private void testZeroTrainingError(
GradientOracle<DynamicFactorGraph, Example<DynamicAssignment, DynamicAssignment>> oracle,
boolean useLbfgs) {
MapReduceConfiguration.setMapReduceExecutor(new LocalMapReduceExecutor(1, 1));
SufficientStatistics parameters = null;
if (useLbfgs) {
RetryingLbfgs trainer = new RetryingLbfgs(50, 10, 0.1, new DefaultLogFunction(1, false));
parameters = trainer.train(oracle, sequenceModel.getNewSufficientStatistics(), trainingData);
} else {
StochasticGradientTrainer trainer =
StochasticGradientTrainer.createWithL2Regularization(
100, 1, 1.0, true, false, 0.1, new DefaultLogFunction());
parameters = trainer.train(oracle, sequenceModel.getNewSufficientStatistics(), trainingData);
}
DynamicFactorGraph trainedModel = sequenceModel.getModelFromParameters(parameters);
// Should be able to get 0 training error.
FactorGraphPredictor predictor =
new FactorGraphPredictor(
trainedModel,
VariableNamePattern.fromTemplateVariables(y, VariableNumMap.EMPTY),
new JunctionTree());
for (Example<DynamicAssignment, DynamicAssignment> trainingDatum : trainingData) {
DynamicAssignment prediction =
predictor.getBestPrediction(trainingDatum.getInput()).getBestPrediction();
assertEquals(trainingDatum.getOutput(), prediction);
}
}
@Override
public Predictor<I, O> train(Iterable<Example<I, O>> trainingData) {
ParametricFactorGraph model = constructGraphicalModel(trainingData);
Converter<I, DynamicAssignment> inputConverter = getInputConverter(model);
Converter<O, DynamicAssignment> outputConverter = getOutputConverter(model);
Converter<Example<I, O>, Example<DynamicAssignment, DynamicAssignment>> exampleConverter =
Example.converter(inputConverter, outputConverter);
List<Example<DynamicAssignment, DynamicAssignment>> trainingDataAssignments =
Lists.newArrayList(Iterables.transform(trainingData, exampleConverter));
SufficientStatistics initialParameters = model.getNewSufficientStatistics();
initialParameters.perturb(parameterPerturbation);
System.out.println(initialParameters);
SufficientStatistics finalParameters =
trainer.train(model, initialParameters, trainingDataAssignments);
Predictor<DynamicAssignment, DynamicAssignment> assignmentPredictor =
new FactorGraphPredictor(
model.getModelFromParameters(finalParameters),
getOutputVariables(model),
new JunctionTree());
return ForwardingPredictor.create(assignmentPredictor, inputConverter, outputConverter);
}