SOURCE/TARGET SPANS:"; - os << std::endl << " SOURCE:"; - for (; si != sourceMap.end(); ++si) { - os << " " << *si; - } - os << std::endl << " TARGET:"; - for (; ti != targetMap.end(); ++ti) { - os << " " << *ti; - } - os << std::endl << std::endl; - if (doLMStats && lmCalls > 0) { - std::vector<unsigned int>::iterator acc = lmAcc.begin(); - - const std::vector<const StatefulFeatureFunction*> &statefulFFs = StatefulFeatureFunction::GetStatefulFeatureFunctions(); - for (size_t i = 0; i < statefulFFs.size(); ++i) { - const StatefulFeatureFunction *ff = statefulFFs[i]; - const LanguageModel *lm = dynamic_cast<const LanguageModel*>(ff); - - if (lm) { - char buf[256]; - sprintf(buf, "%.4f", (float)(*acc)/(float)lmCalls); - os << lm->GetScoreProducerDescription() <<", AVG N-GRAM LENGTH: " << buf << std::endl; - - ++acc; - } - } - } - - if (droppedWords.size() > 0) { - std::vector<std::string>::iterator dwi = droppedWords.begin(); - os << std::endl << "WORDS/PHRASES DROPPED:" << std::endl; - for (; dwi != droppedWords.end(); ++dwi) { - os << "\tdropped=" << *dwi << std::endl; - } - } - os << std::endl << "SCORES (UNWEIGHTED/WEIGHTED): "; - os << translationPath.back()->GetScoreBreakdown(); - os << " weighted(TODO)"; - os << std::endl; -} - -} diff --git a/moses-cmd/TranslationAnalysis.h b/moses-cmd/TranslationAnalysis.h deleted file mode 100644 index 348cfe512..000000000 --- a/moses-cmd/TranslationAnalysis.h +++ /dev/null @@ -1,24 +0,0 @@ -// $Id$ - -/* - * also see moses/SentenceStats - */ - -#ifndef moses_cmd_TranslationAnalysis_h -#define moses_cmd_TranslationAnalysis_h - -#include <iostream> -#include "moses/Hypothesis.h" - -namespace TranslationAnalysis -{ - -/*** - * print details about the translation represented in hypothesis to - * os. Included information: phrase alignment, words dropped, scores - */ -void PrintTranslationAnalysis(std::ostream &os, const Moses::Hypothesis* hypo); - -} - -#endif diff --git a/moses-cmd/mbr.cpp b/moses-cmd/mbr.cpp deleted file mode 100644 index 6a8dfa823..000000000 --- a/moses-cmd/mbr.cpp +++ /dev/null @@ -1,178 +0,0 @@ -#include <iostream> -#include <fstream> -#include <sstream> -#include <iomanip> -#include <vector> -#include <map> -#include <stdlib.h> -#include <math.h> -#include <algorithm> -#include <stdio.h> -#include "moses/TrellisPathList.h" -#include "moses/TrellisPath.h" -#include "moses/StaticData.h" -#include "moses/Util.h" -#include "mbr.h" - -using namespace std ; -using namespace Moses; - - -/* Input : - 1. a sorted n-best list, with duplicates filtered out in the following format - 0 ||| amr moussa is currently on a visit to libya , tomorrow , sunday , to hold talks with regard to the in sudan . ||| 0 -4.94418 0 0 -2.16036 0 0 -81.4462 -106.593 -114.43 -105.55 -12.7873 -26.9057 -25.3715 -52.9336 7.99917 -24 ||| -4.58432 - - 2. a weight vector - 3. bleu order ( default = 4) - 4. scaling factor to weigh the weight vector (default = 1.0) - - Output : - translations that minimise the Bayes Risk of the n-best list - - -*/ - -int BLEU_ORDER = 4; -int SMOOTH = 1; -float min_interval = 1e-4; -void extract_ngrams(const vector<const Factor* >& sentence, map < vector < const Factor* >, int > & allngrams) -{ - vector< const Factor* > ngram; - for (int k = 0; k < BLEU_ORDER; k++) { - for(int i =0; i < max((int)sentence.size()-k,0); i++) { - for ( int j = i; j<= i+k; j++) { - ngram.push_back(sentence[j]); - } - ++allngrams[ngram]; - ngram.clear(); - } - } -} - -float calculate_score(const vector< vector<const Factor*> > & sents, int ref, int hyp, vector < map < vector < const Factor *>, int > > & ngram_stats ) -{ - int comps_n = 2*BLEU_ORDER+1; - vector<int> comps(comps_n); - float logbleu = 0.0, brevity; - - int hyp_length = sents[hyp].size(); - - for (int i =0; i<BLEU_ORDER; i++) { - comps[2*i] = 0; - comps[2*i+1] = max(hyp_length-i,0); - } - - map< vector < const Factor * > ,int > & hyp_ngrams = ngram_stats[hyp] ; - map< vector < const Factor * >, int > & ref_ngrams = ngram_stats[ref] ; - - for (map< vector< const Factor * >, int >::iterator it = hyp_ngrams.begin(); - it != hyp_ngrams.end(); it++) { - map< vector< const Factor * >, int >::iterator ref_it = ref_ngrams.find(it->first); - if(ref_it != ref_ngrams.end()) { - comps[2* (it->first.size()-1)] += min(ref_it->second,it->second); - } - } - comps[comps_n-1] = sents[ref].size(); - - for (int i=0; i<BLEU_ORDER; i++) { - if (comps[0] == 0) - return 0.0; - if ( i > 0 ) - logbleu += log((float)comps[2*i]+SMOOTH)-log((float)comps[2*i+1]+SMOOTH); - else - logbleu += log((float)comps[2*i])-log((float)comps[2*i+1]); - } - logbleu /= BLEU_ORDER; - brevity = 1.0-(float)comps[comps_n-1]/comps[1]; // comps[comps_n-1] is the ref length, comps[1] is the test length - if (brevity < 0.0) - logbleu += brevity; - return exp(logbleu); -} - -const TrellisPath doMBR(const TrellisPathList& nBestList) -{ - float marginal = 0; - - vector<float> joint_prob_vec; - vector< vector<const Factor*> > translations; - float joint_prob; - vector< map < vector <const Factor *>, int > > ngram_stats; - - TrellisPathList::const_iterator iter; - - // get max score to prevent underflow - float maxScore = -1e20; - for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) { - const TrellisPath &path = **iter; - float score = StaticData::Instance().GetMBRScale() - * path.GetScoreBreakdown().GetWeightedScore(); - if (maxScore < score) maxScore = score; - } - - for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) { - const TrellisPath &path = **iter; - joint_prob = UntransformScore(StaticData::Instance().GetMBRScale() * path.GetScoreBreakdown().GetWeightedScore() - maxScore); - marginal += joint_prob; - joint_prob_vec.push_back(joint_prob); - - // get words in translation - vector<const Factor*> translation; - GetOutputFactors(path, translation); - - // collect n-gram counts - map < vector < const Factor *>, int > counts; - extract_ngrams(translation,counts); - - ngram_stats.push_back(counts); - translations.push_back(translation); - } - - vector<float> mbr_loss; - float bleu, weightedLoss; - float weightedLossCumul = 0; - float minMBRLoss = 1000000; - int minMBRLossIdx = -1; - - /* Main MBR computation done here */ - iter = nBestList.begin(); - for (unsigned int i = 0; i < nBestList.GetSize(); i++) { - weightedLossCumul = 0; - for (unsigned int j = 0; j < nBestList.GetSize(); j++) { - if ( i != j) { - bleu = calculate_score(translations, j, i,ngram_stats ); - weightedLoss = ( 1 - bleu) * ( joint_prob_vec[j]/marginal); - weightedLossCumul += weightedLoss; - if (weightedLossCumul > minMBRLoss) - break; - } - } - if (weightedLossCumul < minMBRLoss) { - minMBRLoss = weightedLossCumul; - minMBRLossIdx = i; - } - iter++; - } - /* Find sentence that minimises Bayes Risk under 1- BLEU loss */ - return nBestList.at(minMBRLossIdx); - //return translations[minMBRLossIdx]; -} - -void GetOutputFactors(const TrellisPath &path, vector <const Factor*> &translation) -{ - const std::vector<const Hypothesis *> &edges = path.GetEdges(); - const std::vector<FactorType>& outputFactorOrder = StaticData::Instance().GetOutputFactorOrder(); - assert (outputFactorOrder.size() == 1); - - // print the surface factor of the translation - for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) { - const Hypothesis &edge = *edges[currEdge]; - const Phrase &phrase = edge.GetCurrTargetPhrase(); - size_t size = phrase.GetSize(); - for (size_t pos = 0 ; pos < size ; pos++) { - - const Factor *factor = phrase.GetFactor(pos, outputFactorOrder[0]); - translation.push_back(factor); - } - } -} - diff --git a/moses-cmd/mbr.h b/moses-cmd/mbr.h deleted file mode 100644 index d08b11a98..000000000 --- a/moses-cmd/mbr.h +++ /dev/null @@ -1,28 +0,0 @@ -// $Id$ - -/*********************************************************************** -Moses - factored phrase-based language decoder -Copyright (C) 2006 University of Edinburgh - -This library is free software; you can redistribute it and/or -modify it under the terms of the GNU Lesser General Public -License as published by the Free Software Foundation; either -version 2.1 of the License, or (at your option) any later version. - -This library is distributed in the hope that it will be useful, -but WITHOUT ANY WARRANTY; without even the implied warranty of -MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU -Lesser General Public License for more details. - -You should have received a copy of the GNU Lesser General Public -License along with this library; if not, write to the Free Software -Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA -***********************************************************************/ - -#ifndef moses_cmd_mbr_h -#define moses_cmd_mbr_h - -const Moses::TrellisPath doMBR(const Moses::TrellisPathList& nBestList); -void GetOutputFactors(const Moses::TrellisPath &path, std::vector <const Moses::Factor*> &translation); -float calculate_score(const std::vector< std::vector<const Moses::Factor*> > & sents, int ref, int hyp, std::vector < std::map < std::vector < const Moses::Factor *>, int > > & ngram_stats ); -#endif diff --git a/moses-cmd/simulate-pe.cc b/moses-cmd/simulate-pe.cc index 5384d9886..daba1a715 100644 --- a/moses-cmd/simulate-pe.cc +++ b/moses-cmd/simulate-pe.cc @@ -1,5 +1,4 @@ -// Fork of Main.cpp, to simulate post-editing sessions. -// Written by Ulrich Germann. +// $Id: MainMT.cpp 3045 2010-04-05 13:07:29Z hieuhoang1972 $ /*********************************************************************** Moses - factored phrase-based language decoder @@ -20,44 +19,41 @@ License along with this library; if not, write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA ***********************************************************************/ -#include <boost/algorithm/string/predicate.hpp> -#include <boost/filesystem.hpp> -#include <boost/iostreams/device/file.hpp> -#include <boost/iostreams/filter/bzip2.hpp> -#include <boost/iostreams/filter/gzip.hpp> -#include <boost/iostreams/filtering_stream.hpp> -#include <boost/foreach.hpp> - +/** + * Moses main, for single-threaded and multi-threaded. + **/ #include <exception> #include <fstream> #include <sstream> #include <vector> +#include <boost/program_options.hpp> #include "util/usage.hh" #include "util/exception.hh" #include "moses/Util.h" + +#ifdef PT_UG #include "moses/TranslationModel/UG/mmsapt.h" #include "moses/TranslationModel/UG/generic/program_options/ug_splice_arglist.h" +#endif #ifdef WIN32 // Include Visual Leak Detector //#include <vld.h> #endif -#include "TranslationAnalysis.h" -#include "IOWrapper.h" -#include "mbr.h" - +#include "moses/IOWrapper.h" #include "moses/Hypothesis.h" +#include "moses/HypergraphOutput.h" #include "moses/Manager.h" #include "moses/StaticData.h" +#include "moses/TypeDef.h" #include "moses/Util.h" #include "moses/Timer.h" -#include "moses/ThreadPool.h" -#include "moses/OutputCollector.h" #include "moses/TranslationModel/PhraseDictionary.h" #include "moses/FF/StatefulFeatureFunction.h" #include "moses/FF/StatelessFeatureFunction.h" +#include "moses/TranslationTask.h" #ifdef HAVE_PROTOBUF #include "hypergraph.pb.h" @@ -65,522 +61,15 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA using namespace std; using namespace Moses; -using namespace MosesCmd; - -namespace MosesCmd -{ -// output floats with five significant digits -static const size_t PRECISION = 3; - -/** Enforce rounding */ -void fix(std::ostream& stream, size_t size) -{ - stream.setf(std::ios::fixed); - stream.precision(size); -} -/** Translates a sentence. - * - calls the search (Manager) - * - applies the decision rule - * - outputs best translation and additional reporting - **/ -class TranslationTask : public Task +namespace Moses { -public: - - TranslationTask(size_t lineNumber, - InputType* source, - OutputCollector* outputCollector, - OutputCollector* nbestCollector, - OutputCollector* latticeSamplesCollector, - OutputCollector* wordGraphCollector, - OutputCollector* searchGraphCollector, - OutputCollector* detailedTranslationCollector, - OutputCollector* alignmentInfoCollector, - OutputCollector* unknownsCollector, - bool outputSearchGraphSLF, - bool outputSearchGraphHypergraph) - : m_source(source) - , m_lineNumber(lineNumber) - , m_outputCollector(outputCollector) - , m_nbestCollector(nbestCollector) - , m_latticeSamplesCollector(latticeSamplesCollector) - , m_wordGraphCollector(wordGraphCollector) - , m_searchGraphCollector(searchGraphCollector) - , m_detailedTranslationCollector(detailedTranslationCollector) - , m_alignmentInfoCollector(alignmentInfoCollector) - , m_unknownsCollector(unknownsCollector) - , m_outputSearchGraphSLF(outputSearchGraphSLF) - , m_outputSearchGraphHypergraph(outputSearchGraphHypergraph) - { } - - /** Translate one sentence - * gets called by main function implemented at end of this source file */ - void Run() { - // shorthand for "global data" - const StaticData &staticData = StaticData::Instance(); - - // input sentence - Sentence sentence; - - // report wall time spent on translation - Timer translationTime; - translationTime.start(); - - // report thread number -#if defined(WITH_THREADS) && defined(BOOST_HAS_PTHREADS) - TRACE_ERR("Translating line " << m_lineNumber << " in thread id " << pthread_self() << std::endl); -#endif - - - // execute the translation - // note: this executes the search, resulting in a search graph - // we still need to apply the decision rule (MAP, MBR, ...) - Timer initTime; - initTime.start(); - Manager manager(m_lineNumber, *m_source,staticData.GetSearchAlgorithm()); - VERBOSE(1, "Line " << m_lineNumber << ": Initialize search took " << initTime << " seconds total" << endl); - manager.ProcessSentence(); - - // we are done with search, let's look what we got - Timer additionalReportingTime; - additionalReportingTime.start(); - - // output word graph - if (m_wordGraphCollector) { - ostringstream out; - fix(out,PRECISION); - manager.GetWordGraph(m_lineNumber, out); - m_wordGraphCollector->Write(m_lineNumber, out.str()); - } - - // output search graph - if (m_searchGraphCollector) { - ostringstream out; - fix(out,PRECISION); - manager.OutputSearchGraph(m_lineNumber, out); - m_searchGraphCollector->Write(m_lineNumber, out.str()); - -#ifdef HAVE_PROTOBUF - if (staticData.GetOutputSearchGraphPB()) { - ostringstream sfn; - sfn << staticData.GetParam("output-search-graph-pb")[0] << '/' << m_lineNumber << ".pb" << ends; - string fn = sfn.str(); - VERBOSE(2, "Writing search graph to " << fn << endl); - fstream output(fn.c_str(), ios::trunc | ios::binary | ios::out); - manager.SerializeSearchGraphPB(m_lineNumber, output); - } -#endif - } - - // Output search graph in HTK standard lattice format (SLF) - if (m_outputSearchGraphSLF) { - stringstream fileName; - fileName << staticData.GetParam("output-search-graph-slf")[0] << "/" << m_lineNumber << ".slf"; - std::ofstream *file = new std::ofstream; - file->open(fileName.str().c_str()); - if (file->is_open() && file->good()) { - ostringstream out; - fix(out,PRECISION); - manager.OutputSearchGraphAsSLF(m_lineNumber, out); - *file << out.str(); - file -> flush(); - } else { - TRACE_ERR("Cannot output HTK standard lattice for line " << m_lineNumber << " because the output file is not open or not ready for writing" << std::endl); - } - delete file; - } - - // Output search graph in hypergraph format for Kenneth Heafield's lazy hypergraph decoder - if (m_outputSearchGraphHypergraph) { - - vector<string> hypergraphParameters = staticData.GetParam("output-search-graph-hypergraph"); - - bool appendSuffix; - if (hypergraphParameters.size() > 0 && hypergraphParameters[0] == "true") { - appendSuffix = true; - } else { - appendSuffix = false; - } - - string compression; - if (hypergraphParameters.size() > 1) { - compression = hypergraphParameters[1]; - } else { - compression = "txt"; - } - - string hypergraphDir; - if ( hypergraphParameters.size() > 2 ) { - hypergraphDir = hypergraphParameters[2]; - } else { - string nbestFile = staticData.GetNBestFilePath(); - if ( ! nbestFile.empty() && nbestFile!="-" && !boost::starts_with(nbestFile,"/dev/stdout") ) { - boost::filesystem::path nbestPath(nbestFile); - - // In the Boost filesystem API version 2, - // which was the default prior to Boost 1.46, - // the filename() method returned a string. - // - // In the Boost filesystem API version 3, - // which is the default starting with Boost 1.46, - // the filename() method returns a path object. - // - // To get a string from the path object, - // the native() method must be called. - // hypergraphDir = nbestPath.parent_path().filename() - //#if BOOST_VERSION >= 104600 - // .native() - //#endif - //; - - // Hopefully the following compiles under all versions of Boost. - // - // If this line gives you compile errors, - // contact Lane Schwartz on the Moses mailing list - hypergraphDir = nbestPath.parent_path().string(); - - } else { - stringstream hypergraphDirName; - hypergraphDirName << boost::filesystem::current_path().string() << "/hypergraph"; - hypergraphDir = hypergraphDirName.str(); - } - } - - if ( ! boost::filesystem::exists(hypergraphDir) ) { - boost::filesystem::create_directory(hypergraphDir); - } - - if ( ! boost::filesystem::exists(hypergraphDir) ) { - TRACE_ERR("Cannot output hypergraphs to " << hypergraphDir << " because the directory does not exist" << std::endl); - } else if ( ! boost::filesystem::is_directory(hypergraphDir) ) { - TRACE_ERR("Cannot output hypergraphs to " << hypergraphDir << " because that path exists, but is not a directory" << std::endl); - } else { - stringstream fileName; - fileName << hypergraphDir << "/" << m_lineNumber; - if ( appendSuffix ) { - fileName << "." << compression; - } - boost::iostreams::filtering_ostream *file - = new boost::iostreams::filtering_ostream; - - if ( compression == "gz" ) { - file->push( boost::iostreams::gzip_compressor() ); - } else if ( compression == "bz2" ) { - file->push( boost::iostreams::bzip2_compressor() ); - } else if ( compression != "txt" ) { - TRACE_ERR("Unrecognized hypergraph compression format (" - << compression - << ") - using uncompressed plain txt" << std::endl); - compression = "txt"; - } - - file->push( boost::iostreams::file_sink(fileName.str(), ios_base::out) ); - - if (file->is_complete() && file->good()) { - fix(*file,PRECISION); - manager.OutputSearchGraphAsHypergraph(*file); - file -> flush(); - } else { - TRACE_ERR("Cannot output hypergraph for line " << m_lineNumber - << " because the output file " << fileName.str() - << " is not open or not ready for writing" - << std::endl); - } - file -> pop(); - delete file; - } - } - additionalReportingTime.stop(); - - // apply decision rule and output best translation(s) - if (m_outputCollector) { - ostringstream out; - ostringstream debug; - fix(debug,PRECISION); - - // all derivations - send them to debug stream - if (staticData.PrintAllDerivations()) { - additionalReportingTime.start(); - manager.PrintAllDerivations(m_lineNumber, debug); - additionalReportingTime.stop(); - } - - Timer decisionRuleTime; - decisionRuleTime.start(); - - // MAP decoding: best hypothesis - const Hypothesis* bestHypo = NULL; - if (!staticData.UseMBR()) { - bestHypo = manager.GetBestHypothesis(); - if (bestHypo) { - if (StaticData::Instance().GetOutputHypoScore()) { - out << bestHypo->GetTotalScore() << ' '; - } - if (staticData.IsPathRecoveryEnabled()) { - OutputInput(out, bestHypo); - out << "||| "; - } - if (staticData.GetParam("print-id").size() && Scan<bool>(staticData.GetParam("print-id")[0]) ) { - out << m_source->GetTranslationId() << " "; - } - - if (staticData.GetReportSegmentation() == 2) { - manager.GetOutputLanguageModelOrder(out, bestHypo); - } - OutputBestSurface( - out, - bestHypo, - staticData.GetOutputFactorOrder(), - staticData.GetReportSegmentation(), - staticData.GetReportAllFactors()); - if (staticData.PrintAlignmentInfo()) { - out << "||| "; - OutputAlignment(out, bestHypo); - } - - OutputAlignment(m_alignmentInfoCollector, m_lineNumber, bestHypo); - IFVERBOSE(1) { - debug << "BEST TRANSLATION: " << *bestHypo << endl; - } - } else { - VERBOSE(1, "NO BEST TRANSLATION" << endl); - } - - out << endl; - } - - // MBR decoding (n-best MBR, lattice MBR, consensus) - else { - // we first need the n-best translations - size_t nBestSize = staticData.GetMBRSize(); - if (nBestSize <= 0) { - cerr << "ERROR: negative size for number of MBR candidate translations not allowed (option mbr-size)" << endl; - exit(1); - } - TrellisPathList nBestList; - manager.CalcNBest(nBestSize, nBestList,true); - VERBOSE(2,"size of n-best: " << nBestList.GetSize() << " (" << nBestSize << ")" << endl); - IFVERBOSE(2) { - PrintUserTime("calculated n-best list for (L)MBR decoding"); - } - - // lattice MBR - if (staticData.UseLatticeMBR()) { - if (m_nbestCollector) { - //lattice mbr nbest - vector<LatticeMBRSolution> solutions; - size_t n = min(nBestSize, staticData.GetNBestSize()); - getLatticeMBRNBest(manager,nBestList,solutions,n); - ostringstream out; - OutputLatticeMBRNBest(out, solutions,m_lineNumber); - m_nbestCollector->Write(m_lineNumber, out.str()); - } else { - //Lattice MBR decoding - vector<Word> mbrBestHypo = doLatticeMBR(manager,nBestList); - OutputBestHypo(mbrBestHypo, m_lineNumber, staticData.GetReportSegmentation(), - staticData.GetReportAllFactors(),out); - IFVERBOSE(2) { - PrintUserTime("finished Lattice MBR decoding"); - } - } - } - - // consensus decoding - else if (staticData.UseConsensusDecoding()) { - const TrellisPath &conBestHypo = doConsensusDecoding(manager,nBestList); - OutputBestHypo(conBestHypo, m_lineNumber, - staticData.GetReportSegmentation(), - staticData.GetReportAllFactors(),out); - OutputAlignment(m_alignmentInfoCollector, m_lineNumber, conBestHypo); - IFVERBOSE(2) { - PrintUserTime("finished Consensus decoding"); - } - } - - // n-best MBR decoding - else { - const Moses::TrellisPath &mbrBestHypo = doMBR(nBestList); - OutputBestHypo(mbrBestHypo, m_lineNumber, - staticData.GetReportSegmentation(), - staticData.GetReportAllFactors(),out); - OutputAlignment(m_alignmentInfoCollector, m_lineNumber, mbrBestHypo); - IFVERBOSE(2) { - PrintUserTime("finished MBR decoding"); - } - } - } - - // report best translation to output collector - m_outputCollector->Write(m_lineNumber,out.str(),debug.str()); - - decisionRuleTime.stop(); - VERBOSE(1, "Line " << m_lineNumber << ": Decision rule took " << decisionRuleTime << " seconds total" << endl); - } - - additionalReportingTime.start(); - - // output n-best list - if (m_nbestCollector && !staticData.UseLatticeMBR()) { - TrellisPathList nBestList; - ostringstream out; - manager.CalcNBest(staticData.GetNBestSize(), nBestList,staticData.GetDistinctNBest()); - OutputNBest(out, nBestList, staticData.GetOutputFactorOrder(), m_lineNumber, - staticData.GetReportSegmentation()); - m_nbestCollector->Write(m_lineNumber, out.str()); - } - - //lattice samples - if (m_latticeSamplesCollector) { - TrellisPathList latticeSamples; - ostringstream out; - manager.CalcLatticeSamples(staticData.GetLatticeSamplesSize(), latticeSamples); - OutputNBest(out,latticeSamples, staticData.GetOutputFactorOrder(), m_lineNumber, - staticData.GetReportSegmentation()); - m_latticeSamplesCollector->Write(m_lineNumber, out.str()); - } - - // detailed translation reporting - if (m_detailedTranslationCollector) { - ostringstream out; - fix(out,PRECISION); - TranslationAnalysis::PrintTranslationAnalysis(out, manager.GetBestHypothesis()); - m_detailedTranslationCollector->Write(m_lineNumber,out.str()); - } - - //list of unknown words - if (m_unknownsCollector) { - const vector<const Phrase*>& unknowns = manager.getSntTranslationOptions()->GetUnknownSources(); - ostringstream out; - for (size_t i = 0; i < unknowns.size(); ++i) { - out << *(unknowns[i]); - } - out << endl; - m_unknownsCollector->Write(m_lineNumber, out.str()); - } - - // report additional statistics - manager.CalcDecoderStatistics(); - VERBOSE(1, "Line " << m_lineNumber << ": Additional reporting took " << additionalReportingTime << " seconds total" << endl); - VERBOSE(1, "Line " << m_lineNumber << ": Translation took " << translationTime << " seconds total" << endl); - IFVERBOSE(2) { - PrintUserTime("Sentence Decoding Time:"); - } - } - - ~TranslationTask() { - delete m_source; - } - -private: - InputType* m_source; - size_t m_lineNumber; - OutputCollector* m_outputCollector; - OutputCollector* m_nbestCollector; - OutputCollector* m_latticeSamplesCollector; - OutputCollector* m_wordGraphCollector; - OutputCollector* m_searchGraphCollector; - OutputCollector* m_detailedTranslationCollector; - OutputCollector* m_alignmentInfoCollector; - OutputCollector* m_unknownsCollector; - bool m_outputSearchGraphSLF; - bool m_outputSearchGraphHypergraph; - std::ofstream *m_alignmentStream; - - -}; - -static void PrintFeatureWeight(const FeatureFunction* ff) -{ - cout << ff->GetScoreProducerDescription() << "="; - size_t numScoreComps = ff->GetNumScoreComponents(); - vector<float> values = StaticData::Instance().GetAllWeights().GetScoresForProducer(ff); - for (size_t i = 0; i < numScoreComps; ++i) { - cout << " " << values[i]; - } - cout << endl; -} - -static void ShowWeights() -{ - //TODO: Find a way of ensuring this order is synced with the nbest - fix(cout,6); - const vector<const StatelessFeatureFunction*>& slf = StatelessFeatureFunction::GetStatelessFeatureFunctions(); - const vector<const StatefulFeatureFunction*>& sff = StatefulFeatureFunction::GetStatefulFeatureFunctions(); - - for (size_t i = 0; i < sff.size(); ++i) { - const StatefulFeatureFunction *ff = sff[i]; - if (ff->IsTuneable()) { - PrintFeatureWeight(ff); - } - else { - cout << ff->GetScoreProducerDescription() << " UNTUNEABLE" << endl; - } - } - for (size_t i = 0; i < slf.size(); ++i) { - const StatelessFeatureFunction *ff = slf[i]; - if (ff->IsTuneable()) { - PrintFeatureWeight(ff); - } - else { - cout << ff->GetScoreProducerDescription() << " UNTUNEABLE" << endl; - } - } -} - -size_t OutputFeatureWeightsForHypergraph(size_t index, const FeatureFunction* ff, std::ostream &outputSearchGraphStream) -{ - size_t numScoreComps = ff->GetNumScoreComponents(); - if (numScoreComps != 0) { - vector<float> values = StaticData::Instance().GetAllWeights().GetScoresForProducer(ff); - if (numScoreComps > 1) { - for (size_t i = 0; i < numScoreComps; ++i) { - outputSearchGraphStream << ff->GetScoreProducerDescription() - << i - << "=" << values[i] << endl; - } - } else { - outputSearchGraphStream << ff->GetScoreProducerDescription() - << "=" << values[0] << endl; - } - return index+numScoreComps; - } else { - UTIL_THROW2("Sparse features are not yet supported when outputting hypergraph format"); - } -} - void OutputFeatureWeightsForHypergraph(std::ostream &outputSearchGraphStream) { outputSearchGraphStream.setf(std::ios::fixed); outputSearchGraphStream.precision(6); - - const vector<const StatelessFeatureFunction*>& slf =StatelessFeatureFunction::GetStatelessFeatureFunctions(); - const vector<const StatefulFeatureFunction*>& sff = StatefulFeatureFunction::GetStatefulFeatureFunctions(); - size_t featureIndex = 1; - for (size_t i = 0; i < sff.size(); ++i) { - featureIndex = OutputFeatureWeightsForHypergraph(featureIndex, sff[i], outputSearchGraphStream); - } - for (size_t i = 0; i < slf.size(); ++i) { - /* - if (slf[i]->GetScoreProducerWeightShortName() != "u" && - slf[i]->GetScoreProducerWeightShortName() != "tm" && - slf[i]->GetScoreProducerWeightShortName() != "I" && - slf[i]->GetScoreProducerWeightShortName() != "g") - */ - { - featureIndex = OutputFeatureWeightsForHypergraph(featureIndex, slf[i], outputSearchGraphStream); - } - } - const vector<PhraseDictionary*>& pds = PhraseDictionary::GetColl(); - for( size_t i=0; i<pds.size(); i++ ) { - featureIndex = OutputFeatureWeightsForHypergraph(featureIndex, pds[i], outputSearchGraphStream); - } - const vector<GenerationDictionary*>& gds = GenerationDictionary::GetColl(); - for( size_t i=0; i<gds.size(); i++ ) { - featureIndex = OutputFeatureWeightsForHypergraph(featureIndex, gds[i], outputSearchGraphStream); - } - + StaticData::Instance().GetAllWeights().Save(outputSearchGraphStream); } @@ -603,9 +92,36 @@ int main(int argc, char** argv) } // set number of significant decimals in output - fix(cout,PRECISION); - fix(cerr,PRECISION); - + // By the way, having a static member function in an Moses-specific class + // just to do + // cout.setf(ios::fixed);cout.precision(3); + // doesn't make sense. + // 1. It makes the program harder to understand. If I see + // cout.setf(ios::fixed);cout.precision(3); + // I know immediately what's going on. If I see, + // IOWrapper::FixPrecision(cout); + // I assume that something much more elaborate is going on (side effects?) + // and need to look it up. + // After all, one reads programs in order to understand them. + // 2. If you want to enforce a certain default behaviour on stdout, + // that should happen only once. Other components should not do that. + // 3. If you want to format locally, use + // string IOWrapper::score_default_formatting = "%.3f"; + // cout << boost::format(IOWrapper::score_default_formatting) % my_variable; + // This even allows you to do pretty printing by default if you wish. + // (Actually, the format variable should be constructed + // from an int IOWrapper::score_output_default_precision itself.) + // 4. If output speed is an issue and you are afraid of the overhead that + // boost::format may add, don't use streams to begin with. + // IOWrapper::FixPrecision(cout); + // IOWrapper::FixPrecision(cerr); + cout.setf(ios::fixed); cout.precision(3); + cout.setf(ios::fixed); cout.precision(3); + +#ifdef PT_UG + // see my notes on the moses support mailing list, this is a hack around + // the misguided implementation of LoadParameter(), which should leave + // arguments it doesn't know alone. vector<pair<string,int> > argfilter(4); argfilter[0] = std::make_pair(string("--spe-src"),1); argfilter[1] = std::make_pair(string("--spe-trg"),1); @@ -616,16 +132,16 @@ int main(int argc, char** argv) filter_arguments(argc, argv, mo_acnt, &mo_args, my_acnt, &my_args, argfilter); ifstream spe_src,spe_trg,spe_aln; - // instead of translating show coverage by phrase tables for (int i = 0; i < my_acnt; i += 2) { - if (!strcmp(my_args[i],"--spe-src")) - spe_src.open(my_args[i+1]); - else if (!strcmp(my_args[i],"--spe-trg")) - spe_trg.open(my_args[i+1]); - else if (!strcmp(my_args[i],"--spe-aln")) - spe_aln.open(my_args[i+1]); + if (!strcmp(my_args[i],"--spe-src")) + spe_src.open(my_args[i+1]); + else if (!strcmp(my_args[i],"--spe-trg")) + spe_trg.open(my_args[i+1]); + else if (!strcmp(my_args[i],"--spe-aln")) + spe_aln.open(my_args[i+1]); } +#endif // load all the settings into the Parameter class // (stores them as strings, or array of strings) @@ -634,7 +150,6 @@ int main(int argc, char** argv) exit(1); } - // initialize all "global" variables, which are stored in StaticData // note: this also loads models such as the language model, etc. if (!StaticData::LoadDataStatic(¶ms, argv[0])) { @@ -650,12 +165,11 @@ int main(int argc, char** argv) // shorthand for accessing information in StaticData const StaticData& staticData = StaticData::Instance(); - //initialise random numbers srand(time(NULL)); // set up read/writing class - IOWrapper* ioWrapper = GetIOWrapper(staticData); + IOWrapper* ioWrapper = IOWrapper::GetIOWrapper(staticData); if (!ioWrapper) { cerr << "Error; Failed to create IO object" << endl; exit(1); @@ -668,113 +182,17 @@ int main(int argc, char** argv) TRACE_ERR(weights); TRACE_ERR("\n"); } - if (staticData.GetOutputSearchGraphHypergraph()) { - ofstream* weightsOut = new std::ofstream; - stringstream weightsFilename; - if (staticData.GetParam("output-search-graph-hypergraph").size() > 3) { - weightsFilename << staticData.GetParam("output-search-graph-hypergraph")[3]; - } else { - string nbestFile = staticData.GetNBestFilePath(); - if ( ! nbestFile.empty() && nbestFile!="-" && !boost::starts_with(nbestFile,"/dev/stdout") ) { - boost::filesystem::path nbestPath(nbestFile); - weightsFilename << nbestPath.parent_path().filename() << "/weights"; - } else { - weightsFilename << boost::filesystem::current_path().string() << "/hypergraph/weights"; - } - } - boost::filesystem::path weightsFilePath(weightsFilename.str()); - if ( ! boost::filesystem::exists(weightsFilePath.parent_path()) ) { - boost::filesystem::create_directory(weightsFilePath.parent_path()); - } - TRACE_ERR("The weights file is " << weightsFilename.str() << "\n"); - weightsOut->open(weightsFilename.str().c_str()); - OutputFeatureWeightsForHypergraph(*weightsOut); - weightsOut->flush(); - weightsOut->close(); - delete weightsOut; - } - - - // initialize output streams - // note: we can't just write to STDOUT or files - // because multithreading may return sentences in shuffled order - auto_ptr<OutputCollector> outputCollector; // for translations - auto_ptr<OutputCollector> nbestCollector; // for n-best lists - auto_ptr<OutputCollector> latticeSamplesCollector; //for lattice samples - auto_ptr<ofstream> nbestOut; - auto_ptr<ofstream> latticeSamplesOut; - size_t nbestSize = staticData.GetNBestSize(); - string nbestFile = staticData.GetNBestFilePath(); - bool output1best = true; - if (nbestSize) { - if (nbestFile == "-" || nbestFile == "/dev/stdout") { - // nbest to stdout, no 1-best - nbestCollector.reset(new OutputCollector()); - output1best = false; - } else { - // nbest to file, 1-best to stdout - nbestOut.reset(new ofstream(nbestFile.c_str())); - if (!nbestOut->good()) { - TRACE_ERR("ERROR: Failed to open " << nbestFile << " for nbest lists" << endl); - exit(1); - } - nbestCollector.reset(new OutputCollector(nbestOut.get())); - } - } - size_t latticeSamplesSize = staticData.GetLatticeSamplesSize(); - string latticeSamplesFile = staticData.GetLatticeSamplesFilePath(); - if (latticeSamplesSize) { - if (latticeSamplesFile == "-" || latticeSamplesFile == "/dev/stdout") { - latticeSamplesCollector.reset(new OutputCollector()); - output1best = false; - } else { - latticeSamplesOut.reset(new ofstream(latticeSamplesFile.c_str())); - if (!latticeSamplesOut->good()) { - TRACE_ERR("ERROR: Failed to open " << latticeSamplesFile << " for lattice samples" << endl); - exit(1); - } - latticeSamplesCollector.reset(new OutputCollector(latticeSamplesOut.get())); - } - } - if (output1best) { - outputCollector.reset(new OutputCollector()); - } - - // initialize stream for word graph (aka: output lattice) - auto_ptr<OutputCollector> wordGraphCollector; - if (staticData.GetOutputWordGraph()) { - wordGraphCollector.reset(new OutputCollector(&(ioWrapper->GetOutputWordGraphStream()))); - } - - // initialize stream for search graph - // note: this is essentially the same as above, but in a different format - auto_ptr<OutputCollector> searchGraphCollector; - if (staticData.GetOutputSearchGraph()) { - searchGraphCollector.reset(new OutputCollector(&(ioWrapper->GetOutputSearchGraphStream()))); - } - - // initialize stram for details about the decoder run - auto_ptr<OutputCollector> detailedTranslationCollector; - if (staticData.IsDetailedTranslationReportingEnabled()) { - detailedTranslationCollector.reset(new OutputCollector(&(ioWrapper->GetDetailedTranslationReportingStream()))); - } - // initialize stram for word alignment between input and output - auto_ptr<OutputCollector> alignmentInfoCollector; - if (!staticData.GetAlignmentOutputFile().empty()) { - alignmentInfoCollector.reset(new OutputCollector(ioWrapper->GetAlignmentOutputStream())); - } + boost::shared_ptr<HypergraphOutput<Manager> > hypergraphOutput; + boost::shared_ptr<HypergraphOutput<ChartManager> > hypergraphOutputChart; - //initialise stream for unknown (oov) words - auto_ptr<OutputCollector> unknownsCollector; - auto_ptr<ofstream> unknownsStream; - if (!staticData.GetOutputUnknownsFile().empty()) { - unknownsStream.reset(new ofstream(staticData.GetOutputUnknownsFile().c_str())); - if (!unknownsStream->good()) { - TRACE_ERR("Unable to open " << staticData.GetOutputUnknownsFile() << " for unknowns"); - exit(1); - } - unknownsCollector.reset(new OutputCollector(unknownsStream.get())); + if (staticData.GetOutputSearchGraphHypergraph()) { + if (staticData.IsChart()) { + hypergraphOutputChart.reset(new HypergraphOutput<ChartManager>(PRECISION)); + } + else { + hypergraphOutput.reset(new HypergraphOutput<Manager>(PRECISION)); + } } #ifdef WITH_THREADS @@ -784,45 +202,54 @@ int main(int argc, char** argv) // main loop over set of input sentences InputType* source = NULL; size_t lineCount = staticData.GetStartTranslationId(); - while(ReadInput(*ioWrapper,staticData.GetInputType(),source)) { + while(ioWrapper->ReadInput(staticData.GetInputType(),source)) { + source->SetTranslationId(lineCount); IFVERBOSE(1) { ResetUserTime(); } + + FeatureFunction::CallChangeSource(source); + // set up task of translating one sentence - TranslationTask* task = - new TranslationTask(lineCount,source, outputCollector.get(), - nbestCollector.get(), - latticeSamplesCollector.get(), - wordGraphCollector.get(), - searchGraphCollector.get(), - detailedTranslationCollector.get(), - alignmentInfoCollector.get(), - unknownsCollector.get(), - staticData.GetOutputSearchGraphSLF(), - staticData.GetOutputSearchGraphHypergraph()); + TranslationTask* task; + if (staticData.IsChart()) { + // scfg + task = new TranslationTask(source, *ioWrapper, hypergraphOutputChart); + } + else { + // pb + task = new TranslationTask(source, *ioWrapper, + staticData.GetOutputSearchGraphSLF(), + hypergraphOutput); + } + // execute task #ifdef WITH_THREADS +#ifdef PT_UG if (my_acnt) - { - task->Run(); - delete task; - string src,trg,aln; - UTIL_THROW_IF2(!getline(spe_src,src), "[" << HERE << "] " - << "missing update data for simulated post-editing."); - UTIL_THROW_IF2(!getline(spe_trg,trg), "[" << HERE << "] " - << "missing update data for simulated post-editing."); - UTIL_THROW_IF2(!getline(spe_aln,aln), "[" << HERE << "] " - << "missing update data for simulated post-editing."); - BOOST_FOREACH (PhraseDictionary* pd, PhraseDictionary::GetColl()) - { - Mmsapt* sapt = dynamic_cast<Mmsapt*>(pd); - if (sapt) sapt->add(src,trg,aln); - VERBOSE(1,"[" << HERE << " added src] " << src << endl); - VERBOSE(1,"[" << HERE << " added trg] " << trg << endl); - VERBOSE(1,"[" << HERE << " added aln] " << aln << endl); - } - } - else pool.Submit(task); + { + // simulated post-editing: always run single-threaded! + task->Run(); + delete task; + string src,trg,aln; + UTIL_THROW_IF2(!getline(spe_src,src), "[" << HERE << "] " + << "missing update data for simulated post-editing."); + UTIL_THROW_IF2(!getline(spe_trg,trg), "[" << HERE << "] " + << "missing update data for simulated post-editing."); + UTIL_THROW_IF2(!getline(spe_aln,aln), "[" << HERE << "] " + << "missing update data for simulated post-editing."); + BOOST_FOREACH (PhraseDictionary* pd, PhraseDictionary::GetColl()) + { + Mmsapt* sapt = dynamic_cast<Mmsapt*>(pd); + if (sapt) sapt->add(src,trg,aln); + VERBOSE(1,"[" << HERE << " added src] " << src << endl); + VERBOSE(1,"[" << HERE << " added trg] " << trg << endl); + VERBOSE(1,"[" << HERE << " added aln] " << aln << endl); + } + } + else +#endif + pool.Submit(task); #else task->Run(); delete task; |