private BayBoostModel retrainLastWeight(
BayBoostModel ensemble, ExampleSet exampleSet, Vector holdOutSet) throws OperatorException {
this.prepareExtendedBatch(exampleSet); // method fits by chance
int modelNum = ensemble.getNumberOfModels();
Vector<BayBoostBaseModelInfo> modelInfo = new Vector<BayBoostBaseModelInfo>();
double[] priors = ensemble.getPriors();
for (int i = 0; i < modelNum - 1; i++) {
Model model = ensemble.getModel(i);
ContingencyMatrix cm = ensemble.getContingencyMatrix(i);
modelInfo.add(new BayBoostBaseModelInfo(model, cm));
exampleSet = model.apply(exampleSet);
WeightedPerformanceMeasures.reweightExamples(exampleSet, cm, false);
}
Model latestModel = ensemble.getModel(modelNum - 1);
exampleSet = latestModel.apply(exampleSet);
// quite ugly:
double[] weights = new double[holdOutSet.size()];
Iterator it = holdOutSet.iterator();
int index = 0;
while (it.hasNext()) {
Example example = (Example) it.next();
weights[index++] = example.getWeight();
}
Iterator<Example> reader = exampleSet.iterator();
while (reader.hasNext()) {
reader.next().setWeight(0);
}
it = holdOutSet.iterator();
index = 0;
while (it.hasNext()) {
Example example = (Example) it.next();
example.setWeight(weights[index++]);
}
WeightedPerformanceMeasures wp = new WeightedPerformanceMeasures(exampleSet);
modelInfo.add(new BayBoostBaseModelInfo(latestModel, wp.getContingencyMatrix()));
return new BayBoostModel(exampleSet, modelInfo, priors);
}
/**
* Constructs a <code>Model</code> repeatedly running a weak learner, reweighting the training
* example set accordingly, and combining the hypothesis using the available weighted performance
* values.
*/
public Model learn(ExampleSet exampleSet) throws OperatorException {
this.runVector = new RunVector();
BayBoostModel ensembleNewBatch = null;
BayBoostModel ensembleExtBatch = null;
final Vector<BayBoostBaseModelInfo> modelInfo = new Vector<BayBoostBaseModelInfo>(); // for
// models
// and
// their
// probability
// estimates
Vector<BayBoostBaseModelInfo> modelInfo2 = new Vector<BayBoostBaseModelInfo>();
this.currentIteration = 0;
int firstOpenBatch = 1;
// prepare the stream control attribute
final Attribute streamControlAttribute;
{
Attribute attr = null;
if ((attr = exampleSet.getAttributes().get(STREAM_CONTROL_ATTRIB_NAME)) == null)
streamControlAttribute =
com.rapidminer.example.Tools.createSpecialAttribute(
exampleSet, STREAM_CONTROL_ATTRIB_NAME, Ontology.INTEGER);
else {
streamControlAttribute = attr;
logWarning(
"Attribute with the (reserved) name of the stream control attribute exists. It is probably an old version created by this operator. Trying to recycle it... ");
// Resetting the stream control attribute values by overwriting
// them with 0 avoids (unlikely)
// problems in case the same ExampleSet is passed to this
// operator over and over again:
Iterator<Example> e = exampleSet.iterator();
while (e.hasNext()) {
e.next().setValue(streamControlAttribute, 0);
}
}
}
// and the weight attribute
if (exampleSet.getAttributes().getWeight() == null) {
this.prepareWeights(exampleSet);
}
boolean estimateFavoursExtBatch = true;
// *** The main loop, one iteration per batch: ***
Iterator<Example> reader = exampleSet.iterator();
while (reader.hasNext()) {
// increment batch number, collect batch and evaluate performance of
// current model on batch
double[] classPriors =
this.prepareBatch(++this.currentIteration, reader, streamControlAttribute);
ConditionedExampleSet trainingSet =
new ConditionedExampleSet(
exampleSet, new BatchFilterCondition(streamControlAttribute, this.currentIteration));
final EstimatedPerformance estPerf;
// Step 1: apply the ensemble model to the current batch (prediction
// phase), evaluate and store result
if (ensembleExtBatch != null) {
// apply extended batch model first:
trainingSet = (ConditionedExampleSet) ensembleExtBatch.apply(trainingSet);
this.performance = evaluatePredictions(trainingSet); // unweighted
// performance;
// then apply new batch model:
trainingSet = (ConditionedExampleSet) ensembleNewBatch.apply(trainingSet);
double newBatchPerformance = evaluatePredictions(trainingSet);
// heuristic: use extended batch model for predicting
// unclassified instances
if (estimateFavoursExtBatch == true)
estPerf =
new EstimatedPerformance("accuracy", this.performance, trainingSet.size(), false);
else
estPerf =
new EstimatedPerformance("accuracy", newBatchPerformance, trainingSet.size(), false);
// final double[] ensembleWeights;
// continue with the better model:
if (newBatchPerformance > this.performance) {
this.performance = newBatchPerformance;
firstOpenBatch = Math.max(1, this.currentIteration - 1);
// ensembleWeights = ensembleNewBatch.getModelWeights();
} else {
modelInfo.clear();
modelInfo.addAll(modelInfo2);
// ensembleWeights = ensembleExtBatch.getModelWeights();
}
} else if (ensembleNewBatch != null) {
trainingSet = (ConditionedExampleSet) ensembleNewBatch.apply(trainingSet);
this.performance = evaluatePredictions(trainingSet);
firstOpenBatch = Math.max(1, this.currentIteration - 1);
estPerf = new EstimatedPerformance("accuracy", this.performance, trainingSet.size(), false);
} else estPerf = null; // no model ==> no prediction performance
if (estPerf != null) {
PerformanceVector perf = new PerformanceVector();
perf.addAveragable(estPerf);
this.runVector.addVector(perf);
}
// *** retraining phase ***
// Step 2: First reconstruct the initial weighting, if necessary.
if (this.getParameterAsBoolean(PARAMETER_RESCALE_LABEL_PRIORS) == true) {
this.rescalePriors(trainingSet, classPriors);
}
estimateFavoursExtBatch = true;
// Step 3: Find better weights for existing models and continue
// training
if (modelInfo.size() > 0) {
modelInfo2 = new Vector<BayBoostBaseModelInfo>();
for (BayBoostBaseModelInfo bbbmi : modelInfo) {
modelInfo2.add(bbbmi); // BayBoostBaseModelInfo objects
// cannot be changed, no deep copy
// required
}
// separate hold out set
final double holdOutRatio = this.getParameterAsDouble(PARAMETER_FRACTION_HOLD_OUT_SET);
Vector<Example> holdOutExamples = new Vector<Example>();
if (holdOutRatio > 0) {
RandomGenerator random = RandomGenerator.getRandomGenerator(this);
Iterator<Example> randBatchReader = trainingSet.iterator();
while (randBatchReader.hasNext()) {
Example example = randBatchReader.next();
if (random.nextDoubleInRange(0, 1) <= holdOutRatio) {
example.setValue(streamControlAttribute, 0);
holdOutExamples.add(example);
}
}
// TODO: create new example set
// trainingSet.updateCondition();
}
// model 1: train one more base classifier
boolean trainingExamplesLeft = this.adjustBaseModelWeights(trainingSet, modelInfo);
if (trainingExamplesLeft) {
// "trainingExamplesLeft" needs to be checked to avoid
// exceptions.
// Anyway, learning does not make sense, otherwise.
if (!this.trainAdditionalModel(trainingSet, modelInfo)) {}
}
ensembleNewBatch = new BayBoostModel(exampleSet, modelInfo, classPriors);
// model 2: remove last classifier, extend batch, train on
// extended batch
ExampleSet extendedBatch = // because of the ">=" condition it
// is sufficient to remember the
// opening batch
new ConditionedExampleSet(
exampleSet, new BatchFilterCondition(streamControlAttribute, firstOpenBatch));
classPriors = this.prepareExtendedBatch(extendedBatch);
if (this.getParameterAsBoolean(PARAMETER_RESCALE_LABEL_PRIORS) == true) {
this.rescalePriors(extendedBatch, classPriors);
}
modelInfo2.remove(modelInfo2.size() - 1);
trainingExamplesLeft = this.adjustBaseModelWeights(extendedBatch, modelInfo2);
// If no training examples are left: no need and chance to
// continue training.
if (trainingExamplesLeft == false) {
ensembleExtBatch = new BayBoostModel(exampleSet, modelInfo2, classPriors);
} else {
boolean success = this.trainAdditionalModel(extendedBatch, modelInfo2);
if (success) {
ensembleExtBatch = new BayBoostModel(exampleSet, modelInfo2, classPriors);
} else {
ensembleExtBatch = null;
estimateFavoursExtBatch = false;
}
}
if (holdOutRatio > 0) {
Iterator hoEit = holdOutExamples.iterator();
while (hoEit.hasNext()) {
((Example) hoEit.next()).setValue(streamControlAttribute, this.currentIteration);
}
// TODO: create new example set
// trainingSet.updateCondition();
if (ensembleExtBatch != null) {
trainingSet = (ConditionedExampleSet) ensembleNewBatch.apply(trainingSet);
hoEit = holdOutExamples.iterator();
int errors = 0;
while (hoEit.hasNext()) {
Example example = (Example) hoEit.next();
if (example.getPredictedLabel() != example.getLabel()) errors++;
}
double newBatchErr = ((double) errors) / holdOutExamples.size();
trainingSet = (ConditionedExampleSet) ensembleExtBatch.apply(trainingSet);
hoEit = holdOutExamples.iterator();
errors = 0;
while (hoEit.hasNext()) {
Example example = (Example) hoEit.next();
if (example.getPredictedLabel() != example.getLabel()) errors++;
}
double extBatchErr = ((double) errors) / holdOutExamples.size();
estimateFavoursExtBatch = (extBatchErr <= newBatchErr);
if (estimateFavoursExtBatch) {
ensembleExtBatch =
this.retrainLastWeight(ensembleExtBatch, trainingSet, holdOutExamples);
} else
ensembleNewBatch =
this.retrainLastWeight(ensembleNewBatch, trainingSet, holdOutExamples);
} else
ensembleNewBatch =
this.retrainLastWeight(ensembleNewBatch, trainingSet, holdOutExamples);
}
} else {
this.trainAdditionalModel(trainingSet, modelInfo);
ensembleNewBatch = new BayBoostModel(exampleSet, modelInfo, classPriors);
ensembleExtBatch = null;
estimateFavoursExtBatch = false;
}
}
this.restoreOldWeights(exampleSet);
return (ensembleExtBatch == null ? ensembleNewBatch : ensembleExtBatch);
}
