public double[][] trainWeights(
GeneralDataset<L, F> dataset,
double[] initial,
boolean bypassTuneSigma,
Minimizer<DiffFunction> minimizer) {
if (minimizer == null) minimizer = minimizerCreator.create();
if (dataset instanceof RVFDataset) ((RVFDataset<L, F>) dataset).ensureRealValues();
double[] interimWeights = null;
if (!bypassTuneSigma) {
if (tuneSigmaHeldOut) {
interimWeights =
heldOutSetSigma(
dataset); // the optimum interim weights from held-out training data have already
// been found.
} else if (tuneSigmaCV) {
crossValidateSetSigma(
dataset, folds); // TODO: assign optimum interim weights as part of this process.
}
}
LogConditionalObjectiveFunction<L, F> objective =
new LogConditionalObjectiveFunction<L, F>(dataset, logPrior);
if (initial == null && interimWeights != null && !retrainFromScratchAfterSigmaTuning) {
// System.err.println("## taking advantage of interim weights as starting point.");
initial = interimWeights;
}
if (initial == null) {
initial = objective.initial();
}
double[] weights = minimizer.minimize(objective, TOL, initial);
return objective.to2D(weights);
}
public Classifier<L, F> trainClassifier(
GeneralDataset<L, F> dataset, float[] dataWeights, LogPrior prior) {
Minimizer<DiffFunction> minimizer = getMinimizer();
if (dataset instanceof RVFDataset) ((RVFDataset<L, F>) dataset).ensureRealValues();
LogConditionalObjectiveFunction<L, F> objective =
new LogConditionalObjectiveFunction<L, F>(dataset, dataWeights, logPrior);
double[] initial = objective.initial();
double[] weights = minimizer.minimize(objective, TOL, initial);
LinearClassifier<L, F> classifier =
new LinearClassifier<L, F>(
objective.to2D(weights), dataset.featureIndex(), dataset.labelIndex());
return classifier;
}
public double[][] trainWeightsSemiSup(
GeneralDataset<L, F> data,
GeneralDataset<L, F> biasedData,
double[][] confusionMatrix,
double[] initial) {
Minimizer<DiffFunction> minimizer = minimizerCreator.create();
LogConditionalObjectiveFunction<L, F> objective =
new LogConditionalObjectiveFunction<L, F>(data, new LogPrior(LogPrior.LogPriorType.NULL));
BiasedLogConditionalObjectiveFunction biasedObjective =
new BiasedLogConditionalObjectiveFunction(
biasedData, confusionMatrix, new LogPrior(LogPrior.LogPriorType.NULL));
SemiSupervisedLogConditionalObjectiveFunction semiSupObjective =
new SemiSupervisedLogConditionalObjectiveFunction(objective, biasedObjective, logPrior);
if (initial == null) {
initial = objective.initial();
}
double[] weights = minimizer.minimize(semiSupObjective, TOL, initial);
return objective.to2D(weights);
}
/**
* Trains the linear classifier using Generalized Expectation criteria as described in
* <tt>Generalized Expectation Criteria for Semi Supervised Learning of Conditional Random
* Fields</tt>, Mann and McCallum, ACL 2008. The original algorithm is proposed for CRFs but has
* been adopted to LinearClassifier (which is a simpler special case of a CRF). IMPORTANT: the
* labeled features that are passed as an argument are assumed to be binary valued, although other
* features are allowed to be real valued.
*/
public LinearClassifier<L, F> trainSemiSupGE(
GeneralDataset<L, F> labeledDataset,
List<? extends Datum<L, F>> unlabeledDataList,
List<F> GEFeatures,
double convexComboCoeff) {
Minimizer<DiffFunction> minimizer = minimizerCreator.create();
LogConditionalObjectiveFunction<L, F> objective =
new LogConditionalObjectiveFunction<L, F>(
labeledDataset, new LogPrior(LogPrior.LogPriorType.NULL));
GeneralizedExpectationObjectiveFunction<L, F> geObjective =
new GeneralizedExpectationObjectiveFunction<L, F>(
labeledDataset, unlabeledDataList, GEFeatures);
SemiSupervisedLogConditionalObjectiveFunction semiSupObjective =
new SemiSupervisedLogConditionalObjectiveFunction(
objective, geObjective, null, convexComboCoeff);
double[] initial = objective.initial();
double[] weights = minimizer.minimize(semiSupObjective, TOL, initial);
return new LinearClassifier<L, F>(
objective.to2D(weights), labeledDataset.featureIndex(), labeledDataset.labelIndex());
}
private NBWeights trainWeightsUCL(int[][] data, int[] labels, int numFeatures, int numClasses) {
int[] numValues = numberValues(data, numFeatures);
int[] sumValues = new int[numFeatures]; // how many feature-values are before this feature
for (int j = 1; j < numFeatures; j++) {
sumValues[j] = sumValues[j - 1] + numValues[j - 1];
}
int[][] newdata = new int[data.length][numFeatures + 1];
for (int i = 0; i < data.length; i++) {
newdata[i][0] = 0;
for (int j = 0; j < numFeatures; j++) {
newdata[i][j + 1] = sumValues[j] + data[i][j] + 1;
}
}
int totalFeatures = sumValues[numFeatures - 1] + numValues[numFeatures - 1] + 1;
System.err.println("total feats " + totalFeatures);
LogConditionalObjectiveFunction<L, F> objective =
new LogConditionalObjectiveFunction<L, F>(
totalFeatures, numClasses, newdata, labels, prior, sigma, 0.0);
Minimizer<DiffFunction> min = new QNMinimizer();
double[] argmin = min.minimize(objective, 1e-4, objective.initial());
double[][] wts = objective.to2D(argmin);
System.out.println("weights have dimension " + wts.length);
return new NBWeights(wts, numValues);
}
public Classifier<L, F> trainClassifier(Iterable<Datum<L, F>> dataIterable) {
Minimizer<DiffFunction> minimizer = getMinimizer();
Index<F> featureIndex = Generics.newIndex();
Index<L> labelIndex = Generics.newIndex();
for (Datum<L, F> d : dataIterable) {
labelIndex.add(d.label());
featureIndex.addAll(d.asFeatures()); // If there are duplicates, it doesn't add them again.
}
System.err.println(
String.format(
"Training linear classifier with %d features and %d labels",
featureIndex.size(), labelIndex.size()));
LogConditionalObjectiveFunction<L, F> objective =
new LogConditionalObjectiveFunction<L, F>(dataIterable, logPrior, featureIndex, labelIndex);
objective.setPrior(new LogPrior(LogPrior.LogPriorType.QUADRATIC));
double[] initial = objective.initial();
double[] weights = minimizer.minimize(objective, TOL, initial);
LinearClassifier<L, F> classifier =
new LinearClassifier<L, F>(objective.to2D(weights), featureIndex, labelIndex);
return classifier;
}