@Override
protected ComputeResult doCompute(List<Data> dataset, Parameter parameter) {
AnnModel annModel = (AnnModel) ((Object[]) parameter.getSerializable())[0];
int accurate = 0;
double totalLoss = 0;
double activationFunctionCenterValue = 0.5;
if (annModel.getConfiguration().activationFunctionOfLayer(annModel.getLayerCount() - 1)
== ActivationFunction.TANH) {
activationFunctionCenterValue = 0;
}
AnnTrainer trainer = new AnnTrainer();
double[][] sum = new double[annModel.getLayerCount()][];
for (Data data : dataset) {
double loss = 0;
boolean error = false;
AnnData annData = (AnnData) data.getSerializable();
double[][] output = trainer.run(annModel, annData.getInput(), sum);
for (int i = 0; i < output[output.length - 1].length; i++) {
double expected = annData.getOutput()[i];
double actual = output[output.length - 1][i];
loss += Math.abs(expected - actual);
if ((expected - activationFunctionCenterValue) * (actual - activationFunctionCenterValue)
< 0) {
error = true;
break;
}
}
if (!error) {
accurate++;
}
totalLoss += loss / output[output.length - 1].length;
}
LogUtility.logAnnModel(getLogger(), annModel);
System.out.println("Loss = " + (totalLoss / dataset.size()));
String message = "Classification precision: " + 100.0 * accurate / dataset.size() + "%";
getLogger().info(message);
System.out.println(message);
// LogUtility.logAnnModel(getLogger(), annModel);
// return computation result
ComputeResult result = new ComputeResult();
result.setRepeat(true);
result.setRepeatDelayInSeconds(EVALUATION_INTERVAL_IN_SECONDS);
result.setGradient(null);
result.setAudit(false);
return result;
}
@Override
public List<Data> produceData(int size) throws DataProducerException {
// add negative instances
List<Data> dataset = Lists.newArrayList();
while (dataset.size() < size) {
for (Data data : super.produceData(size / 2)) {
WordEmbeddingTrainingInstance positiveInstance =
(WordEmbeddingTrainingInstance) data.getSerializable();
// validate positive instance
boolean invalidWindow = false;
for (int index : positiveInstance.getInput()) {
invalidWindow = isWordInvalid(index);
if (invalidWindow) {
break;
}
}
if (invalidWindow) {
continue;
}
positiveInstance.setOutput(1);
dataset.add(data);
WordEmbeddingTrainingInstance negativeInstance = new WordEmbeddingTrainingInstance();
negativeInstance.setInput(Lists.newArrayList(positiveInstance.getInput()));
negativeInstance.setOutput(-1);
// generate negative word
int negativeWord =
this.dictionary.sampleWordUniformlyAboveFrequenceRank(this.frequencyRankBound);
while (negativeWord
== negativeInstance.getInput().get(negativeInstance.getInput().size() / 2)) {
negativeWord =
this.dictionary.sampleWordUniformlyAboveFrequenceRank(this.frequencyRankBound);
}
// set negative word
negativeInstance.getInput().set(negativeInstance.getInput().size() / 2, negativeWord);
dataset.add(new Data(negativeInstance));
}
}
Collections.shuffle(dataset);
return dataset;
}
