public PairwiseModel(PairwiseModel.Builder builder)
{
name = builder.name;
meta = builder.meta;
trainingExamples = builder.trainingExamples;
epochs = builder.epochs;
singletonRatio = builder.singletonRatio;
classifier = new SimpleLinearClassifier(builder.loss, builder.learningRateSchedule, builder.regularizationStrength, builder.modelFile == null ? null : ((builder.modelFile.EndsWith(".ser") || builder.modelFile.EndsWith(".gz")) ? builder.modelFile
: StatisticalCorefTrainer.pairwiseModelsPath + builder.modelFile + "/model.ser"));
str = StatisticalCorefTrainer.FieldValues(builder);
}
public MetaFeatureExtractor(MetaFeatureExtractor.Builder builder)
{
anaphoricityClassifier = builder.anaphoricityClassifier;
if (anaphoricityClassifier)
{
pairConjunctions = new HashSet <MetaFeatureExtractor.PairConjunction>();
}
else
{
pairConjunctions = new HashSet <MetaFeatureExtractor.PairConjunction>(builder.pairConjunctions);
}
singleConjunctions = new HashSet <MetaFeatureExtractor.SingleConjunction>(builder.singleConjunctions);
disallowedPrefixes = builder.disallowedPrefixes;
neTypeConjuntion = builder.useNEType;
str = StatisticalCorefTrainer.FieldValues(builder);
}
public virtual void DoTraining(string modelName)
{
classifier.SetWeight("bias", -0.3);
classifier.SetWeight("anaphorSeen", -1);
classifier.SetWeight("max-ranking", 1);
classifier.SetWeight("bias-single", -0.3);
classifier.SetWeight("anaphorSeen-single", -1);
classifier.SetWeight("max-ranking-single", 1);
string outputPath = StatisticalCorefTrainer.clusteringModelsPath + modelName + "/";
File outDir = new File(outputPath);
if (!outDir.Exists())
{
outDir.Mkdir();
}
PrintWriter progressWriter;
IList <ClustererDataLoader.ClustererDoc> trainDocs;
try
{
PrintWriter configWriter = new PrintWriter(outputPath + "config", "UTF-8");
configWriter.Print(StatisticalCorefTrainer.FieldValues(this));
configWriter.Close();
progressWriter = new PrintWriter(outputPath + "progress", "UTF-8");
Redwood.Log("scoref.train", "Loading training data");
StatisticalCorefTrainer.SetDataPath("dev");
trainDocs = ClustererDataLoader.LoadDocuments(MaxDocs);
}
catch (Exception e)
{
throw new Exception("Error setting up training", e);
}
double bestTrainScore = 0;
IList <IList <Pair <Clusterer.CandidateAction, Clusterer.CandidateAction> > > examples = new List <IList <Pair <Clusterer.CandidateAction, Clusterer.CandidateAction> > >();
for (int iteration = 0; iteration < RetrainIterations; iteration++)
{
Redwood.Log("scoref.train", "ITERATION " + iteration);
classifier.PrintWeightVector(null);
Redwood.Log("scoref.train", string.Empty);
try
{
classifier.WriteWeights(outputPath + "model");
classifier.PrintWeightVector(IOUtils.GetPrintWriter(outputPath + "weights"));
}
catch (Exception)
{
throw new Exception();
}
long start = Runtime.CurrentTimeMillis();
Java.Util.Collections.Shuffle(trainDocs, random);
examples = examples.SubList(Math.Max(0, examples.Count - BufferSizeMultiplier * trainDocs.Count), examples.Count);
TrainPolicy(examples);
if (iteration % EvalFrequency == 0)
{
double trainScore = EvaluatePolicy(trainDocs, true);
if (trainScore > bestTrainScore)
{
bestTrainScore = trainScore;
WriteModel("best", outputPath);
}
if (iteration % 10 == 0)
{
WriteModel("iter_" + iteration, outputPath);
}
WriteModel("last", outputPath);
double timeElapsed = (Runtime.CurrentTimeMillis() - start) / 1000.0;
double ffhr = Clusterer.State.ffHits / (double)(Clusterer.State.ffHits + Clusterer.State.ffMisses);
double shr = Clusterer.State.sHits / (double)(Clusterer.State.sHits + Clusterer.State.sMisses);
double fhr = featuresCacheHits / (double)(featuresCacheHits + featuresCacheMisses);
Redwood.Log("scoref.train", modelName);
Redwood.Log("scoref.train", string.Format("Best train: %.4f", bestTrainScore));
Redwood.Log("scoref.train", string.Format("Time elapsed: %.2f", timeElapsed));
Redwood.Log("scoref.train", string.Format("Cost hit rate: %.4f", ffhr));
Redwood.Log("scoref.train", string.Format("Score hit rate: %.4f", shr));
Redwood.Log("scoref.train", string.Format("Features hit rate: %.4f", fhr));
Redwood.Log("scoref.train", string.Empty);
progressWriter.Write(iteration + " " + trainScore + " " + " " + timeElapsed + " " + ffhr + " " + shr + " " + fhr + "\n");
progressWriter.Flush();
}
foreach (ClustererDataLoader.ClustererDoc trainDoc in trainDocs)
{
examples.Add(RunPolicy(trainDoc, Math.Pow(ExpertDecay, (iteration + 1))));
}
}
progressWriter.Close();
}
