@Override
public void trainC(ClassificationDataSet dataSet, ExecutorService threadPool) {
final int models = baseClassifiers.size();
final int C = dataSet.getClassSize();
weightsPerModel = C == 2 ? 1 : C;
ClassificationDataSet metaSet =
new ClassificationDataSet(
weightsPerModel * models, new CategoricalData[0], dataSet.getPredicting());
List<ClassificationDataSet> dataFolds = dataSet.cvSet(folds);
// iterate in the order of the folds so we get the right dataum weights
for (ClassificationDataSet cds : dataFolds)
for (int i = 0; i < cds.getSampleSize(); i++)
metaSet.addDataPoint(
new DenseVector(weightsPerModel * models),
cds.getDataPointCategory(i),
cds.getDataPoint(i).getWeight());
// create the meta training set
for (int c = 0; c < baseClassifiers.size(); c++) {
Classifier cl = baseClassifiers.get(c);
int pos = 0;
for (int f = 0; f < dataFolds.size(); f++) {
ClassificationDataSet train = ClassificationDataSet.comineAllBut(dataFolds, f);
ClassificationDataSet test = dataFolds.get(f);
if (threadPool == null) cl.trainC(train);
else cl.trainC(train, threadPool);
for (int i = 0;
i < test.getSampleSize();
i++) // evaluate and mark each point in the held out fold.
{
CategoricalResults pred = cl.classify(test.getDataPoint(i));
if (C == 2)
metaSet.getDataPoint(pos).getNumericalValues().set(c, pred.getProb(0) * 2 - 1);
else {
Vec toSet = metaSet.getDataPoint(pos).getNumericalValues();
for (int j = weightsPerModel * c; j < weightsPerModel * (c + 1); j++)
toSet.set(j, pred.getProb(j - weightsPerModel * c));
}
pos++;
}
}
}
// train the meta model
if (threadPool == null) aggregatingClassifier.trainC(metaSet);
else aggregatingClassifier.trainC(metaSet, threadPool);
// train the final classifiers, unless folds=1. In that case they are already trained
if (folds != 1) {
for (Classifier cl : baseClassifiers)
if (threadPool == null) cl.trainC(dataSet);
else cl.trainC(dataSet, threadPool);
}
}
/**
* Copy constructor
*
* @param toCopy the object to copy
*/
public Stacking(Stacking toCopy) {
this.folds = toCopy.folds;
this.weightsPerModel = toCopy.weightsPerModel;
if (toCopy.aggregatingClassifier != null) {
this.aggregatingClassifier = toCopy.aggregatingClassifier.clone();
this.baseClassifiers = new ArrayList<Classifier>(toCopy.baseClassifiers.size());
for (Classifier bc : toCopy.baseClassifiers) this.baseClassifiers.add(bc.clone());
if (toCopy.aggregatingRegressor == toCopy.aggregatingClassifier) // supports both
{
aggregatingRegressor = (Regressor) aggregatingClassifier;
baseRegressors = (List) baseClassifiers; // ugly type easure exploitation...
}
} else // we are doing with regressors only
{
this.aggregatingRegressor = toCopy.aggregatingRegressor.clone();
this.baseRegressors = new ArrayList<Regressor>(toCopy.baseRegressors.size());
for (Regressor br : toCopy.baseRegressors) this.baseRegressors.add(br.clone());
}
}
@Override
public CategoricalResults classify(DataPoint data) {
Vec w = new DenseVector(weightsPerModel * baseClassifiers.size());
if (weightsPerModel == 1)
for (int i = 0; i < baseClassifiers.size(); i++)
w.set(i, baseClassifiers.get(i).classify(data).getProb(0) * 2 - 1);
else {
for (int i = 0; i < baseClassifiers.size(); i++) {
CategoricalResults pred = baseClassifiers.get(i).classify(data);
for (int j = 0; j < weightsPerModel; j++) w.set(i * weightsPerModel + j, pred.getProb(j));
}
}
return aggregatingClassifier.classify(new DataPoint(w));
}
@Override
public boolean supportsWeightedData() {
if (aggregatingClassifier != null) return aggregatingClassifier.supportsWeightedData();
else return aggregatingRegressor.supportsWeightedData();
}
@Override
public void trainC(final ClassificationDataSet dataSet, final ExecutorService threadPool) {
final PriorityQueue<ClassificationModelEvaluation> bestModels =
new PriorityQueue<ClassificationModelEvaluation>(
folds,
new Comparator<ClassificationModelEvaluation>() {
@Override
public int compare(
ClassificationModelEvaluation t, ClassificationModelEvaluation t1) {
double v0 = t.getScoreStats(classificationTargetScore).getMean();
double v1 = t1.getScoreStats(classificationTargetScore).getMean();
int order = classificationTargetScore.lowerIsBetter() ? 1 : -1;
return order * Double.compare(v0, v1);
}
});
/**
* Use this to keep track of which parameter we are altering. Index correspondence to the
* parameter, and its value corresponds to which value has been used. Increment and carry counts
* to iterate over all possible combinations.
*/
int[] setTo = new int[searchParams.size()];
/**
* Each model is set to have different combination of parameters. We then train each model to
* determine the best one.
*/
final List<Classifier> paramsToEval = new ArrayList<Classifier>();
while (true) {
setParameters(setTo);
paramsToEval.add(baseClassifier.clone());
if (incrementCombination(setTo)) break;
}
/*
* This is the Executor used for training the models in parallel. If we
* are not supposed to do that, it will be an executor that executes
* them sequentually.
*/
final ExecutorService modelService;
if (trainModelsInParallel) modelService = threadPool;
else modelService = new FakeExecutor();
final CountDownLatch latch; // used for stopping in both cases
// if we are doing our CV splits ahead of time, get them done now
final List<ClassificationDataSet> preFolded;
/** Pre-combine our training combinations so that any caching can be re-used */
final List<ClassificationDataSet> trainCombinations;
if (reuseSameCVFolds) {
preFolded = dataSet.cvSet(folds);
trainCombinations = new ArrayList<ClassificationDataSet>(preFolded.size());
for (int i = 0; i < preFolded.size(); i++)
trainCombinations.add(ClassificationDataSet.comineAllBut(preFolded, i));
} else {
preFolded = null;
trainCombinations = null;
}
boolean considerWarm = useWarmStarts && baseClassifier instanceof WarmClassifier;
/**
* make sure we don't do a warm start if its only supported when trained on the same data but we
* aren't reuse-ing the same CV splits So we get the truth table
*
* <p>a | b | (a&&b)||¬a T | T | T T | F | F F | T | T F | F | T
*
* <p>where a = warmFromSameDataOnly and b = reuseSameSplit So we can instead use ¬ a || b
*/
if (considerWarm
&& (!((WarmClassifier) baseClassifier).warmFromSameDataOnly() || reuseSameCVFolds)) {
/* we want all of the first parameter (which is the warm paramter,
* taken care of for us) values done in a group. So We can get this
* by just dividing up the larger list into sub lists, each sub list
* is adjacent in the original and is the number of parameter values
* we wanted to try
*/
int stepSize = searchValues.get(0).size();
int totalJobs = paramsToEval.size() / stepSize;
latch = new CountDownLatch(totalJobs);
for (int startPos = 0; startPos < paramsToEval.size(); startPos += stepSize) {
final List<Classifier> subSet = paramsToEval.subList(startPos, startPos + stepSize);
modelService.submit(
new Runnable() {
@Override
public void run() {
Classifier[] prevModels = null;
for (Classifier c : subSet) {
ClassificationModelEvaluation cme =
trainModelsInParallel
? new ClassificationModelEvaluation(c, dataSet)
: new ClassificationModelEvaluation(c, dataSet, threadPool);
cme.setKeepModels(true); // we need these to do warm starts!
cme.setWarmModels(prevModels);
cme.addScorer(classificationTargetScore.clone());
if (reuseSameCVFolds) cme.evaluateCrossValidation(preFolded, trainCombinations);
else cme.evaluateCrossValidation(folds);
prevModels = cme.getKeptModels();
synchronized (bestModels) {
bestModels.add(cme);
}
}
latch.countDown();
}
});
}
} else // regular CV, train a new model from scratch at every step
{
latch = new CountDownLatch(paramsToEval.size());
for (final Classifier toTrain : paramsToEval) {
modelService.submit(
new Runnable() {
@Override
public void run() {
ClassificationModelEvaluation cme =
trainModelsInParallel
? new ClassificationModelEvaluation(toTrain, dataSet)
: new ClassificationModelEvaluation(toTrain, dataSet, threadPool);
cme.addScorer(classificationTargetScore.clone());
if (reuseSameCVFolds) cme.evaluateCrossValidation(preFolded, trainCombinations);
else cme.evaluateCrossValidation(folds);
synchronized (bestModels) {
bestModels.add(cme);
}
latch.countDown();
}
});
}
}
// now wait for everyone to finish
try {
latch.await();
// Now we know the best classifier, we need to train one on the whole data set.
Classifier bestClassifier =
bestModels.peek().getClassifier(); // Just re-train it on the whole set
if (trainFinalModel) {
// try and warm start the final model if we can
if (useWarmStarts
&& bestClassifier instanceof WarmClassifier
&& !((WarmClassifier) bestClassifier)
.warmFromSameDataOnly()) // last line here needed to make sure we can do this warm
// train
{
WarmClassifier wc = (WarmClassifier) bestClassifier;
if (threadPool instanceof FakeExecutor) wc.trainC(dataSet, wc.clone());
else wc.trainC(dataSet, wc.clone(), threadPool);
} else {
if (threadPool instanceof FakeExecutor) bestClassifier.trainC(dataSet);
else bestClassifier.trainC(dataSet, threadPool);
}
}
trainedClassifier = bestClassifier;
} catch (InterruptedException ex) {
Logger.getLogger(GridSearch.class.getName()).log(Level.SEVERE, null, ex);
}
}