mosesdecoder/mert/HopeFearDecoder.cpp
2015-05-15 18:09:38 +01:00

453 lines
14 KiB
C++

/***********************************************************************
Moses - factored phrase-based language decoder
Copyright (C) 2014- 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
***********************************************************************/
#include <algorithm>
#include <cmath>
#include <iterator>
#define BOOST_FILESYSTEM_VERSION 3
#include <boost/filesystem.hpp>
#include <boost/lexical_cast.hpp>
#include "util/exception.hh"
#include "util/file_piece.hh"
#include "Scorer.h"
#include "HopeFearDecoder.h"
using namespace std;
namespace fs = boost::filesystem;
namespace MosesTuning
{
static const ValType BLEU_RATIO = 5;
std::pair<MiraWeightVector*,size_t>
InitialiseWeights(const string& denseInitFile, const string& sparseInitFile,
const string& type, bool verbose)
{
// Dense
vector<parameter_t> initParams;
if(!denseInitFile.empty()) {
ifstream opt(denseInitFile.c_str());
string buffer;
if (opt.fail()) {
cerr << "could not open dense initfile: " << denseInitFile << endl;
exit(3);
}
if (verbose) cerr << "Reading dense features:" << endl;
parameter_t val;
getline(opt,buffer);
if (buffer.find_first_of("=") == buffer.npos) {
UTIL_THROW_IF(type == "hypergraph", util::Exception, "For hypergraph version, require dense features in 'name= value' format");
cerr << "WARN: dense features in deprecated Moses mert format. Prefer 'name= value' format." << endl;
istringstream strstrm(buffer);
while(strstrm >> val) {
initParams.push_back(val);
if(verbose) cerr << val << endl;
}
} else {
vector<string> names;
string last_name = "";
size_t feature_ctr = 1;
do {
size_t equals = buffer.find_last_of("=");
UTIL_THROW_IF(equals == buffer.npos, util::Exception, "Incorrect format in dense feature file: '"
<< buffer << "'");
string name = buffer.substr(0,equals);
names.push_back(name);
initParams.push_back(boost::lexical_cast<ValType>(buffer.substr(equals+2)));
//Names for features with several values need to have their id added
if (name != last_name) feature_ctr = 1;
last_name = name;
if (feature_ctr>1) {
stringstream namestr;
namestr << names.back() << "_" << feature_ctr;
names[names.size()-1] = namestr.str();
if (feature_ctr == 2) {
stringstream namestr;
namestr << names[names.size()-2] << "_" << (feature_ctr-1);
names[names.size()-2] = namestr.str();
}
}
++feature_ctr;
} while(getline(opt,buffer));
//Make sure that SparseVector encodes dense feature names as 0..n-1.
for (size_t i = 0; i < names.size(); ++i) {
size_t id = SparseVector::encode(names[i]);
assert(id == i);
if (verbose) cerr << names[i] << " " << initParams[i] << endl;
}
}
opt.close();
}
size_t initDenseSize = initParams.size();
// Sparse
if(!sparseInitFile.empty()) {
if(initDenseSize==0) {
cerr << "sparse initialization requires dense initialization" << endl;
exit(3);
}
ifstream opt(sparseInitFile.c_str());
if(opt.fail()) {
cerr << "could not open sparse initfile: " << sparseInitFile << endl;
exit(3);
}
int sparseCount=0;
parameter_t val;
std::string name;
while(opt >> name >> val) {
size_t id = SparseVector::encode(name) + initDenseSize;
while(initParams.size()<=id) initParams.push_back(0.0);
initParams[id] = val;
sparseCount++;
}
cerr << "Found " << sparseCount << " initial sparse features" << endl;
opt.close();
}
return pair<MiraWeightVector*,size_t>(new MiraWeightVector(initParams), initDenseSize);
}
ValType HopeFearDecoder::Evaluate(const AvgWeightVector& wv)
{
vector<ValType> stats(scorer_->NumberOfScores(),0);
for(reset(); !finished(); next()) {
vector<ValType> sent;
MaxModel(wv,&sent);
for(size_t i=0; i<sent.size(); i++) {
stats[i]+=sent[i];
}
}
return scorer_->calculateScore(stats);
}
NbestHopeFearDecoder::NbestHopeFearDecoder(
const vector<string>& featureFiles,
const vector<string>& scoreFiles,
bool streaming,
bool no_shuffle,
bool safe_hope,
Scorer* scorer
) : safe_hope_(safe_hope)
{
scorer_ = scorer;
if (streaming) {
train_.reset(new StreamingHypPackEnumerator(featureFiles, scoreFiles));
} else {
train_.reset(new RandomAccessHypPackEnumerator(featureFiles, scoreFiles, no_shuffle));
}
}
void NbestHopeFearDecoder::next()
{
train_->next();
}
bool NbestHopeFearDecoder::finished()
{
return train_->finished();
}
void NbestHopeFearDecoder::reset()
{
train_->reset();
}
void NbestHopeFearDecoder::HopeFear(
const std::vector<ValType>& backgroundBleu,
const MiraWeightVector& wv,
HopeFearData* hopeFear
)
{
// Hope / fear decode
ValType hope_scale = 1.0;
size_t hope_index=0, fear_index=0, model_index=0;
ValType hope_score=0, fear_score=0, model_score=0;
for(size_t safe_loop=0; safe_loop<2; safe_loop++) {
ValType hope_bleu=0, hope_model=0;
for(size_t i=0; i< train_->cur_size(); i++) {
const MiraFeatureVector& vec=train_->featuresAt(i);
ValType score = wv.score(vec);
ValType bleu = scorer_->calculateSentenceLevelBackgroundScore(train_->scoresAt(i),backgroundBleu);
// Hope
if(i==0 || (hope_scale*score + bleu) > hope_score) {
hope_score = hope_scale*score + bleu;
hope_index = i;
hope_bleu = bleu;
hope_model = score;
}
// Fear
if(i==0 || (score - bleu) > fear_score) {
fear_score = score - bleu;
fear_index = i;
}
// Model
if(i==0 || score > model_score) {
model_score = score;
model_index = i;
}
}
// Outer loop rescales the contribution of model score to 'hope' in antagonistic cases
// where model score is having far more influence than BLEU
hope_bleu *= BLEU_RATIO; // We only care about cases where model has MUCH more influence than BLEU
if(safe_hope_ && safe_loop==0 && abs(hope_model)>1e-8 && abs(hope_bleu)/abs(hope_model)<hope_scale)
hope_scale = abs(hope_bleu) / abs(hope_model);
else break;
}
hopeFear->modelFeatures = train_->featuresAt(model_index);
hopeFear->hopeFeatures = train_->featuresAt(hope_index);
hopeFear->fearFeatures = train_->featuresAt(fear_index);
hopeFear->hopeStats = train_->scoresAt(hope_index);
hopeFear->hopeBleu = scorer_->calculateSentenceLevelBackgroundScore(hopeFear->hopeStats, backgroundBleu);
const vector<float>& fear_stats = train_->scoresAt(fear_index);
hopeFear->fearBleu = scorer_->calculateSentenceLevelBackgroundScore(fear_stats, backgroundBleu);
hopeFear->modelStats = train_->scoresAt(model_index);
hopeFear->hopeFearEqual = (hope_index == fear_index);
}
void NbestHopeFearDecoder::MaxModel(const AvgWeightVector& wv, std::vector<ValType>* stats)
{
// Find max model
size_t max_index=0;
ValType max_score=0;
for(size_t i=0; i<train_->cur_size(); i++) {
MiraFeatureVector vec(train_->featuresAt(i));
ValType score = wv.score(vec);
if(i==0 || score > max_score) {
max_index = i;
max_score = score;
}
}
*stats = train_->scoresAt(max_index);
}
HypergraphHopeFearDecoder::HypergraphHopeFearDecoder
(
const string& hypergraphDir,
const vector<string>& referenceFiles,
size_t num_dense,
bool streaming,
bool no_shuffle,
bool safe_hope,
size_t hg_pruning,
const MiraWeightVector& wv,
Scorer* scorer
) :
num_dense_(num_dense)
{
UTIL_THROW_IF(streaming, util::Exception, "Streaming not currently supported for hypergraphs");
UTIL_THROW_IF(!fs::exists(hypergraphDir), HypergraphException, "Directory '" << hypergraphDir << "' does not exist");
UTIL_THROW_IF(!referenceFiles.size(), util::Exception, "No reference files supplied");
references_.Load(referenceFiles, vocab_);
SparseVector weights;
wv.ToSparse(&weights,num_dense_);
scorer_ = scorer;
static const string kWeights = "weights";
fs::directory_iterator dend;
size_t fileCount = 0;
cerr << "Reading hypergraphs" << endl;
for (fs::directory_iterator di(hypergraphDir); di != dend; ++di) {
const fs::path& hgpath = di->path();
if (hgpath.filename() == kWeights) continue;
// cerr << "Reading " << hgpath.filename() << endl;
Graph graph(vocab_);
size_t id = boost::lexical_cast<size_t>(hgpath.stem().string());
util::scoped_fd fd(util::OpenReadOrThrow(hgpath.string().c_str()));
//util::FilePiece file(di->path().string().c_str());
util::FilePiece file(fd.release());
ReadGraph(file,graph);
//cerr << "ref length " << references_.Length(id) << endl;
size_t edgeCount = hg_pruning * references_.Length(id);
boost::shared_ptr<Graph> prunedGraph;
prunedGraph.reset(new Graph(vocab_));
graph.Prune(prunedGraph.get(), weights, edgeCount);
graphs_[id] = prunedGraph;
// cerr << "Pruning to v=" << graphs_[id]->VertexSize() << " e=" << graphs_[id]->EdgeSize() << endl;
++fileCount;
if (fileCount % 10 == 0) cerr << ".";
if (fileCount % 400 == 0) cerr << " [count=" << fileCount << "]\n";
}
cerr << endl << "Done" << endl;
sentenceIds_.resize(graphs_.size());
for (size_t i = 0; i < graphs_.size(); ++i) sentenceIds_[i] = i;
if (!no_shuffle) {
random_shuffle(sentenceIds_.begin(), sentenceIds_.end());
}
}
void HypergraphHopeFearDecoder::reset()
{
sentenceIdIter_ = sentenceIds_.begin();
}
void HypergraphHopeFearDecoder::next()
{
sentenceIdIter_++;
}
bool HypergraphHopeFearDecoder::finished()
{
return sentenceIdIter_ == sentenceIds_.end();
}
void HypergraphHopeFearDecoder::HopeFear(
const vector<ValType>& backgroundBleu,
const MiraWeightVector& wv,
HopeFearData* hopeFear
)
{
size_t sentenceId = *sentenceIdIter_;
SparseVector weights;
wv.ToSparse(&weights, num_dense_);
const Graph& graph = *(graphs_[sentenceId]);
// ValType hope_scale = 1.0;
HgHypothesis hopeHypo, fearHypo, modelHypo;
for(size_t safe_loop=0; safe_loop<2; safe_loop++) {
//hope decode
Viterbi(graph, weights, 1, references_, sentenceId, backgroundBleu, &hopeHypo);
//fear decode
Viterbi(graph, weights, -1, references_, sentenceId, backgroundBleu, &fearHypo);
//Model decode
Viterbi(graph, weights, 0, references_, sentenceId, backgroundBleu, &modelHypo);
// Outer loop rescales the contribution of model score to 'hope' in antagonistic cases
// where model score is having far more influence than BLEU
// hope_bleu *= BLEU_RATIO; // We only care about cases where model has MUCH more influence than BLEU
// if(safe_hope_ && safe_loop==0 && abs(hope_model)>1e-8 && abs(hope_bleu)/abs(hope_model)<hope_scale)
// hope_scale = abs(hope_bleu) / abs(hope_model);
// else break;
//TODO: Don't currently get model and bleu so commented this out for now.
break;
}
//modelFeatures, hopeFeatures and fearFeatures
hopeFear->modelFeatures = MiraFeatureVector(modelHypo.featureVector, num_dense_);
hopeFear->hopeFeatures = MiraFeatureVector(hopeHypo.featureVector, num_dense_);
hopeFear->fearFeatures = MiraFeatureVector(fearHypo.featureVector, num_dense_);
//Need to know which are to be mapped to dense features!
//Only C++11
//hopeFear->modelStats.assign(std::begin(modelHypo.bleuStats), std::end(modelHypo.bleuStats));
vector<ValType> fearStats(scorer_->NumberOfScores());
hopeFear->hopeStats.reserve(scorer_->NumberOfScores());
hopeFear->modelStats.reserve(scorer_->NumberOfScores());
for (size_t i = 0; i < fearStats.size(); ++i) {
hopeFear->modelStats.push_back(modelHypo.bleuStats[i]);
hopeFear->hopeStats.push_back(hopeHypo.bleuStats[i]);
fearStats[i] = fearHypo.bleuStats[i];
}
/*
cerr << "hope" << endl;;
for (size_t i = 0; i < hopeHypo.text.size(); ++i) {
cerr << hopeHypo.text[i]->first << " ";
}
cerr << endl;
for (size_t i = 0; i < fearStats.size(); ++i) {
cerr << hopeHypo.bleuStats[i] << " ";
}
cerr << endl;
cerr << "fear";
for (size_t i = 0; i < fearHypo.text.size(); ++i) {
cerr << fearHypo.text[i]->first << " ";
}
cerr << endl;
for (size_t i = 0; i < fearStats.size(); ++i) {
cerr << fearHypo.bleuStats[i] << " ";
}
cerr << endl;
cerr << "model";
for (size_t i = 0; i < modelHypo.text.size(); ++i) {
cerr << modelHypo.text[i]->first << " ";
}
cerr << endl;
for (size_t i = 0; i < fearStats.size(); ++i) {
cerr << modelHypo.bleuStats[i] << " ";
}
cerr << endl;
*/
hopeFear->hopeBleu = sentenceLevelBackgroundBleu(hopeFear->hopeStats, backgroundBleu);
hopeFear->fearBleu = sentenceLevelBackgroundBleu(fearStats, backgroundBleu);
//If fv and bleu stats are equal, then assume equal
hopeFear->hopeFearEqual = true; //(hopeFear->hopeBleu - hopeFear->fearBleu) >= 1e-8;
if (hopeFear->hopeFearEqual) {
for (size_t i = 0; i < fearStats.size(); ++i) {
if (fearStats[i] != hopeFear->hopeStats[i]) {
hopeFear->hopeFearEqual = false;
break;
}
}
}
hopeFear->hopeFearEqual = hopeFear->hopeFearEqual && (hopeFear->fearFeatures == hopeFear->hopeFeatures);
}
void HypergraphHopeFearDecoder::MaxModel(const AvgWeightVector& wv, vector<ValType>* stats)
{
assert(!finished());
HgHypothesis bestHypo;
size_t sentenceId = *sentenceIdIter_;
SparseVector weights;
wv.ToSparse(&weights, num_dense_);
vector<ValType> bg(scorer_->NumberOfScores());
//cerr << "Calculating bleu on " << sentenceId << endl;
Viterbi(*(graphs_[sentenceId]), weights, 0, references_, sentenceId, bg, &bestHypo);
stats->resize(bestHypo.bleuStats.size());
/*
for (size_t i = 0; i < bestHypo.text.size(); ++i) {
cerr << bestHypo.text[i]->first << " ";
}
cerr << endl;
*/
for (size_t i = 0; i < bestHypo.bleuStats.size(); ++i) {
(*stats)[i] = bestHypo.bleuStats[i];
}
}
};