moses-smt/mosesdecoder
1
1
Fork 0
mirror of https://github.com/moses-smt/mosesdecoder.git synced 2024-11-10 00:47:31 +03:00
Code Issues Packages Projects Releases Wiki Activity

introduce parameter --update-after-epoch

Browse Source 
git-svn-id: http://svn.statmt.org/repository/mira@3845 cc96ff50-19ce-11e0-b349-13d7f0bd23df
This commit is contained in:
ehasler 2011-03-23 12:13:38 +00:00 committed by Ondrej Bojar
parent 41d2d28408
commit d04066ee0e
5 changed files with 995 additions and 893 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
mira/Decoder.cpp
View File

@ -101,7 +101,7 @@ namespace Mira {
vector< float>& bleuScores,
bool oracle,
bool distinct,
size_t rank)
size_t rank)
{
StaticData &staticData = StaticData::InstanceNonConst();
@ -138,7 +138,7 @@ namespace Mira {
//std::cout << "Score breakdown: " << path.GetScoreBreakdown() << endl;
float scoreWithoutBleu = path.GetTotalScore() - bleuObjectiveWeight * bleuScore;
cerr << "Total score: " << path.GetTotalScore() << ", Score w/o bleu: " << scoreWithoutBleu << ", Bleu: " << bleuScore << endl;
cerr << "Rank " << rank << ", total score: " << path.GetTotalScore() << ", Score w/o bleu: " << scoreWithoutBleu << ", Bleu: " << bleuScore << endl;
Phrase bestPhrase = path.GetTargetPhrase();

1682
mira/Main.cpp
View File

File diff suppressed because it is too large Load Diff

143
mira/MiraOptimiser.cpp
View File

@ -7,15 +7,13 @@ using namespace std;
namespace Mira {
int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
const vector< vector<ScoreComponentCollection> >& featureValues,
const vector< vector<float> >& losses,
const vector<std::vector<float> >& bleuScores,
const vector< ScoreComponentCollection>& oracleFeatureValues,
const vector< float> oracleBleuScores,
const vector< size_t> sentenceIds,
float learning_rate,
float max_sentence_update,
size_t rank) {
const vector<vector<ScoreComponentCollection> >& featureValues,
const vector<vector<float> >& losses,
const vector<std::vector<float> >& bleuScores, const vector<
ScoreComponentCollection>& oracleFeatureValues,
const vector<float> oracleBleuScores, const vector<size_t> sentenceIds,
float learning_rate, float max_sentence_update, size_t rank,
bool update_after_epoch) {
// add every oracle in batch to list of oracles (under certain conditions)
for (size_t i = 0; i < oracleFeatureValues.size(); ++i) {
@ -35,8 +33,8 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
m_bleu_of_oracles[sentenceId][j] = oracleBleuScores[j];
updated = true;
break;
}
else if (worstWeightedScore == 0 || currentWeightedScore > worstWeightedScore){
} else if (worstWeightedScore == 0 || currentWeightedScore
> worstWeightedScore) {
worstWeightedScore = currentWeightedScore;
indexOfWorst = j;
}
@ -47,8 +45,7 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
if (m_max_number_oracles > m_oracles[sentenceId].size()) {
m_oracles[sentenceId].push_back(oracleFeatureValues[i]);
m_bleu_of_oracles[sentenceId].push_back(oracleBleuScores[i]);
}
else {
} else {
m_oracles[sentenceId][indexOfWorst] = oracleFeatureValues[i];
m_bleu_of_oracles[sentenceId][indexOfWorst] = oracleBleuScores[i];
}
@ -56,8 +53,8 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
}
size_t violatedConstraintsBefore = 0;
vector< ScoreComponentCollection> featureValueDiffs;
vector< float> lossMarginDistances;
vector<ScoreComponentCollection> featureValueDiffs;
vector<float> lossMarginDistances;
// find most violated constraint
float maxViolationLossMarginDistance;
@ -67,8 +64,9 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
for (size_t i = 0; i < featureValues.size(); ++i) {
size_t sentenceId = sentenceIds[i];
if (m_oracles[sentenceId].size() > 1)
cerr << "Available oracles for source sentence " << sentenceId << ": " << m_oracles[sentenceId].size() << endl;
for (size_t j = 0; j < featureValues[i].size(); ++j) {
cerr << "Available oracles for source sentence " << sentenceId << ": "
<< m_oracles[sentenceId].size() << endl;
for (size_t j = 0; j < featureValues[i].size(); ++j) {
// check if optimisation criterion is violated for one hypothesis and the oracle
// h(e*) >= h(e_ij) + loss(e_ij)
// h(e*) - h(e_ij) >= loss(e_ij)
@ -84,15 +82,15 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
loss *= log10(bleuScores[i][j]);
}
cerr << "Rank " << rank << ", loss: " << loss << ", model score diff: " << modelScoreDiff << endl;
//cerr << "Rank " << rank << ", loss: " << loss << ", model score diff: "
// << modelScoreDiff << endl;
bool addConstraint = true;
if (modelScoreDiff < loss) {
// constraint violated
++violatedConstraintsBefore;
oldDistanceFromOptimum += (loss - modelScoreDiff);
}
else if (m_onlyViolatedConstraints) {
} else if (m_onlyViolatedConstraints) {
// constraint not violated
addConstraint = false;
}
@ -100,13 +98,13 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
if (addConstraint) {
float lossMarginDistance = loss - modelScoreDiff;
if (m_accumulateMostViolatedConstraints && !m_pastAndCurrentConstraints) {
if (m_accumulateMostViolatedConstraints
&& !m_pastAndCurrentConstraints) {
if (lossMarginDistance > maxViolationLossMarginDistance) {
maxViolationLossMarginDistance = lossMarginDistance;
maxViolationfeatureValueDiff = featureValueDiff;
}
}
else if (m_pastAndCurrentConstraints) {
} else if (m_pastAndCurrentConstraints) {
if (lossMarginDistance > maxViolationLossMarginDistance) {
maxViolationLossMarginDistance = lossMarginDistance;
maxViolationfeatureValueDiff = featureValueDiff;
@ -114,8 +112,7 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
featureValueDiffs.push_back(featureValueDiff);
lossMarginDistances.push_back(lossMarginDistance);
}
else {
} else {
// Objective: 1/2 * ||w' - w||^2 + C * SUM_1_m[ max_1_n (l_ij - Delta_h_ij.w')]
// To add a constraint for the optimiser for each sentence i and hypothesis j, we need:
// 1. vector Delta_h_ij of the feature value differences (oracle - hypothesis)
@ -135,22 +132,24 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
}
}
cerr << "Number of violated constraints before optimisation: " << violatedConstraintsBefore << endl;
cerr << "Number of violated constraints before optimisation: "
<< violatedConstraintsBefore << endl;
if (featureValueDiffs.size() != 30) {
cerr << "Number of constraints passed to optimiser: " << featureValueDiffs.size() << endl;
cerr << "Number of constraints passed to optimiser: "
<< featureValueDiffs.size() << endl;
}
// run optimisation: compute alphas for all given constraints
vector< float> alphas;
vector<float> alphas;
ScoreComponentCollection totalUpdate;
if (m_accumulateMostViolatedConstraints && !m_pastAndCurrentConstraints) {
m_featureValueDiffs.push_back(maxViolationfeatureValueDiff);
m_lossMarginDistances.push_back(maxViolationLossMarginDistance);
if (m_slack != 0) {
alphas = Hildreth::optimise(m_featureValueDiffs, m_lossMarginDistances, m_slack);
}
else {
alphas = Hildreth::optimise(m_featureValueDiffs, m_lossMarginDistances,
m_slack);
} else {
alphas = Hildreth::optimise(m_featureValueDiffs, m_lossMarginDistances);
}
@ -164,8 +163,7 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
// accumulate update
totalUpdate.PlusEquals(m_featureValueDiffs[k]);
}
}
else if (violatedConstraintsBefore > 0) {
} else if (violatedConstraintsBefore > 0) {
if (m_pastAndCurrentConstraints) {
// add all (most violated) past constraints to the list of current constraints
for (size_t i = 0; i < m_featureValueDiffs.size(); ++i) {
@ -179,9 +177,9 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
}
if (m_slack != 0) {
alphas = Hildreth::optimise(featureValueDiffs, lossMarginDistances, m_slack);
}
else {
alphas = Hildreth::optimise(featureValueDiffs, lossMarginDistances,
m_slack);
} else {
alphas = Hildreth::optimise(featureValueDiffs, lossMarginDistances);
}
@ -195,59 +193,70 @@ int MiraOptimiser::updateWeights(ScoreComponentCollection& currWeights,
// accumulate update
totalUpdate.PlusEquals(featureValueDiffs[k]);
}
}
else {
} else {
cerr << "No constraint violated for this batch" << endl;
return 0;
}
// apply learning rate (fixed or flexible)
if (learning_rate != 1) {
cerr << "Rank " << rank << ", update before applying learning rate: " << totalUpdate << endl;
cerr << "Rank " << rank << ", update before applying learning rate: "
<< totalUpdate << endl;
totalUpdate.MultiplyEquals(learning_rate);
cerr << "Rank " << rank << ", update after applying learning rate: " << totalUpdate << endl;
cerr << "Rank " << rank << ", update after applying learning rate: "
<< totalUpdate << endl;
}
// apply threshold scaling
if (max_sentence_update != -1) {
cerr << "Rank " << rank << ", update before scaling to max-sentence-update: " << totalUpdate << endl;
cerr << "Rank " << rank
<< ", update before scaling to max-sentence-update: " << totalUpdate
<< endl;
totalUpdate.ThresholdScaling(max_sentence_update);
cerr << "Rank " << rank << ", update after scaling to max-sentence-update: " << totalUpdate << endl;
cerr << "Rank " << rank
<< ", update after scaling to max-sentence-update: " << totalUpdate
<< endl;
}
// apply update to weight vector
cerr << "Rank " << rank << ", weights before update: " << currWeights << endl;
currWeights.PlusEquals(totalUpdate);
cerr << "Rank " << rank << ", weights after update: " << currWeights << endl;
if (update_after_epoch) {
m_accumulatedUpdates.PlusEquals(totalUpdate);
cerr << "Rank " << rank << ", new accumulated updates:" << m_accumulatedUpdates << endl;
} else {
// apply update to weight vector
cerr << "Rank " << rank << ", weights before update: " << currWeights
<< endl;
currWeights.PlusEquals(totalUpdate);
cerr << "Rank " << rank << ", weights after update: " << currWeights
<< endl;
// sanity check: how many constraints violated after optimisation?
size_t violatedConstraintsAfter = 0;
float newDistanceFromOptimum = 0;
for (size_t i = 0; i < featureValues.size(); ++i) {
for (size_t j = 0; j < featureValues[i].size(); ++j) {
ScoreComponentCollection featureValueDiff = oracleFeatureValues[i];
featureValueDiff.MinusEquals(featureValues[i][j]);
float modelScoreDiff = featureValueDiff.InnerProduct(currWeights);
float loss = losses[i][j] * m_marginScaleFactor;
if (modelScoreDiff < loss) {
++violatedConstraintsAfter;
newDistanceFromOptimum += (loss - modelScoreDiff);
// sanity check: how many constraints violated after optimisation?
size_t violatedConstraintsAfter = 0;
float newDistanceFromOptimum = 0;
for (size_t i = 0; i < featureValues.size(); ++i) {
for (size_t j = 0; j < featureValues[i].size(); ++j) {
ScoreComponentCollection featureValueDiff = oracleFeatureValues[i];
featureValueDiff.MinusEquals(featureValues[i][j]);
float modelScoreDiff = featureValueDiff.InnerProduct(currWeights);
float loss = losses[i][j] * m_marginScaleFactor;
if (modelScoreDiff < loss) {
++violatedConstraintsAfter;
newDistanceFromOptimum += (loss - modelScoreDiff);
}
}
}
}
int constraintChange = violatedConstraintsBefore - violatedConstraintsAfter;
cerr << "Rank " << rank << ", constraint change: " << constraintChange << " (before: " << violatedConstraintsBefore << ")" << endl;
float distanceChange = oldDistanceFromOptimum - newDistanceFromOptimum;
cerr << "Rank " << rank << ", distance change: " << distanceChange << endl;
if (constraintChange < 0 && distanceChange < 0) {
return -1;
int constraintChange = violatedConstraintsBefore - violatedConstraintsAfter;
cerr << "Rank " << rank << ", constraint change: " << constraintChange
<< " (before: " << violatedConstraintsBefore << ")" << endl;
float distanceChange = oldDistanceFromOptimum - newDistanceFromOptimum;
cerr << "Rank " << rank << ", distance change: " << distanceChange << endl;
if (constraintChange < 0 && distanceChange < 0) {
return -1;
}
}
return 0;
}
}

56
mira/Optimiser.h
View File

@ -30,37 +30,21 @@ namespace Mira {
public:
Optimiser() {}
virtual int updateWeights(Moses::ScoreComponentCollection& weights,
const std::vector<std::vector<Moses::ScoreComponentCollection> >& featureValues,
const std::vector<std::vector<float> >& losses,
const std::vector<std::vector<float> >& bleuScores,
const std::vector<Moses::ScoreComponentCollection>& oracleFeatureValues,
const std::vector< float> oracleBleuScores,
const std::vector< size_t> dummy,
float learning_rate,
float max_sentence_update,
size_t rank) = 0;
};
class DummyOptimiser : public Optimiser {
public:
virtual int updateWeights(Moses::ScoreComponentCollection& weights,
const std::vector< std::vector<Moses::ScoreComponentCollection> >& featureValues,
const std::vector< std::vector<float> >& losses,
const std::vector<std::vector<float> >& bleuScores,
const std::vector<Moses::ScoreComponentCollection>& oracleFeatureValues,
const std::vector< float> oracleBleuScores,
const std::vector< size_t> dummy,
float learning_rate,
float max_sentence_update,
size_t rank)
{ return 0; }
const std::vector< std::vector<Moses::ScoreComponentCollection> >& featureValues,
const std::vector< std::vector<float> >& losses,
const std::vector<std::vector<float> >& bleuScores,
const std::vector< Moses::ScoreComponentCollection>& oracleFeatureValues,
const std::vector< float> oracleBleuScores,
const std::vector< size_t> sentenceId,
float learning_rate,
float max_sentence_update,
size_t rank,
bool update_after_epoch) = 0;
};
class Perceptron : public Optimiser {
public:
virtual int updateWeights(Moses::ScoreComponentCollection& weights,
virtual int updateWeights(Moses::ScoreComponentCollection& weights,
const std::vector< std::vector<Moses::ScoreComponentCollection> >& featureValues,
const std::vector< std::vector<float> >& losses,
const std::vector<std::vector<float> >& bleuScores,
@ -69,7 +53,8 @@ namespace Mira {
const std::vector< size_t> dummy,
float learning_rate,
float max_sentence_update,
size_t rank);
size_t rank,
bool update_after_epoch);
};
class MiraOptimiser : public Optimiser {
@ -93,7 +78,7 @@ namespace Mira {
~MiraOptimiser() {}
virtual int updateWeights(Moses::ScoreComponentCollection& weights,
virtual int updateWeights(Moses::ScoreComponentCollection& weights,
const std::vector< std::vector<Moses::ScoreComponentCollection> >& featureValues,
const std::vector< std::vector<float> >& losses,
const std::vector<std::vector<float> >& bleuScores,
@ -102,7 +87,8 @@ namespace Mira {
const std::vector< size_t> sentenceId,
float learning_rate,
float max_sentence_update,
size_t rank);
size_t rank,
bool update_after_epoch);
void setOracleIndices(std::vector<size_t> oracleIndices) {
m_oracleIndices= oracleIndices;
@ -115,6 +101,14 @@ namespace Mira {
void setMarginScaleFactor(float msf) {
m_marginScaleFactor = msf;
}
Moses::ScoreComponentCollection getAccumulatedUpdates() {
return m_accumulatedUpdates;
}
void resetAccumulatedUpdates() {
m_accumulatedUpdates.ZeroAll();
}
private:
// number of hypotheses used for each nbest list (number of hope, fear, best model translations)
@ -151,6 +145,8 @@ namespace Mira {
bool m_accumulateMostViolatedConstraints;
bool m_pastAndCurrentConstraints;
Moses::ScoreComponentCollection m_accumulatedUpdates;
};
}

3
mira/Perceptron.cpp
View File

@ -33,7 +33,8 @@ int Perceptron::updateWeights(ScoreComponentCollection& currWeights,
const vector< size_t> dummy,
float learning_rate,
float max_sentence_update,
size_t rank)
size_t rank,
bool update_after_epoch)
{
for (size_t i = 0; i < featureValues.size(); ++i) {
for (size_t j = 0; j < featureValues[i].size(); ++j) {