private static <T> Pair<List<T>, List<T>> splitData(List<T> original, double splitForSecond) {
List<T> first = new ArrayList<>();
List<T> second = new ArrayList<>();
if (splitForSecond > 0.0) {
Collections.shuffle(original, new Random(0L));
int numFirst = (int) ((1.0 - splitForSecond) * original.size());
first.addAll(original.subList(0, numFirst));
second.addAll(original.subList(numFirst, original.size()));
} else {
first.addAll(original);
// second stays empty
}
return Tuples.pair(first, second);
}
@SneakyThrows
public static void trainAndSaveModel(Opts opts) {
// Load labeled data
List<String> templateLines = linesFromPath(opts.templateFile).collect(toList());
val predExtractor = ConllFormat.predicatesFromTemplate(templateLines.stream());
List<List<Pair<ConllFormat.Row, String>>> labeledData =
ConllFormat.readData(linesFromPath(opts.trainPath), true)
.stream()
.map(x -> x.stream().map(y -> y.asLabeledPair().swap()).collect(Collectors.toList()))
.collect(Collectors.toList());
// Split train/test data
logger.info(
"CRF training with {} threads and {} labeled examples",
opts.numThreads,
labeledData.size());
val trainTestPair = splitData(labeledData, opts.testSplitRatio);
List<List<Pair<ConllFormat.Row, String>>> trainLabeledData = trainTestPair.getOne();
List<List<Pair<ConllFormat.Row, String>>> testLabeledData = trainTestPair.getTwo();
// Set up Train options
CRFTrainer.Opts trainOpts = new CRFTrainer.Opts();
trainOpts.sigmaSq = opts.sigmaSquared;
trainOpts.lbfgsHistorySize = opts.lbfgsHistorySize;
trainOpts.minExpectedFeatureCount = (int) (1.0 / opts.featureKeepProb);
trainOpts.numThreads = opts.numThreads;
// Trainer
CRFTrainer<String, ConllFormat.Row, String> trainer =
new CRFTrainer<>(trainLabeledData, predExtractor, trainOpts);
// Setup iteration callback, weird trick here where you require
// the trainer to make a model for each iteration but then need
// to modify the iteration-callback to use it
Parallel.MROpts evalMrOpts =
Parallel.MROpts.withIdAndThreads("mr-crf-train-eval", opts.numThreads);
trainOpts.optimizerOpts.iterCallback =
(weights) -> {
CRFModel<String, ConllFormat.Row, String> crfModel = trainer.modelForWeights(weights);
long start = System.currentTimeMillis();
List<List<Pair<String, ConllFormat.Row>>> trainEvalData =
trainLabeledData
.stream()
.map(x -> x.stream().map(Pair::swap).collect(toList()))
.collect(toList());
Evaluation<String> eval = Evaluation.compute(crfModel, trainEvalData, evalMrOpts);
long stop = System.currentTimeMillis();
logger.info(
"Train Accuracy: {} (took {} ms)", eval.tokenAccuracy.accuracy(), stop - start);
if (!testLabeledData.isEmpty()) {
start = System.currentTimeMillis();
List<List<Pair<String, ConllFormat.Row>>> testEvalData =
testLabeledData
.stream()
.map(x -> x.stream().map(Pair::swap).collect(toList()))
.collect(toList());
eval = Evaluation.compute(crfModel, testEvalData, evalMrOpts);
stop = System.currentTimeMillis();
logger.info(
"Test Accuracy: {} (took {} ms)", eval.tokenAccuracy.accuracy(), stop - start);
}
};
CRFModel<String, ConllFormat.Row, String> crfModel = trainer.train(trainLabeledData);
Parallel.shutdownExecutor(evalMrOpts.executorService, Long.MAX_VALUE);
Vector weights = crfModel.weights();
val dos = new DataOutputStream(new FileOutputStream(opts.modelPath));
logger.info("Writing model to {}", opts.modelPath);
ConllFormat.saveModel(dos, templateLines, crfModel.featureEncoder, weights);
}
