/**
* Saves the feature divide model settings .fsm file.
*
* @throws MaltChainedException
*/
protected void save() throws MaltChainedException {
try {
final BufferedWriter out =
new BufferedWriter(
getGuide()
.getConfiguration()
.getConfigurationDir()
.getOutputStreamWriter(getModelName() + ".dsm"));
out.write(masterModel.getIndex() + "\t" + masterModel.getFrequency() + "\n");
if (divideModels != null) {
for (AtomicModel divideModel : divideModels.values()) {
out.write(divideModel.getIndex() + "\t" + divideModel.getFrequency() + "\n");
}
}
out.close();
} catch (IOException e) {
throw new GuideException(
"Could not write to the guide model settings file '"
+ getModelName()
+ ".dsm"
+ "', when "
+ "saving the guide model settings to file. ",
e);
}
}
public boolean predict(SingleDecision decision) throws MaltChainedException {
if (getGuide().getGuideMode() == ClassifierGuide.GuideMode.BATCH) {
throw new GuideException("Can only predict during parsing. ");
} else if (!(divideFeature.getFeatureValue() instanceof SingleFeatureValue)) {
throw new GuideException("The divide feature does not have a single value. ");
}
// divideFeature.update();
if (divideModels != null
&& divideModels.containsKey(
((SingleFeatureValue) divideFeature.getFeatureValue()).getCode())) {
return divideModels
.get(((SingleFeatureValue) divideFeature.getFeatureValue()).getCode())
.predict(decision);
} else if (masterModel != null && masterModel.getFrequency() > 0) {
return masterModel.predict(decision);
} else {
getGuide()
.getConfiguration()
.getConfigLogger()
.info(
"Could not predict the next parser decision because there is "
+ "no divide or master model that covers the divide value '"
+ ((SingleFeatureValue) divideFeature.getFeatureValue()).getCode()
+ "', as default"
+ " class code '1' is used. ");
decision.addDecision(1); // default prediction
// classCodeTable.getEmptyKBestList().addKBestItem(1);
}
return true;
}
public void terminate() throws MaltChainedException {
if (divideModels != null) {
for (AtomicModel divideModel : divideModels.values()) {
divideModel.terminate();
}
}
if (masterModel != null) {
masterModel.terminate();
}
}
public void finalizeSentence(DependencyStructure dependencyGraph) throws MaltChainedException {
// if (getGuide().getGuideMode() == Guide.GuideMode.CLASSIFY) {
// throw new GuideException("Can only finish sentence during learning. ");
// }
if (divideModels != null) {
for (AtomicModel divideModel : divideModels.values()) {
divideModel.finalizeSentence(dependencyGraph);
}
} else {
throw new GuideException("The feature divide models cannot be found. ");
}
}
public void noMoreInstances() throws MaltChainedException {
// if (getGuide().getGuideMode() == Guide.GuideMode.CLASSIFY) {
// throw new GuideException("Can only finish all data during learning. ");
// }
if (divideModels != null) {
divideFeature.updateCardinality();
for (Integer index : divideModels.keySet()) {
divideModels.get(index).noMoreInstances();
}
final TreeSet<Integer> removeSet = new TreeSet<Integer>();
for (Integer index : divideModels.keySet()) {
if (divideModels.get(index).getFrequency() <= divideThreshold) {
divideModels
.get(index)
.moveAllInstances(masterModel, divideFeature, divideFeatureIndexVector);
removeSet.add(index);
}
}
for (Integer index : removeSet) {
divideModels.remove(index);
}
masterModel.noMoreInstances();
} else {
throw new GuideException("The feature divide models cannot be found. ");
}
}
private AtomicModel getAtomicModel() throws MaltChainedException {
if (getGuide().getGuideMode() == ClassifierGuide.GuideMode.BATCH) {
throw new GuideException("Can only predict during parsing. ");
} else if (!(divideFeature.getFeatureValue() instanceof SingleFeatureValue)) {
throw new GuideException("The divide feature does not have a single value. ");
}
if (divideModels != null
&& divideModels.containsKey(
((SingleFeatureValue) divideFeature.getFeatureValue()).getCode())) {
return divideModels.get(((SingleFeatureValue) divideFeature.getFeatureValue()).getCode());
} else if (masterModel != null && masterModel.getFrequency() > 0) {
return masterModel;
} else {
getGuide()
.getConfiguration()
.getConfigLogger()
.info(
"Could not predict the next parser decision because there is "
+ "no divide or master model that covers the divide value '"
+ ((SingleFeatureValue) divideFeature.getFeatureValue()).getCode()
+ "', as default"
+ " class code '1' is used. ");
}
return null;
}
public void train() throws MaltChainedException {
for (AtomicModel divideModel : divideModels.values()) {
divideModel.train();
}
masterModel.train();
save();
for (AtomicModel divideModel : divideModels.values()) {
divideModel.terminate();
}
masterModel.terminate();
}
/**
* Moves all instance from this atomic model into the destination atomic model and add the divide
* feature. This method is used by the feature divide model to sum up all model below a certain
* threshold.
*
* @param model the destination atomic model
* @param divideFeature the divide feature
* @param divideFeatureIndexVector the divide feature index vector
* @throws MaltChainedException
*/
public void moveAllInstances(
AtomicModel model, FeatureFunction divideFeature, ArrayList<Integer> divideFeatureIndexVector)
throws MaltChainedException {
if (method == null) {
throw new GuideException("The learner cannot be found. ");
} else if (model == null) {
throw new GuideException("The guide model cannot be found. ");
} else if (divideFeature == null) {
throw new GuideException("The divide feature cannot be found. ");
} else if (divideFeatureIndexVector == null) {
throw new GuideException("The divide feature index vector cannot be found. ");
}
((Modifiable) divideFeature).setFeatureValue(index);
method.moveAllInstances(model.getMethod(), divideFeature, divideFeatureIndexVector);
method.terminate();
method = null;
}
/**
* Loads the feature divide model settings .fsm file.
*
* @throws MaltChainedException
*/
protected void load() throws MaltChainedException {
try {
final BufferedReader in =
new BufferedReader(
getGuide()
.getConfiguration()
.getConfigurationDir()
.getInputStreamReader(getModelName() + ".dsm"));
final Pattern tabPattern = Pattern.compile("\t");
while (true) {
String line = in.readLine();
if (line == null) break;
String[] cols = tabPattern.split(line);
if (cols.length != 2) {
throw new GuideException("");
}
int code = -1;
int freq = 0;
try {
code = Integer.parseInt(cols[0]);
freq = Integer.parseInt(cols[1]);
} catch (NumberFormatException e) {
throw new GuideException(
"Could not convert a string value into an integer value when loading the feature divide model settings (.fsm). ",
e);
}
if (code == -1) {
masterModel = new AtomicModel(-1, masterFeatureVector, this);
masterModel.setFrequency(freq);
} else if (divideModels != null) {
divideModels.put(code, new AtomicModel(code, divideFeatureVector, this));
divideModels.get(code).setFrequency(freq);
}
setFrequency(getFrequency() + freq);
}
in.close();
} catch (IOException e) {
throw new GuideException(
"Could not read from the guide model settings file '"
+ getModelName()
+ ".dsm"
+ "', when "
+ "loading the guide model settings. ",
e);
}
}