/** This is (mostly) copied from CRF4.java */
public boolean[][] labelConnectionsIn(
Alphabet outputAlphabet, InstanceList trainingSet, String start) {
int numLabels = outputAlphabet.size();
boolean[][] connections = new boolean[numLabels][numLabels];
for (int i = 0; i < trainingSet.size(); i++) {
Instance instance = trainingSet.getInstance(i);
FeatureSequence output = (FeatureSequence) instance.getTarget();
for (int j = 1; j < output.size(); j++) {
int sourceIndex = outputAlphabet.lookupIndex(output.get(j - 1));
int destIndex = outputAlphabet.lookupIndex(output.get(j));
assert (sourceIndex >= 0 && destIndex >= 0);
connections[sourceIndex][destIndex] = true;
}
}
// Handle start state
if (start != null) {
int startIndex = outputAlphabet.lookupIndex(start);
for (int j = 0; j < outputAlphabet.size(); j++) {
connections[startIndex][j] = true;
}
}
return connections;
}
public static void main(String[] args) throws bsh.EvalError, java.io.IOException {
// Process the command-line options
commandOptions.process(args);
System.out.println("Trainer = " + trainerConstructorOption.value.toString());
ClassifierTrainer trainer = (ClassifierTrainer) trainerConstructorOption.value;
InstanceList ilist = InstanceList.load(new File(instanceListFilenameOption.value));
Random r = randomSeedOption.wasInvoked() ? new Random(randomSeedOption.value) : new Random();
double t = trainingProportionOption.value;
double v = validationProportionOption.value;
InstanceList[] ilists = ilist.split(r, new double[] {t, v, 1 - t - v});
System.err.println("Training...");
Classifier c =
trainer.train(
ilists[0],
ilists[1],
null,
(ClassifierEvaluating) classifierEvaluatorOption.value,
null);
if (printTrainAccuracyOption.value)
System.out.print("Train accuracy = " + c.getAccuracy(ilists[0]) + " ");
if (printTestAccuracyOption.value)
System.out.print("Test accuracy = " + c.getAccuracy(ilists[2]));
if (printTrainAccuracyOption.value || printTestAccuracyOption.value) System.out.println("");
if (outputFilenameOption.wasInvoked()) {
try {
ObjectOutputStream oos =
new ObjectOutputStream(new FileOutputStream(instanceListFilenameOption.value));
oos.writeObject(c);
oos.close();
} catch (Exception e) {
e.printStackTrace();
throw new IllegalArgumentException(
"Couldn't write classifier to filename " + instanceListFilenameOption.value);
}
}
}
public static CRF4 createCRF(File trainingFile, CRFInfo crfInfo) throws FileNotFoundException {
Reader trainingFileReader = new FileReader(trainingFile);
// Create a pipe that we can use to convert the training
// file to a feature vector sequence.
Pipe p = new SimpleTagger.SimpleTaggerSentence2FeatureVectorSequence();
// The training file does contain tags (aka targets)
p.setTargetProcessing(true);
// Register the default tag with the pipe, by looking it up
// in the targetAlphabet before we look up any other tag.
p.getTargetAlphabet().lookupIndex(crfInfo.defaultLabel);
// Create a new instancelist to hold the training data.
InstanceList trainingData = new InstanceList(p);
// Read in the training data.
trainingData.add(new LineGroupIterator(trainingFileReader, Pattern.compile("^\\s*$"), true));
// Create the CRF model.
CRF4 crf = new CRF4(p, null);
// Set various config options
crf.setGaussianPriorVariance(crfInfo.gaussianVariance);
crf.setTransductionType(crfInfo.transductionType);
// Set up the model's states.
if (crfInfo.stateInfoList != null) {
Iterator stateIter = crfInfo.stateInfoList.iterator();
while (stateIter.hasNext()) {
CRFInfo.StateInfo state = (CRFInfo.StateInfo) stateIter.next();
crf.addState(
state.name,
state.initialCost,
state.finalCost,
state.destinationNames,
state.labelNames,
state.weightNames);
}
} else if (crfInfo.stateStructure == CRFInfo.FULLY_CONNECTED_STRUCTURE)
crf.addStatesForLabelsConnectedAsIn(trainingData);
else if (crfInfo.stateStructure == CRFInfo.HALF_CONNECTED_STRUCTURE)
crf.addStatesForHalfLabelsConnectedAsIn(trainingData);
else if (crfInfo.stateStructure == CRFInfo.THREE_QUARTERS_CONNECTED_STRUCTURE)
crf.addStatesForThreeQuarterLabelsConnectedAsIn(trainingData);
else if (crfInfo.stateStructure == CRFInfo.BILABELS_STRUCTURE)
crf.addStatesForBiLabelsConnectedAsIn(trainingData);
else throw new RuntimeException("Unexpected state structure " + crfInfo.stateStructure);
// Set up the weight groups.
if (crfInfo.weightGroupInfoList != null) {
Iterator wgIter = crfInfo.weightGroupInfoList.iterator();
while (wgIter.hasNext()) {
CRFInfo.WeightGroupInfo wg = (CRFInfo.WeightGroupInfo) wgIter.next();
FeatureSelection fs =
FeatureSelection.createFromRegex(
crf.getInputAlphabet(), Pattern.compile(wg.featureSelectionRegex));
crf.setFeatureSelection(crf.getWeightsIndex(wg.name), fs);
}
}
// Train the CRF.
crf.train(trainingData, null, null, null, crfInfo.maxIterations);
return crf;
}
public void test(
Transducer transducer,
InstanceList data,
String description,
PrintStream viterbiOutputStream) {
int[] ntrue = new int[segmentTags.length];
int[] npred = new int[segmentTags.length];
int[] ncorr = new int[segmentTags.length];
LabelAlphabet dict = (LabelAlphabet) transducer.getInputPipe().getTargetAlphabet();
for (int i = 0; i < data.size(); i++) {
Instance instance = data.getInstance(i);
Sequence input = (Sequence) instance.getData();
Sequence trueOutput = (Sequence) instance.getTarget();
assert (input.size() == trueOutput.size());
Sequence predOutput = transducer.viterbiPath(input).output();
assert (predOutput.size() == trueOutput.size());
List trueSegs = new ArrayList();
List predSegs = new ArrayList();
addSegs(trueSegs, trueOutput);
addSegs(predSegs, predOutput);
// System.out.println("FieldF1Evaluator instance "+instance.getName ());
// printSegs(dict, trueSegs, "True");
// printSegs(dict, predSegs, "Pred");
for (Iterator it = predSegs.iterator(); it.hasNext(); ) {
Segment seg = (Segment) it.next();
npred[seg.tag]++;
if (trueSegs.contains(seg)) {
ncorr[seg.tag]++;
}
}
for (Iterator it = trueSegs.iterator(); it.hasNext(); ) {
Segment seg = (Segment) it.next();
ntrue[seg.tag]++;
}
}
DecimalFormat f = new DecimalFormat("0.####");
logger.info(description + " per-field F1");
for (int tag = 0; tag < segmentTags.length; tag++) {
double precision = ((double) ncorr[tag]) / npred[tag];
double recall = ((double) ncorr[tag]) / ntrue[tag];
double f1 = (2 * precision * recall) / (precision + recall);
Label name = dict.lookupLabel(segmentTags[tag]);
logger.info(
" segments "
+ name
+ " true = "
+ ntrue[tag]
+ " pred = "
+ npred[tag]
+ " correct = "
+ ncorr[tag]);
logger.info(
" precision="
+ f.format(precision)
+ " recall="
+ f.format(recall)
+ " f1="
+ f.format(f1));
}
}
