private IEnumerable <IList <TranslationInfo> > GetNBestLists(ThotSmtParameters parameters,
IReadOnlyList <IReadOnlyList <string> > sourceCorpus)
{
var results = new IList <TranslationInfo> [sourceCorpus.Count];
Parallel.ForEach(Partitioner.Create(0, sourceCorpus.Count), range =>
{
IntPtr smtModelHandle = IntPtr.Zero, decoderHandle = IntPtr.Zero;
try
{
smtModelHandle = Thot.LoadSmtModel(parameters);
decoderHandle = Thot.LoadDecoder(smtModelHandle, parameters);
for (int i = range.Item1; i < range.Item2; i++)
{
IReadOnlyList <string> sourceSegment = sourceCorpus[i];
results[i] = Thot.DoTranslateNBest(decoderHandle, Thot.decoder_translateNBest, K,
sourceSegment, false, sourceSegment, CreateTranslationInfo).ToArray();
}
}
finally
{
if (decoderHandle != IntPtr.Zero)
{
Thot.decoder_close(decoderHandle);
}
if (smtModelHandle != IntPtr.Zero)
{
Thot.smtModel_close(smtModelHandle);
}
}
});
return(results);
}
public BatchTrainer(ThotSmtModel smtModel, ThotSmtParameters parameters,
Func <string, string> sourcePreprocessor, Func <string, string> targetPreprocessor,
ParallelTextCorpus corpus)
: base(parameters, sourcePreprocessor, targetPreprocessor, corpus)
{
_smtModel = smtModel;
}
public void TrainModels_EmptyCorpus_GeneratesModels()
{
using (var tempDir = new TempDirectory("ThotSmtEngineTests"))
{
var sourceCorpus = new DictionaryTextCorpus(Enumerable.Empty <MemoryText>());
var targetCorpus = new DictionaryTextCorpus(Enumerable.Empty <MemoryText>());
var alignmentCorpus = new DictionaryTextAlignmentCorpus(
Enumerable.Empty <MemoryTextAlignmentCollection>());
var corpus = new ParallelTextCorpus(sourceCorpus, targetCorpus, alignmentCorpus);
var parameters = new ThotSmtParameters
{
TranslationModelFileNamePrefix = Path.Combine(tempDir.Path, "tm", "src_trg"),
LanguageModelFileNamePrefix = Path.Combine(tempDir.Path, "lm", "trg.lm")
};
using (var trainer = new ThotSmtBatchTrainer(parameters, s => s, s => s, corpus))
{
trainer.Train();
trainer.Save();
}
Assert.That(File.Exists(Path.Combine(tempDir.Path, "lm", "trg.lm")), Is.True);
Assert.That(File.Exists(Path.Combine(tempDir.Path, "tm", "src_trg_swm.hmm_alignd")), Is.True);
Assert.That(File.Exists(Path.Combine(tempDir.Path, "tm", "src_trg_invswm.hmm_alignd")), Is.True);
Assert.That(File.Exists(Path.Combine(tempDir.Path, "tm", "src_trg.ttable")), Is.True);
// TODO: test for more than just existence of files
}
}
public static IntPtr LoadDecoder(IntPtr smtModelHandle, ThotSmtParameters parameters)
{
IntPtr handle = decoder_create(smtModelHandle);
decoder_setS(handle, parameters.DecoderS);
decoder_setBreadthFirst(handle, parameters.DecoderBreadthFirst);
decoder_setG(handle, parameters.DecoderG);
return(handle);
}
public ThotSmtModel(ThotSmtParameters parameters)
{
Parameters = parameters;
Parameters.Freeze();
Handle = Thot.LoadSmtModel(Parameters);
_directWordAlignmentModel = new ThotWordAlignmentModel(
Thot.smtModel_getSingleWordAlignmentModel(Handle));
_inverseWordAlignmentModel = new ThotWordAlignmentModel(
Thot.smtModel_getInverseSingleWordAlignmentModel(Handle));
}
public static IntPtr LoadSmtModel(ThotSmtParameters parameters)
{
IntPtr handle = smtModel_create();
smtModel_loadTranslationModel(handle, parameters.TranslationModelFileNamePrefix);
smtModel_loadLanguageModel(handle, parameters.LanguageModelFileNamePrefix);
smtModel_setNonMonotonicity(handle, parameters.ModelNonMonotonicity);
smtModel_setW(handle, parameters.ModelW);
smtModel_setA(handle, parameters.ModelA);
smtModel_setE(handle, parameters.ModelE);
smtModel_setHeuristic(handle, (uint)parameters.ModelHeuristic);
smtModel_setOnlineTrainingParameters(handle, (uint)parameters.LearningAlgorithm,
(uint)parameters.LearningRatePolicy, parameters.LearningStepSize, parameters.LearningEMIters,
parameters.LearningE, parameters.LearningR);
if (parameters.ModelWeights != null)
{
smtModel_setWeights(handle, parameters.ModelWeights.ToArray(), (uint)parameters.ModelWeights.Count);
}
return(handle);
}
private static double CalculateBleu(ThotSmtParameters parameters,
IReadOnlyList <IReadOnlyList <string> > sourceCorpus, IReadOnlyList <IReadOnlyList <string> > tuneTargetCorpus)
{
IEnumerable <IReadOnlyList <string> > translations = GenerateTranslations(parameters, sourceCorpus);
double bleu = Evaluation.CalculateBleu(translations, tuneTargetCorpus);
double penalty = 0;
for (int i = 0; i < parameters.ModelWeights.Count; i++)
{
if (i == 0 || i == 2 || i == 7)
{
continue;
}
if (parameters.ModelWeights[i] < 0)
{
penalty += parameters.ModelWeights[i] * 1000 * -1;
}
}
return((1.0 - bleu) + penalty);
}
private static IEnumerable <IReadOnlyList <string> > GenerateTranslations(ThotSmtParameters parameters,
IReadOnlyList <IReadOnlyList <string> > sourceCorpus)
{
IntPtr smtModelHandle = IntPtr.Zero;
try
{
smtModelHandle = Thot.LoadSmtModel(parameters);
var results = new IReadOnlyList <string> [sourceCorpus.Count];
Parallel.ForEach(Partitioner.Create(0, sourceCorpus.Count), range =>
{
IntPtr decoderHandle = IntPtr.Zero;
try
{
decoderHandle = Thot.LoadDecoder(smtModelHandle, parameters);
for (int i = range.Item1; i < range.Item2; i++)
{
IReadOnlyList <string> segment = sourceCorpus[i];
results[i] = Thot.DoTranslate(decoderHandle, Thot.decoder_translate, segment, false,
segment, (s, t, d) => t);
}
}
finally
{
if (decoderHandle != IntPtr.Zero)
{
Thot.decoder_close(decoderHandle);
}
}
});
return(results);
}
finally
{
if (smtModelHandle != IntPtr.Zero)
{
Thot.smtModel_close(smtModelHandle);
}
}
}
public ThotSmtParameters Tune(ThotSmtParameters parameters,
IReadOnlyList <IReadOnlyList <string> > tuneSourceCorpus,
IReadOnlyList <IReadOnlyList <string> > tuneTargetCorpus, ThotTrainProgressReporter reporter,
SmtBatchTrainStats stats)
{
float sentLenWeight = parameters.ModelWeights[7];
int numFuncEvals = 0;
double Evaluate(Vector weights)
{
ThotSmtParameters newParameters = parameters.Clone();
newParameters.ModelWeights = weights.Select(w => (float)w).Concat(sentLenWeight).ToArray();
newParameters.Freeze();
double quality = CalculateBleu(newParameters, tuneSourceCorpus, tuneTargetCorpus);
numFuncEvals++;
if (numFuncEvals < MaxFunctionEvaluations && ProgressIncrementInterval > 0 &&
numFuncEvals % ProgressIncrementInterval == 0)
{
reporter.Step();
}
else
{
reporter.CheckCanceled();
}
return(quality);
};
var simplex = new NelderMeadSimplex(ConvergenceTolerance, MaxFunctionEvaluations, 1.0);
MinimizationResult result = simplex.FindMinimum(Evaluate,
parameters.ModelWeights.Select(w => (double)w).Take(7));
stats.TranslationModelBleu = 1.0 - result.ErrorValue;
ThotSmtParameters bestParameters = parameters.Clone();
bestParameters.ModelWeights = result.MinimizingPoint.Select(w => (float)w).Concat(sentLenWeight).ToArray();
bestParameters.Freeze();
return(bestParameters);
}
public ThotSmtParameters Tune(ThotSmtParameters parameters,
IReadOnlyList <IReadOnlyList <string> > tuneSourceCorpus,
IReadOnlyList <IReadOnlyList <string> > tuneTargetCorpus, SmtBatchTrainStats stats,
IProgress <ProgressStatus> progress)
{
float sentLenWeight = parameters.ModelWeights[7];
int numFuncEvals = 0;
double Evaluate(Vector weights)
{
ThotSmtParameters newParameters = parameters.Clone();
newParameters.ModelWeights = weights.Select(w => (float)w).Concat(sentLenWeight).ToArray();
newParameters.Freeze();
double quality = CalculateBleu(newParameters, tuneSourceCorpus, tuneTargetCorpus);
numFuncEvals++;
int currentStep = Math.Min(numFuncEvals, MaxProgressFunctionEvaluations);
progress.Report(new ProgressStatus(currentStep, MaxProgressFunctionEvaluations));
return(quality);
};
progress.Report(new ProgressStatus(0, MaxFunctionEvaluations));
var simplex = new NelderMeadSimplex(ConvergenceTolerance, MaxFunctionEvaluations, 1.0);
MinimizationResult result = simplex.FindMinimum(Evaluate,
parameters.ModelWeights.Select(w => (double)w).Take(7));
stats.TranslationModelBleu = 1.0 - result.ErrorValue;
ThotSmtParameters bestParameters = parameters.Clone();
bestParameters.ModelWeights = result.MinimizingPoint.Select(w => (float)w).Concat(sentLenWeight).ToArray();
bestParameters.Freeze();
if (result.FunctionEvaluationCount < MaxProgressFunctionEvaluations)
{
progress.Report(new ProgressStatus(1.0));
}
return(bestParameters);
}
private void TrainTuneCorpus(string trainTMPrefix, string trainLMPrefix,
IReadOnlyList <IReadOnlyList <string> > tuneSourceCorpus,
IReadOnlyList <IReadOnlyList <string> > tuneTargetCorpus, ThotTrainProgressReporter reporter)
{
reporter.Step("Finalizing", TrainingStepCount - 1);
if (tuneSourceCorpus.Count == 0)
{
return;
}
ThotSmtParameters parameters = Parameters.Clone();
parameters.TranslationModelFileNamePrefix = trainTMPrefix;
parameters.LanguageModelFileNamePrefix = trainLMPrefix;
using (var smtModel = new ThotSmtModel(parameters))
using (ISmtEngine engine = smtModel.CreateEngine())
{
for (int i = 0; i < tuneSourceCorpus.Count; i++)
{
engine.TrainSegment(tuneSourceCorpus[i], tuneTargetCorpus[i]);
}
}
}
private void TrainTuneCorpus(string trainTMPrefix, string trainLMPrefix,
IReadOnlyList <IReadOnlyList <string> > tuneSourceCorpus,
IReadOnlyList <IReadOnlyList <string> > tuneTargetCorpus, IProgress <ProgressStatus> progress)
{
if (tuneSourceCorpus.Count == 0)
{
return;
}
ThotSmtParameters parameters = Parameters.Clone();
parameters.TranslationModelFileNamePrefix = trainTMPrefix;
parameters.LanguageModelFileNamePrefix = trainLMPrefix;
using (var smtModel = new ThotSmtModel(parameters))
using (ISmtEngine engine = smtModel.CreateEngine())
{
for (int i = 0; i < tuneSourceCorpus.Count; i++)
{
progress.Report(new ProgressStatus(i, tuneSourceCorpus.Count));
engine.TrainSegment(tuneSourceCorpus[i], tuneTargetCorpus[i]);
}
progress.Report(new ProgressStatus(1.0));
}
}
public ThotSmtBatchTrainer(ThotSmtParameters parameters, Func <string, string> sourcePreprocessor,
Func <string, string> targetPreprocessor, ParallelTextCorpus corpus, int maxCorpusCount = int.MaxValue)
{
Parameters = parameters;
Parameters.Freeze();
_sourcePreprocessor = sourcePreprocessor;
_targetPreprocessor = targetPreprocessor;
_maxCorpusCount = maxCorpusCount;
_parallelCorpus = corpus;
//_modelWeightTuner = new MiraModelWeightTuner();
_modelWeightTuner = new SimplexModelWeightTuner();
_tuneCorpusIndices = CreateTuneCorpus();
do
{
_tempDir = Path.Combine(Path.GetTempPath(), "thot-train-" + Guid.NewGuid());
} while (Directory.Exists(_tempDir));
Directory.CreateDirectory(_tempDir);
_lmFilePrefix = Path.GetFileName(Parameters.LanguageModelFileNamePrefix);
_tmFilePrefix = Path.GetFileName(Parameters.TranslationModelFileNamePrefix);
_trainLMDir = Path.Combine(_tempDir, "lm");
_trainTMDir = Path.Combine(_tempDir, "tm_train");
}
public ThotSmtModel(string cfgFileName)
: this(ThotSmtParameters.Load(cfgFileName))
{
ConfigFileName = cfgFileName;
}
public ThotSmtBatchTrainer(string cfgFileName, Func <string, string> sourcePreprocessor,
Func <string, string> targetPreprocessor, ParallelTextCorpus corpus, int maxCorpusCount = int.MaxValue)
: this(ThotSmtParameters.Load(cfgFileName), sourcePreprocessor, targetPreprocessor, corpus, maxCorpusCount)
{
ConfigFileName = cfgFileName;
}
public ThotSmtParameters Tune(ThotSmtParameters parameters,
IReadOnlyList <IReadOnlyList <string> > tuneSourceCorpus,
IReadOnlyList <IReadOnlyList <string> > tuneTargetCorpus, SmtBatchTrainStats stats,
IProgress <ProgressStatus> progress)
{
IntPtr weightUpdaterHandle = Thot.llWeightUpdater_create();
try
{
var iterQualities = new List <double>();
double bestQuality = double.MinValue;
ThotSmtParameters bestParameters = null;
int iter = 0;
HashSet <TranslationInfo>[] curNBestLists = null;
float[] curWeights = parameters.ModelWeights.ToArray();
while (true)
{
progress.Report(new ProgressStatus(iter, MaxIterations));
ThotSmtParameters newParameters = parameters.Clone();
newParameters.ModelWeights = curWeights;
newParameters.Freeze();
IList <TranslationInfo>[] nbestLists = GetNBestLists(newParameters, tuneSourceCorpus).ToArray();
double quality = Evaluation.CalculateBleu(nbestLists.Select(nbl => nbl.First().Translation),
tuneTargetCorpus);
iterQualities.Add(quality);
if (quality > bestQuality)
{
bestQuality = quality;
bestParameters = newParameters;
}
iter++;
if (iter >= MaxIterations || IsTuningConverged(iterQualities))
{
break;
}
if (curNBestLists == null)
{
curNBestLists = nbestLists.Select(nbl => new HashSet <TranslationInfo>(nbl)).ToArray();
}
else
{
for (int i = 0; i < nbestLists.Length; i++)
{
curNBestLists[i].UnionWith(nbestLists[i]);
}
}
UpdateWeights(weightUpdaterHandle, tuneTargetCorpus, curNBestLists, curWeights);
}
if (iter < MaxIterations)
{
progress.Report(new ProgressStatus(1.0));
}
stats.TranslationModelBleu = bestQuality;
return(bestParameters);
}
finally
{
Thot.llWeightUpdater_close(weightUpdaterHandle);
}
}
public void Train_NonEmptyCorpus_GeneratesModels()
{
using (var tempDir = new TempDirectory("ThotSmtEngineTests"))
{
var sourceCorpus = new DictionaryTextCorpus(new[]
{
new MemoryText("text1", new[]
{
new TextSegment(new TextSegmentRef(1, 1),
"¿ le importaría darnos las llaves de la habitación , por favor ?".Split()),
new TextSegment(new TextSegmentRef(1, 2),
"he hecho la reserva de una habitación tranquila doble con teléfono y televisión a nombre de rosario cabedo .".Split()),
new TextSegment(new TextSegmentRef(1, 3),
"¿ le importaría cambiarme a otra habitación más tranquila ?".Split()),
new TextSegment(new TextSegmentRef(1, 4),
"por favor , tengo reservada una habitación .".Split()),
new TextSegment(new TextSegmentRef(1, 5), "me parece que existe un problema .".Split())
})
});
var targetCorpus = new DictionaryTextCorpus(new[]
{
new MemoryText("text1", new[]
{
new TextSegment(new TextSegmentRef(1, 1),
"would you mind giving us the keys to the room , please ?".Split()),
new TextSegment(new TextSegmentRef(1, 2),
"i have made a reservation for a quiet , double room with a telephone and a tv for rosario cabedo .".Split()),
new TextSegment(new TextSegmentRef(1, 3),
"would you mind moving me to a quieter room ?".Split()),
new TextSegment(new TextSegmentRef(1, 4), "i have booked a room .".Split()),
new TextSegment(new TextSegmentRef(1, 5), "i think that there is a problem .".Split())
})
});
var alignmentCorpus = new DictionaryTextAlignmentCorpus(new[]
{
new MemoryTextAlignmentCollection("text1", new[]
{
new TextAlignment(new TextSegmentRef(1, 1), new[] { new AlignedWordPair(8, 9) }),
new TextAlignment(new TextSegmentRef(1, 2), new[] { new AlignedWordPair(6, 10) }),
new TextAlignment(new TextSegmentRef(1, 3), new[] { new AlignedWordPair(6, 8) }),
new TextAlignment(new TextSegmentRef(1, 4), new[] { new AlignedWordPair(6, 4) }),
new TextAlignment(new TextSegmentRef(1, 5), new AlignedWordPair[0])
})
});
var corpus = new ParallelTextCorpus(sourceCorpus, targetCorpus, alignmentCorpus);
var parameters = new ThotSmtParameters
{
TranslationModelFileNamePrefix = Path.Combine(tempDir.Path, "tm", "src_trg"),
LanguageModelFileNamePrefix = Path.Combine(tempDir.Path, "lm", "trg.lm")
};
using (var trainer = new ThotSmtBatchTrainer(parameters, s => s, s => s, corpus))
{
trainer.Train();
trainer.Save();
}
Assert.That(File.Exists(Path.Combine(tempDir.Path, "lm", "trg.lm")), Is.True);
Assert.That(File.Exists(Path.Combine(tempDir.Path, "tm", "src_trg_swm.hmm_alignd")), Is.True);
Assert.That(File.Exists(Path.Combine(tempDir.Path, "tm", "src_trg_invswm.hmm_alignd")), Is.True);
Assert.That(File.Exists(Path.Combine(tempDir.Path, "tm", "src_trg.ttable")), Is.True);
// TODO: test for more than just existence of files
}
}
