Repository: opennlp-sandbox Updated Branches: refs/heads/master 96c088b00 -> 6bfb15f07
OPENNLP-1009 - added initial RNN and StackedRNN impls from Yay lab, minor fixes Project: http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/repo Commit: http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/commit/a63ec16c Tree: http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/tree/a63ec16c Diff: http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/diff/a63ec16c Branch: refs/heads/master Commit: a63ec16ccd602d5e1c468d2b3370caaea3fb5f3f Parents: 96c088b Author: Tommaso Teofili <[email protected]> Authored: Mon May 8 14:59:33 2017 +0200 Committer: Tommaso Teofili <[email protected]> Committed: Mon May 8 14:59:33 2017 +0200 ---------------------------------------------------------------------- opennlp-dl/pom.xml | 136 ++++--- opennlp-dl/src/main/java/NameFinderDL.java | 232 ------------ .../main/java/NameSampleDataSetIterator.java | 225 ------------ .../java/opennlp/tools/dl/NameFinderDL.java | 232 ++++++++++++ .../tools/dl/NameSampleDataSetIterator.java | 225 ++++++++++++ .../src/main/java/opennlp/tools/dl/RNN.java | 366 +++++++++++++++++++ .../main/java/opennlp/tools/dl/StackedRNN.java | 323 ++++++++++++++++ .../src/test/java/opennlp/tools/dl/RNNTest.java | 110 ++++++ .../java/opennlp/tools/dl/StackedRNNTest.java | 109 ++++++ .../src/test/resources/text/sentences.txt | 52 +++ 10 files changed, 1498 insertions(+), 512 deletions(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/pom.xml ---------------------------------------------------------------------- diff --git a/opennlp-dl/pom.xml b/opennlp-dl/pom.xml index f8a6679..3d15d8f 100644 --- a/opennlp-dl/pom.xml +++ b/opennlp-dl/pom.xml @@ -1,56 +1,82 @@ -<?xml version="1.0" encoding="UTF-8"?> -<project xmlns="http://maven.apache.org/POM/4.0.0"; - xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"; +<!-- + Licensed to the Apache Software Foundation (ASF) under one + or more contributor license agreements. See the NOTICE file + distributed with this work for additional information + regarding copyright ownership. The ASF licenses this file + to you under the Apache License, Version 2.0 (the + "License"); you may not use this file except in compliance + with the License. You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, + software distributed under the License is distributed on an + "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + KIND, either express or implied. See the License for the + specific language governing permissions and limitations + under the License. +--> +<project xmlns="http://maven.apache.org/POM/4.0.0"; xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"; xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd";> - <modelVersion>4.0.0</modelVersion> - - <groupId>burn</groupId> - <artifactId>dl4jtest</artifactId> - <version>1.0-SNAPSHOT</version> - - <dependencies> - <dependency> - <groupId>org.apache.opennlp</groupId> - <artifactId>opennlp-tools</artifactId> - <version>1.7.2</version> - </dependency> - - <dependency> - <groupId>org.deeplearning4j</groupId> - <artifactId>deeplearning4j-core</artifactId> - <version>0.7.2</version> - </dependency> - - <dependency> - <groupId>org.nd4j</groupId> - <artifactId>nd4j-native-platform</artifactId> - <!-- artifactId>nd4j-cuda-8.0-platform</artifactId --> - <version>0.7.2</version> - </dependency> - - <dependency> - <groupId>org.deeplearning4j</groupId> - <artifactId>deeplearning4j-nlp</artifactId> - <version>0.7.2</version> - </dependency> - <dependency> - <groupId>org.slf4j</groupId> - <artifactId>slf4j-simple</artifactId> - <version>1.7.12</version> - </dependency> - </dependencies> - - <build> - <plugins> - <plugin> - <groupId>org.apache.maven.plugins</groupId> - <artifactId>maven-compiler-plugin</artifactId> - <version>3.5.1</version> - <configuration> - <source>1.8</source> - <target>1.8</target> - </configuration> - </plugin> - </plugins> - </build> -</project> + <modelVersion>4.0.0</modelVersion> + + <groupId>org.apache.opennlp</groupId> + <artifactId>opennlp-dl</artifactId> + <version>0.1-SNAPSHOT</version> + + <properties> + <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding> + <nd4j.version>0.7.2</nd4j.version> + </properties> + + <dependencies> + <dependency> + <groupId>org.apache.opennlp</groupId> + <artifactId>opennlp-tools</artifactId> + <version>1.7.2</version> + </dependency> + + <dependency> + <groupId>org.deeplearning4j</groupId> + <artifactId>deeplearning4j-core</artifactId> + <version>${nd4j.version}</version> + </dependency> + + + <dependency> + <groupId>org.deeplearning4j</groupId> + <artifactId>deeplearning4j-nlp</artifactId> + <version>${nd4j.version}</version> + </dependency> + <dependency> + <groupId>org.slf4j</groupId> + <artifactId>slf4j-simple</artifactId> + <version>1.7.12</version> + </dependency> + <dependency> + <groupId>junit</groupId> + <artifactId>junit</artifactId> + <version>4.11</version> + <scope>test</scope> + </dependency> + <dependency> + <groupId>org.nd4j</groupId> + <artifactId>nd4j-native-platform</artifactId> + <version>${nd4j.version}</version> + </dependency> + </dependencies> + <build> + <plugins> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-compiler-plugin</artifactId> + <version>2.0.2</version> + <configuration> + <source>1.8</source> + <target>1.8</target> + <encoding>UTF-8</encoding> + </configuration> + </plugin> + </plugins> + </build> +</project> \ No newline at end of file http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/main/java/NameFinderDL.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/main/java/NameFinderDL.java b/opennlp-dl/src/main/java/NameFinderDL.java deleted file mode 100644 index 1184a06..0000000 --- a/opennlp-dl/src/main/java/NameFinderDL.java +++ /dev/null @@ -1,232 +0,0 @@ -/* - * Licensed to the Apache Software Foundation (ASF) under one or more - * contributor license agreements. See the NOTICE file distributed with - * this work for additional information regarding copyright ownership. - * The ASF licenses this file to You under the Apache License, Version 2.0 - * (the "License"); you may not use this file except in compliance with - * the License. You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ - -import java.io.File; -import java.io.IOException; -import java.nio.charset.StandardCharsets; -import java.util.ArrayList; -import java.util.Arrays; -import java.util.List; -import java.util.Map; -import java.util.stream.Collectors; -import java.util.stream.IntStream; - -import org.deeplearning4j.models.embeddings.loader.WordVectorSerializer; -import org.deeplearning4j.models.embeddings.wordvectors.WordVectors; -import org.deeplearning4j.nn.api.OptimizationAlgorithm; -import org.deeplearning4j.nn.conf.MultiLayerConfiguration; -import org.deeplearning4j.nn.conf.NeuralNetConfiguration; -import org.deeplearning4j.nn.conf.Updater; -import org.deeplearning4j.nn.conf.layers.GravesLSTM; -import org.deeplearning4j.nn.conf.layers.RnnOutputLayer; -import org.deeplearning4j.nn.multilayer.MultiLayerNetwork; -import org.deeplearning4j.nn.weights.WeightInit; -import org.deeplearning4j.optimize.listeners.ScoreIterationListener; -import org.nd4j.linalg.activations.Activation; -import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.api.iterator.DataSetIterator; -import org.nd4j.linalg.factory.Nd4j; -import org.nd4j.linalg.indexing.INDArrayIndex; -import org.nd4j.linalg.indexing.NDArrayIndex; -import org.nd4j.linalg.lossfunctions.LossFunctions; - -import opennlp.tools.namefind.BioCodec; -import opennlp.tools.namefind.NameSample; -import opennlp.tools.namefind.NameSampleDataStream; -import opennlp.tools.namefind.TokenNameFinder; -import opennlp.tools.namefind.TokenNameFinderEvaluator; -import opennlp.tools.util.MarkableFileInputStreamFactory; -import opennlp.tools.util.ObjectStream; -import opennlp.tools.util.PlainTextByLineStream; -import opennlp.tools.util.Span; - -// https://github.com/deeplearning4j/dl4j-examples/blob/master/dl4j-examples/src/main/java/org/deeplearning4j/examples/recurrent/word2vecsentiment/Word2VecSentimentRNN.java -public class NameFinderDL implements TokenNameFinder { - - private final MultiLayerNetwork network; - private final WordVectors wordVectors; - private int windowSize; - private String[] labels; - - public NameFinderDL(MultiLayerNetwork network, WordVectors wordVectors, int windowSize, - String[] labels) { - this.network = network; - this.wordVectors = wordVectors; - this.windowSize = windowSize; - this.labels = labels; - } - - static List<INDArray> mapToFeatureMatrices(WordVectors wordVectors, String[] tokens, int windowSize) { - - List<INDArray> matrices = new ArrayList<>(); - - // TODO: Dont' hard code word vector dimension ... - - for (int i = 0; i < tokens.length; i++) { - INDArray features = Nd4j.create(1, 300, windowSize); - for (int vectorIndex = 0; vectorIndex < windowSize; vectorIndex++) { - int tokenIndex = i + vectorIndex - ((windowSize - 1) / 2); - if (tokenIndex >= 0 && tokenIndex < tokens.length) { - String token = tokens[tokenIndex]; - double[] wv = wordVectors.getWordVector(token); - if (wv != null) { - INDArray vector = wordVectors.getWordVectorMatrix(token); - features.put(new INDArrayIndex[]{NDArrayIndex.point(0), NDArrayIndex.all(), - NDArrayIndex.point(vectorIndex)}, vector); - } - } - } - matrices.add(features); - } - - return matrices; - } - - static List<INDArray> mapToLabelVectors(NameSample sample, int windowSize, String[] labelStrings) { - - Map<String, Integer> labelToIndex = IntStream.range(0, labelStrings.length).boxed() - .collect(Collectors.toMap(i -> labelStrings[i], i -> i)); - - List<INDArray> vectors = new ArrayList<INDArray>(); - - for (int i = 0; i < sample.getSentence().length; i++) { - // encode the outcome as one-hot-representation - String outcomes[] = - new BioCodec().encode(sample.getNames(), sample.getSentence().length); - - INDArray labels = Nd4j.create(1, labelStrings.length, windowSize); - labels.putScalar(new int[]{0, labelToIndex.get(outcomes[i]), windowSize - 1}, 1.0d); - vectors.add(labels); - } - - return vectors; - } - - private static int max(INDArray array) { - int best = 0; - for (int i = 0; i < array.size(0); i++) { - if (array.getDouble(i) > array.getDouble(best)) { - best = i; - } - } - return best; - } - - @Override - public Span[] find(String[] tokens) { - List<INDArray> featureMartrices = mapToFeatureMatrices(wordVectors, tokens, windowSize); - - String[] outcomes = new String[tokens.length]; - for (int i = 0; i < tokens.length; i++) { - INDArray predictionMatrix = network.output(featureMartrices.get(i), false); - INDArray outcomeVector = predictionMatrix.get(NDArrayIndex.point(0), NDArrayIndex.all(), - NDArrayIndex.point(windowSize - 1)); - - outcomes[i] = labels[max(outcomeVector)]; - } - - // Delete invalid spans ... - for (int i = 0; i < outcomes.length; i++) { - if (outcomes[i].endsWith("cont") && (i == 0 || "other".equals(outcomes[i - 1]))) { - outcomes[i] = "other"; - } - } - - return new BioCodec().decode(Arrays.asList(outcomes)); - } - - @Override - public void clearAdaptiveData() { - } - - public static MultiLayerNetwork train(WordVectors wordVectors, ObjectStream<NameSample> samples, - int epochs, int windowSize, String[] labels) throws IOException { - int vectorSize = 300; - int layerSize = 256; - - MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder() - .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1) - .updater(Updater.RMSPROP) - .regularization(true).l2(0.001) - .weightInit(WeightInit.XAVIER) - // .gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue).gradientNormalizationThreshold(1.0) - .learningRate(0.01) - .list() - .layer(0, new GravesLSTM.Builder().nIn(vectorSize).nOut(layerSize) - .activation(Activation.TANH).build()) - .layer(1, new RnnOutputLayer.Builder().activation(Activation.SOFTMAX) - .lossFunction(LossFunctions.LossFunction.MCXENT).nIn(layerSize).nOut(3).build()) - .pretrain(false).backprop(true).build(); - - MultiLayerNetwork net = new MultiLayerNetwork(conf); - net.init(); - net.setListeners(new ScoreIterationListener(5)); - - // TODO: Extract labels on the fly from the data - - DataSetIterator train = new NameSampleDataSetIterator(samples, wordVectors, windowSize, labels); - - System.out.println("Starting training"); - - for (int i = 0; i < epochs; i++) { - net.fit(train); - train.reset(); - System.out.println(String.format("Finished epoche %d", i)); - } - - return net; - } - - public static void main(String[] args) throws Exception { - if (args.length != 3) { - System.out.println("Usage: trainFile testFile gloveTxt"); - return; - } - - String[] labels = new String[] { - "default-start", "default-cont", "other" - }; - - System.out.print("Loading vectors ... "); - WordVectors wordVectors = WordVectorSerializer.loadTxtVectors( - new File(args[2])); - System.out.println("Done"); - - int windowSize = 5; - - MultiLayerNetwork net = train(wordVectors, new NameSampleDataStream(new PlainTextByLineStream( - new MarkableFileInputStreamFactory(new File(args[0])), StandardCharsets.UTF_8)), 1, windowSize, labels); - - ObjectStream<NameSample> evalStream = new NameSampleDataStream(new PlainTextByLineStream( - new MarkableFileInputStreamFactory( - new File(args[1])), StandardCharsets.UTF_8)); - - NameFinderDL nameFinder = new NameFinderDL(net, wordVectors, windowSize, labels); - - System.out.print("Evaluating ... "); - TokenNameFinderEvaluator nameFinderEvaluator = new TokenNameFinderEvaluator(nameFinder); - nameFinderEvaluator.evaluate(evalStream); - - System.out.println("Done"); - - System.out.println(); - System.out.println(); - System.out.println("Results"); - - System.out.println(nameFinderEvaluator.getFMeasure().toString()); - } -} http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/main/java/NameSampleDataSetIterator.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/main/java/NameSampleDataSetIterator.java b/opennlp-dl/src/main/java/NameSampleDataSetIterator.java deleted file mode 100644 index f416a1d..0000000 --- a/opennlp-dl/src/main/java/NameSampleDataSetIterator.java +++ /dev/null @@ -1,225 +0,0 @@ -/* - * Licensed to the Apache Software Foundation (ASF) under one or more - * contributor license agreements. See the NOTICE file distributed with - * this work for additional information regarding copyright ownership. - * The ASF licenses this file to You under the Apache License, Version 2.0 - * (the "License"); you may not use this file except in compliance with - * the License. You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ - -import java.io.IOException; -import java.util.ArrayList; -import java.util.Arrays; -import java.util.Collections; -import java.util.Iterator; -import java.util.List; -import java.util.NoSuchElementException; - -import org.deeplearning4j.models.embeddings.wordvectors.WordVectors; -import org.nd4j.linalg.api.ndarray.INDArray; -import org.nd4j.linalg.dataset.DataSet; -import org.nd4j.linalg.dataset.api.DataSetPreProcessor; -import org.nd4j.linalg.dataset.api.iterator.DataSetIterator; -import org.nd4j.linalg.factory.Nd4j; -import org.nd4j.linalg.indexing.INDArrayIndex; -import org.nd4j.linalg.indexing.NDArrayIndex; - -import opennlp.tools.namefind.NameSample; -import opennlp.tools.util.FilterObjectStream; -import opennlp.tools.util.ObjectStream; - -public class NameSampleDataSetIterator implements DataSetIterator { - - private static class NameSampleToDataSetStream extends FilterObjectStream<NameSample, DataSet> { - - private final WordVectors wordVectors; - private final String[] labels; - private int windowSize; - - private Iterator<DataSet> dataSets = Collections.emptyListIterator(); - - NameSampleToDataSetStream(ObjectStream<NameSample> samples, WordVectors wordVectors, int windowSize, String[] labels) { - super(samples); - this.wordVectors = wordVectors; - this.windowSize = windowSize; - this.labels = labels; - } - - private Iterator<DataSet> createDataSets(NameSample sample) { - List<INDArray> features = NameFinderDL.mapToFeatureMatrices(wordVectors, sample.getSentence(), - windowSize); - - List<INDArray> labels = NameFinderDL.mapToLabelVectors(sample, windowSize, this.labels); - - List<DataSet> dataSetList = new ArrayList<>(); - - for (int i = 0; i < features.size(); i++) { - dataSetList.add(new DataSet(features.get(i), labels.get(i))); - } - - return dataSetList.iterator(); - } - - @Override - public final DataSet read() throws IOException { - - if (dataSets.hasNext()) { - return dataSets.next(); - } - else { - NameSample sample; - while (!dataSets.hasNext() && (sample = samples.read()) != null) { - dataSets = createDataSets(sample); - } - - if (dataSets.hasNext()) { - return read(); - } - } - - return null; - } - } - - private final int windowSize; - private final String[] labels; - - private final int batchSize = 128; - private final int vectorSize = 300; - - private final int totalSamples; - - private int cursor = 0; - - private final ObjectStream<DataSet> samples; - - NameSampleDataSetIterator(ObjectStream<NameSample> samples, WordVectors wordVectors, int windowSize, - String labels[]) throws IOException { - this.windowSize = windowSize; - this.labels = labels; - - this.samples = new NameSampleToDataSetStream(samples, wordVectors, windowSize, labels); - - int total = 0; - - DataSet sample; - while ((sample = this.samples.read()) != null) { - total++; - } - - totalSamples = total; - - samples.reset(); - } - - public DataSet next(int num) { - if (cursor >= totalExamples()) throw new NoSuchElementException(); - - INDArray features = Nd4j.create(num, vectorSize, windowSize); - INDArray featuresMask = Nd4j.zeros(num, windowSize); - - INDArray labels = Nd4j.create(num, 3, windowSize); - INDArray labelsMask = Nd4j.zeros(num, windowSize); - - // iterate stream and copy to arrays - - for (int i = 0; i < num; i++) { - DataSet sample; - try { - sample = samples.read(); - } catch (IOException e) { - throw new RuntimeException(e); - } - - if (sample != null) { - INDArray feature = sample.getFeatureMatrix(); - features.put(new INDArrayIndex[] {NDArrayIndex.point(i)}, feature.get(NDArrayIndex.point(0))); - - feature.get(new INDArrayIndex[] {NDArrayIndex.point(0), NDArrayIndex.all(), - NDArrayIndex.point(0)}); - - for (int j = 0; j < windowSize; j++) { - featuresMask.putScalar(new int[] {i, j}, 1.0); - } - - INDArray label = sample.getLabels(); - labels.put(new INDArrayIndex[] {NDArrayIndex.point(i)}, label.get(NDArrayIndex.point(0))); - labelsMask.putScalar(new int[] {i, windowSize - 1}, 1.0); - } - - cursor++; - } - - return new DataSet(features, labels, featuresMask, labelsMask); - } - - public int totalExamples() { - return totalSamples; - } - - public int inputColumns() { - return vectorSize; - } - - public int totalOutcomes() { - return getLabels().size(); - } - - public boolean resetSupported() { - return true; - } - - public boolean asyncSupported() { - return false; - } - - public void reset() { - cursor = 0; - - try { - samples.reset(); - } catch (IOException e) { - throw new RuntimeException(e); - } - } - - public int batch() { - return batchSize; - } - - public int cursor() { - return cursor; - } - - public int numExamples() { - return totalExamples(); - } - - public void setPreProcessor(DataSetPreProcessor dataSetPreProcessor) { - throw new UnsupportedOperationException(); - } - - public DataSetPreProcessor getPreProcessor() { - throw new UnsupportedOperationException(); - } - - public List<String> getLabels() { - return Arrays.asList("start","cont", "other"); - } - - public boolean hasNext() { - return cursor < numExamples(); - } - - public DataSet next() { - return next(batchSize); - } -} http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/main/java/opennlp/tools/dl/NameFinderDL.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/main/java/opennlp/tools/dl/NameFinderDL.java b/opennlp-dl/src/main/java/opennlp/tools/dl/NameFinderDL.java new file mode 100644 index 0000000..7547196 --- /dev/null +++ b/opennlp-dl/src/main/java/opennlp/tools/dl/NameFinderDL.java @@ -0,0 +1,232 @@ +package opennlp.tools.dl;/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +import java.io.File; +import java.io.IOException; +import java.nio.charset.StandardCharsets; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.List; +import java.util.Map; +import java.util.stream.Collectors; +import java.util.stream.IntStream; + +import org.deeplearning4j.models.embeddings.loader.WordVectorSerializer; +import org.deeplearning4j.models.embeddings.wordvectors.WordVectors; +import org.deeplearning4j.nn.api.OptimizationAlgorithm; +import org.deeplearning4j.nn.conf.MultiLayerConfiguration; +import org.deeplearning4j.nn.conf.NeuralNetConfiguration; +import org.deeplearning4j.nn.conf.Updater; +import org.deeplearning4j.nn.conf.layers.GravesLSTM; +import org.deeplearning4j.nn.conf.layers.RnnOutputLayer; +import org.deeplearning4j.nn.multilayer.MultiLayerNetwork; +import org.deeplearning4j.nn.weights.WeightInit; +import org.deeplearning4j.optimize.listeners.ScoreIterationListener; +import org.nd4j.linalg.activations.Activation; +import org.nd4j.linalg.api.ndarray.INDArray; +import org.nd4j.linalg.dataset.api.iterator.DataSetIterator; +import org.nd4j.linalg.factory.Nd4j; +import org.nd4j.linalg.indexing.INDArrayIndex; +import org.nd4j.linalg.indexing.NDArrayIndex; +import org.nd4j.linalg.lossfunctions.LossFunctions; + +import opennlp.tools.namefind.BioCodec; +import opennlp.tools.namefind.NameSample; +import opennlp.tools.namefind.NameSampleDataStream; +import opennlp.tools.namefind.TokenNameFinder; +import opennlp.tools.namefind.TokenNameFinderEvaluator; +import opennlp.tools.util.MarkableFileInputStreamFactory; +import opennlp.tools.util.ObjectStream; +import opennlp.tools.util.PlainTextByLineStream; +import opennlp.tools.util.Span; + +// https://github.com/deeplearning4j/dl4j-examples/blob/master/dl4j-examples/src/main/java/org/deeplearning4j/examples/recurrent/word2vecsentiment/Word2VecSentimentRNN.java +public class NameFinderDL implements TokenNameFinder { + + private final MultiLayerNetwork network; + private final WordVectors wordVectors; + private int windowSize; + private String[] labels; + + public NameFinderDL(MultiLayerNetwork network, WordVectors wordVectors, int windowSize, + String[] labels) { + this.network = network; + this.wordVectors = wordVectors; + this.windowSize = windowSize; + this.labels = labels; + } + + static List<INDArray> mapToFeatureMatrices(WordVectors wordVectors, String[] tokens, int windowSize) { + + List<INDArray> matrices = new ArrayList<>(); + + // TODO: Dont' hard code word vector dimension ... + + for (int i = 0; i < tokens.length; i++) { + INDArray features = Nd4j.create(1, 300, windowSize); + for (int vectorIndex = 0; vectorIndex < windowSize; vectorIndex++) { + int tokenIndex = i + vectorIndex - ((windowSize - 1) / 2); + if (tokenIndex >= 0 && tokenIndex < tokens.length) { + String token = tokens[tokenIndex]; + double[] wv = wordVectors.getWordVector(token); + if (wv != null) { + INDArray vector = wordVectors.getWordVectorMatrix(token); + features.put(new INDArrayIndex[]{NDArrayIndex.point(0), NDArrayIndex.all(), + NDArrayIndex.point(vectorIndex)}, vector); + } + } + } + matrices.add(features); + } + + return matrices; + } + + static List<INDArray> mapToLabelVectors(NameSample sample, int windowSize, String[] labelStrings) { + + Map<String, Integer> labelToIndex = IntStream.range(0, labelStrings.length).boxed() + .collect(Collectors.toMap(i -> labelStrings[i], i -> i)); + + List<INDArray> vectors = new ArrayList<INDArray>(); + + for (int i = 0; i < sample.getSentence().length; i++) { + // encode the outcome as one-hot-representation + String outcomes[] = + new BioCodec().encode(sample.getNames(), sample.getSentence().length); + + INDArray labels = Nd4j.create(1, labelStrings.length, windowSize); + labels.putScalar(new int[]{0, labelToIndex.get(outcomes[i]), windowSize - 1}, 1.0d); + vectors.add(labels); + } + + return vectors; + } + + private static int max(INDArray array) { + int best = 0; + for (int i = 0; i < array.size(0); i++) { + if (array.getDouble(i) > array.getDouble(best)) { + best = i; + } + } + return best; + } + + @Override + public Span[] find(String[] tokens) { + List<INDArray> featureMartrices = mapToFeatureMatrices(wordVectors, tokens, windowSize); + + String[] outcomes = new String[tokens.length]; + for (int i = 0; i < tokens.length; i++) { + INDArray predictionMatrix = network.output(featureMartrices.get(i), false); + INDArray outcomeVector = predictionMatrix.get(NDArrayIndex.point(0), NDArrayIndex.all(), + NDArrayIndex.point(windowSize - 1)); + + outcomes[i] = labels[max(outcomeVector)]; + } + + // Delete invalid spans ... + for (int i = 0; i < outcomes.length; i++) { + if (outcomes[i].endsWith("cont") && (i == 0 || "other".equals(outcomes[i - 1]))) { + outcomes[i] = "other"; + } + } + + return new BioCodec().decode(Arrays.asList(outcomes)); + } + + @Override + public void clearAdaptiveData() { + } + + public static MultiLayerNetwork train(WordVectors wordVectors, ObjectStream<NameSample> samples, + int epochs, int windowSize, String[] labels) throws IOException { + int vectorSize = 300; + int layerSize = 256; + + MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder() + .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1) + .updater(Updater.RMSPROP) + .regularization(true).l2(0.001) + .weightInit(WeightInit.XAVIER) + // .gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue).gradientNormalizationThreshold(1.0) + .learningRate(0.01) + .list() + .layer(0, new GravesLSTM.Builder().nIn(vectorSize).nOut(layerSize) + .activation(Activation.TANH).build()) + .layer(1, new RnnOutputLayer.Builder().activation(Activation.SOFTMAX) + .lossFunction(LossFunctions.LossFunction.MCXENT).nIn(layerSize).nOut(3).build()) + .pretrain(false).backprop(true).build(); + + MultiLayerNetwork net = new MultiLayerNetwork(conf); + net.init(); + net.setListeners(new ScoreIterationListener(5)); + + // TODO: Extract labels on the fly from the data + + DataSetIterator train = new NameSampleDataSetIterator(samples, wordVectors, windowSize, labels); + + System.out.println("Starting training"); + + for (int i = 0; i < epochs; i++) { + net.fit(train); + train.reset(); + System.out.println(String.format("Finished epoch %d", i)); + } + + return net; + } + + public static void main(String[] args) throws Exception { + if (args.length != 3) { + System.out.println("Usage: trainFile testFile gloveTxt"); + return; + } + + String[] labels = new String[] { + "default-start", "default-cont", "other" + }; + + System.out.print("Loading vectors ... "); + WordVectors wordVectors = WordVectorSerializer.loadTxtVectors( + new File(args[2])); + System.out.println("Done"); + + int windowSize = 5; + + MultiLayerNetwork net = train(wordVectors, new NameSampleDataStream(new PlainTextByLineStream( + new MarkableFileInputStreamFactory(new File(args[0])), StandardCharsets.UTF_8)), 1, windowSize, labels); + + ObjectStream<NameSample> evalStream = new NameSampleDataStream(new PlainTextByLineStream( + new MarkableFileInputStreamFactory( + new File(args[1])), StandardCharsets.UTF_8)); + + NameFinderDL nameFinder = new NameFinderDL(net, wordVectors, windowSize, labels); + + System.out.print("Evaluating ... "); + TokenNameFinderEvaluator nameFinderEvaluator = new TokenNameFinderEvaluator(nameFinder); + nameFinderEvaluator.evaluate(evalStream); + + System.out.println("Done"); + + System.out.println(); + System.out.println(); + System.out.println("Results"); + + System.out.println(nameFinderEvaluator.getFMeasure().toString()); + } +} http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/main/java/opennlp/tools/dl/NameSampleDataSetIterator.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/main/java/opennlp/tools/dl/NameSampleDataSetIterator.java b/opennlp-dl/src/main/java/opennlp/tools/dl/NameSampleDataSetIterator.java new file mode 100644 index 0000000..a420220 --- /dev/null +++ b/opennlp-dl/src/main/java/opennlp/tools/dl/NameSampleDataSetIterator.java @@ -0,0 +1,225 @@ +package opennlp.tools.dl;/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +import java.io.IOException; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Collections; +import java.util.Iterator; +import java.util.List; +import java.util.NoSuchElementException; + +import org.deeplearning4j.models.embeddings.wordvectors.WordVectors; +import org.nd4j.linalg.api.ndarray.INDArray; +import org.nd4j.linalg.dataset.DataSet; +import org.nd4j.linalg.dataset.api.DataSetPreProcessor; +import org.nd4j.linalg.dataset.api.iterator.DataSetIterator; +import org.nd4j.linalg.factory.Nd4j; +import org.nd4j.linalg.indexing.INDArrayIndex; +import org.nd4j.linalg.indexing.NDArrayIndex; + +import opennlp.tools.namefind.NameSample; +import opennlp.tools.util.FilterObjectStream; +import opennlp.tools.util.ObjectStream; + +public class NameSampleDataSetIterator implements DataSetIterator { + + private static class NameSampleToDataSetStream extends FilterObjectStream<NameSample, DataSet> { + + private final WordVectors wordVectors; + private final String[] labels; + private int windowSize; + + private Iterator<DataSet> dataSets = Collections.emptyListIterator(); + + NameSampleToDataSetStream(ObjectStream<NameSample> samples, WordVectors wordVectors, int windowSize, String[] labels) { + super(samples); + this.wordVectors = wordVectors; + this.windowSize = windowSize; + this.labels = labels; + } + + private Iterator<DataSet> createDataSets(NameSample sample) { + List<INDArray> features = NameFinderDL.mapToFeatureMatrices(wordVectors, sample.getSentence(), + windowSize); + + List<INDArray> labels = NameFinderDL.mapToLabelVectors(sample, windowSize, this.labels); + + List<DataSet> dataSetList = new ArrayList<>(); + + for (int i = 0; i < features.size(); i++) { + dataSetList.add(new DataSet(features.get(i), labels.get(i))); + } + + return dataSetList.iterator(); + } + + @Override + public final DataSet read() throws IOException { + + if (dataSets.hasNext()) { + return dataSets.next(); + } + else { + NameSample sample; + while (!dataSets.hasNext() && (sample = samples.read()) != null) { + dataSets = createDataSets(sample); + } + + if (dataSets.hasNext()) { + return read(); + } + } + + return null; + } + } + + private final int windowSize; + private final String[] labels; + + private final int batchSize = 128; + private final int vectorSize = 300; + + private final int totalSamples; + + private int cursor = 0; + + private final ObjectStream<DataSet> samples; + + NameSampleDataSetIterator(ObjectStream<NameSample> samples, WordVectors wordVectors, int windowSize, + String labels[]) throws IOException { + this.windowSize = windowSize; + this.labels = labels; + + this.samples = new NameSampleToDataSetStream(samples, wordVectors, windowSize, labels); + + int total = 0; + + DataSet sample; + while ((sample = this.samples.read()) != null) { + total++; + } + + totalSamples = total; + + samples.reset(); + } + + public DataSet next(int num) { + if (cursor >= totalExamples()) throw new NoSuchElementException(); + + INDArray features = Nd4j.create(num, vectorSize, windowSize); + INDArray featuresMask = Nd4j.zeros(num, windowSize); + + INDArray labels = Nd4j.create(num, 3, windowSize); + INDArray labelsMask = Nd4j.zeros(num, windowSize); + + // iterate stream and copy to arrays + + for (int i = 0; i < num; i++) { + DataSet sample; + try { + sample = samples.read(); + } catch (IOException e) { + throw new RuntimeException(e); + } + + if (sample != null) { + INDArray feature = sample.getFeatureMatrix(); + features.put(new INDArrayIndex[] {NDArrayIndex.point(i)}, feature.get(NDArrayIndex.point(0))); + + feature.get(new INDArrayIndex[] {NDArrayIndex.point(0), NDArrayIndex.all(), + NDArrayIndex.point(0)}); + + for (int j = 0; j < windowSize; j++) { + featuresMask.putScalar(new int[] {i, j}, 1.0); + } + + INDArray label = sample.getLabels(); + labels.put(new INDArrayIndex[] {NDArrayIndex.point(i)}, label.get(NDArrayIndex.point(0))); + labelsMask.putScalar(new int[] {i, windowSize - 1}, 1.0); + } + + cursor++; + } + + return new DataSet(features, labels, featuresMask, labelsMask); + } + + public int totalExamples() { + return totalSamples; + } + + public int inputColumns() { + return vectorSize; + } + + public int totalOutcomes() { + return getLabels().size(); + } + + public boolean resetSupported() { + return true; + } + + public boolean asyncSupported() { + return false; + } + + public void reset() { + cursor = 0; + + try { + samples.reset(); + } catch (IOException e) { + throw new RuntimeException(e); + } + } + + public int batch() { + return batchSize; + } + + public int cursor() { + return cursor; + } + + public int numExamples() { + return totalExamples(); + } + + public void setPreProcessor(DataSetPreProcessor dataSetPreProcessor) { + throw new UnsupportedOperationException(); + } + + public DataSetPreProcessor getPreProcessor() { + throw new UnsupportedOperationException(); + } + + public List<String> getLabels() { + return Arrays.asList("start","cont", "other"); + } + + public boolean hasNext() { + return cursor < numExamples(); + } + + public DataSet next() { + return next(batchSize); + } +} http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/main/java/opennlp/tools/dl/RNN.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/main/java/opennlp/tools/dl/RNN.java b/opennlp-dl/src/main/java/opennlp/tools/dl/RNN.java new file mode 100644 index 0000000..155ec03 --- /dev/null +++ b/opennlp-dl/src/main/java/opennlp/tools/dl/RNN.java @@ -0,0 +1,366 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package opennlp.tools.dl; + +import java.io.BufferedWriter; +import java.io.File; +import java.io.FileWriter; +import java.io.IOException; +import java.util.Collections; +import java.util.Date; +import java.util.HashMap; +import java.util.HashSet; +import java.util.LinkedList; +import java.util.List; +import java.util.Map; +import java.util.Set; + +import org.apache.commons.math3.distribution.EnumeratedDistribution; +import org.apache.commons.math3.util.Pair; +import org.nd4j.linalg.api.iter.NdIndexIterator; +import org.nd4j.linalg.api.ndarray.INDArray; +import org.nd4j.linalg.api.ops.impl.transforms.SetRange; +import org.nd4j.linalg.api.ops.impl.transforms.SoftMax; +import org.nd4j.linalg.factory.Nd4j; +import org.nd4j.linalg.ops.transforms.Transforms; + +/** + * A min char/word-level vanilla RNN model, based on Andrej Karpathy's python code. + * See also: + * + * @see <a href="http://karpathy.github.io/2015/05/21/rnn-effectiveness";>The Unreasonable Effectiveness of Recurrent Neural Networks</a> + * @see <a href="https://gist.github.com/karpathy/d4dee566867f8291f086";>Minimal character-level language model with a Vanilla Recurrent Neural Network, in Python/numpy</a> + */ +public class RNN { + + // hyperparameters + protected final float learningRate; // size of hidden layer of neurons + protected final int seqLength; // no. of steps to unroll the RNN for + protected final int hiddenLayerSize; + protected final int epochs; + protected final boolean useChars; + protected final int vocabSize; + protected final Map<String, Integer> charToIx; + protected final Map<Integer, String> ixToChar; + protected final List<String> data; + private final static double reg = 1e-8; + + // model parameters + private final INDArray wxh; // input to hidden + private final INDArray whh; // hidden to hidden + private final INDArray why; // hidden to output + private final INDArray bh; // hidden bias + private final INDArray by; // output bias + + private INDArray hPrev = null; // memory state + + public RNN(float learningRate, int seqLength, int hiddenLayerSize, int epochs, String text) { + this(learningRate, seqLength, hiddenLayerSize, epochs, text, true); + } + + public RNN(float learningRate, int seqLength, int hiddenLayerSize, int epochs, String text, boolean useChars) { + this.learningRate = learningRate; + this.seqLength = seqLength; + this.hiddenLayerSize = hiddenLayerSize; + this.epochs = epochs; + this.useChars = useChars; + + String[] textTokens = useChars ? toStrings(text.toCharArray()) : text.split(" "); + data = new LinkedList<>(); + Collections.addAll(data, textTokens); + Set<String> tokens = new HashSet<>(data); + vocabSize = tokens.size(); + + System.out.printf("data has %d tokens, %d unique.\n", data.size(), vocabSize); + charToIx = new HashMap<>(); + ixToChar = new HashMap<>(); + int i = 0; + for (String c : tokens) { + charToIx.put(c, i); + ixToChar.put(i, c); + i++; + } + + wxh = Nd4j.randn(hiddenLayerSize, vocabSize).mul(0.01); + whh = Nd4j.randn(hiddenLayerSize, hiddenLayerSize).mul(0.01); + why = Nd4j.randn(vocabSize, hiddenLayerSize).mul(0.01); + bh = Nd4j.zeros(hiddenLayerSize, 1); + by = Nd4j.zeros(vocabSize, 1); + } + + private String[] toStrings(char[] chars) { + String[] strings = new String[chars.length]; + for (int i = 0; i < chars.length; i++) { + strings[i] = String.valueOf(chars[i]); + } + return strings; + } + + public void learn() { + + int currentEpoch = 0; + + int n = 0; + int p = 0; + + // memory variables for Adagrad + INDArray mWxh = Nd4j.zerosLike(wxh); + INDArray mWhh = Nd4j.zerosLike(whh); + INDArray mWhy = Nd4j.zerosLike(why); + + INDArray mbh = Nd4j.zerosLike(bh); + INDArray mby = Nd4j.zerosLike(by); + + // loss at iteration 0 + double smoothLoss = -Math.log(1.0 / vocabSize) * seqLength; + + while (true) { + // prepare inputs (we're sweeping from left to right in steps seqLength long) + if (p + seqLength + 1 >= data.size() || n == 0) { + hPrev = Nd4j.zeros(hiddenLayerSize, 1); // reset RNN memory + p = 0; // go from start of data + currentEpoch++; + if (currentEpoch == epochs) { + System.out.println("training finished: e:" + epochs + ", l: " + smoothLoss + ", h:(" + learningRate + ", " + seqLength + ", " + hiddenLayerSize + ")"); + break; + } + } + + INDArray inputs = getSequence(p); + INDArray targets = getSequence(p + 1); + + // sample from the model every now and then + if (n % 1000 == 0 && n > 0) { + String txt = sample(inputs.getInt(0)); + System.out.printf("\n---\n %s \n----\n", txt); + } + + INDArray dWxh = Nd4j.zerosLike(wxh); + INDArray dWhh = Nd4j.zerosLike(whh); + INDArray dWhy = Nd4j.zerosLike(why); + + INDArray dbh = Nd4j.zerosLike(bh); + INDArray dby = Nd4j.zerosLike(by); + + // forward seqLength characters through the net and fetch gradient + double loss = lossFun(inputs, targets, dWxh, dWhh, dWhy, dbh, dby); + smoothLoss = smoothLoss * 0.999 + loss * 0.001; + if (Double.isNaN(smoothLoss)) { + System.out.println("loss is NaN (over/underflow occured, try adjusting hyperparameters)"); + break; + } + if (n % 100 == 0) { + System.out.printf("iter %d, loss: %f\n", n, smoothLoss); // print progress + } + + // perform parameter update with Adagrad + mWxh.addi(dWxh.mul(dWxh)); + wxh.subi((dWxh.mul(learningRate)).div(Transforms.sqrt(mWxh.add(reg)))); + + mWhh.addi(dWhh.mul(dWhh)); + whh.subi(dWhh.mul(learningRate).div(Transforms.sqrt(mWhh.add(reg)))); + + mWhy.addi(dWhy.mul(dWhy)); + why.subi(dWhy.mul(learningRate).div(Transforms.sqrt(mWhy.add(reg)))); + + mbh.addi(dbh.mul(dbh)); + bh.subi(dbh.mul(learningRate).div(Transforms.sqrt(mbh.add(reg)))); + + mby.addi(dby.mul(dby)); + by.subi(dby.mul(learningRate).div(Transforms.sqrt(mby.add(reg)))); + + p += seqLength; // move data pointer + n++; // iteration counter + } + } + + protected INDArray getSequence(int p) { + INDArray inputs = Nd4j.create(seqLength); + int c = 0; + for (String ch : data.subList(p, p + seqLength)) { + Integer ix = charToIx.get(ch); + inputs.putScalar(c, ix); + c++; + } + return inputs; + } + + /** + * inputs, targets are both list of integers + * hprev is Hx1 array of initial hidden state + * returns the modified loss, gradients on model parameters + */ + private double lossFun(INDArray inputs, INDArray targets, INDArray dWxh, INDArray dWhh, INDArray dWhy, INDArray dbh, + INDArray dby) { + + INDArray xs = Nd4j.zeros(inputs.length(), vocabSize); + INDArray hs = null; + INDArray ys = null; + INDArray ps = null; + + INDArray hs1 = Nd4j.create(hPrev.shape()); + Nd4j.copy(hPrev, hs1); + + double loss = 0; + + // forward pass + for (int t = 0; t < inputs.length(); t++) { + int tIndex = inputs.getScalar(t).getInt(0); + xs.putScalar(t, tIndex, 1); // encode in 1-of-k representation + INDArray hsRow = t == 0 ? hs1 : hs.getRow(t - 1); + INDArray hst = Transforms.tanh(wxh.mmul(xs.getRow(t).transpose()).add(whh.mmul(hsRow)).add(bh)); // hidden state + if (hs == null) { + hs = init(inputs.length(), hst.shape()); + } + hs.putRow(t, hst); + + INDArray yst = (why.mmul(hst)).add(by); // unnormalized log probabilities for next chars + if (ys == null) { + ys = init(inputs.length(), yst.shape()); + } + ys.putRow(t, yst); + INDArray pst = Nd4j.getExecutioner().execAndReturn(new SoftMax(yst)); // probabilities for next chars + if (ps == null) { + ps = init(inputs.length(), pst.shape()); + } + ps.putRow(t, pst); + loss += -Math.log(pst.getDouble(targets.getInt(t))); // softmax (cross-entropy loss) + } + + // backward pass: compute gradients going backwards + INDArray dhNext = Nd4j.zerosLike(hs.getRow(0)); + for (int t = inputs.length() - 1; t >= 0; t--) { + INDArray dy = ps.getRow(t); + dy.putRow(targets.getInt(t), dy.getRow(targets.getInt(t)).sub(1)); // backprop into y + INDArray hst = hs.getRow(t); + dWhy.addi(dy.mmul(hst.transpose())); // derivative of hy layer + dby.addi(dy); + INDArray dh = why.transpose().mmul(dy).add(dhNext); // backprop into h + INDArray dhraw = (Nd4j.ones(hst.shape()).sub(hst.mul(hst))).mul(dh); // backprop through tanh nonlinearity + dbh.addi(dhraw); + dWxh.addi(dhraw.mmul(xs.getRow(t))); + INDArray hsRow = t == 0 ? hs1 : hs.getRow(t - 1); + dWhh.addi(dhraw.mmul(hsRow.transpose())); + dhNext = whh.transpose().mmul(dhraw); + } + + this.hPrev = hs.getRow(inputs.length() - 1); + + return loss; + } + + protected INDArray init(int t, int[] aShape) { + INDArray as; + int[] shape = new int[1 + aShape.length]; + shape[0] = t; + System.arraycopy(aShape, 0, shape, 1, aShape.length); + as = Nd4j.create(shape); + return as; + } + + /** + * sample a sequence of integers from the model, using current (hPrev) memory state, seedIx is seed letter for first time step + */ + public String sample(int seedIx) { + + INDArray x = Nd4j.zeros(vocabSize, 1); + x.putScalar(seedIx, 1); + int sampleSize = 2 * seqLength; + INDArray ixes = Nd4j.create(sampleSize); + + INDArray h = hPrev.dup(); + + for (int t = 0; t < sampleSize; t++) { + h = Transforms.tanh(wxh.mmul(x).add(whh.mmul(h)).add(bh)); + INDArray y = (why.mmul(h)).add(by); + INDArray pm = Nd4j.getExecutioner().execAndReturn(new SoftMax(y)).ravel(); + + List<Pair<Integer, Double>> d = new LinkedList<>(); + for (int pi = 0; pi < vocabSize; pi++) { + d.add(new Pair<>(pi, pm.getDouble(0, pi))); + } + EnumeratedDistribution<Integer> distribution = new EnumeratedDistribution<>(d); + + int ix = distribution.sample(); + + x = Nd4j.zeros(vocabSize, 1); + x.putScalar(ix, 1); + ixes.putScalar(t, ix); + } + + return getSampleString(ixes); + } + + protected String getSampleString(INDArray ixes) { + StringBuilder txt = new StringBuilder(); + + NdIndexIterator ndIndexIterator = new NdIndexIterator(ixes.shape()); + while (ndIndexIterator.hasNext()) { + int[] next = ndIndexIterator.next(); + if (!useChars && txt.length() > 0) { + txt.append(' '); + } + txt.append(ixToChar.get(ixes.getInt(next))); + } + return txt.toString(); + } + + public int getVocabSize() { + return vocabSize; + } + + @Override + public String toString() { + return getClass().getName() + "{" + + "learningRate=" + learningRate + + ", seqLength=" + seqLength + + ", hiddenLayerSize=" + hiddenLayerSize + + ", epochs=" + epochs + + ", vocabSize=" + vocabSize + + ", useChars=" + useChars + + '}'; + } + + + public String getHyperparamsString() { + return getClass().getName() + "{" + + "wxh=" + wxh + + ", whh=" + whh + + ", why=" + why + + ", bh=" + bh + + ", by=" + by + + '}'; + } + + public void serialize(String prefix) throws IOException { + BufferedWriter bufferedWriter = new BufferedWriter(new FileWriter(new File(prefix + new Date().toString() + ".txt"))); + bufferedWriter.write("wxh"); + bufferedWriter.write(wxh.toString()); + bufferedWriter.write("whh"); + bufferedWriter.write(whh.toString()); + bufferedWriter.write("why"); + bufferedWriter.write(why.toString()); + bufferedWriter.write("bh"); + bufferedWriter.write(bh.toString()); + bufferedWriter.write("by"); + bufferedWriter.write(by.toString()); + bufferedWriter.flush(); + bufferedWriter.close(); + } +} http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/main/java/opennlp/tools/dl/StackedRNN.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/main/java/opennlp/tools/dl/StackedRNN.java b/opennlp-dl/src/main/java/opennlp/tools/dl/StackedRNN.java new file mode 100644 index 0000000..e7a49d7 --- /dev/null +++ b/opennlp-dl/src/main/java/opennlp/tools/dl/StackedRNN.java @@ -0,0 +1,323 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package opennlp.tools.dl; + +import org.apache.commons.math3.distribution.EnumeratedDistribution; +import org.apache.commons.math3.util.Pair; +import org.nd4j.linalg.api.ndarray.INDArray; +import org.nd4j.linalg.api.ops.impl.transforms.SetRange; +import org.nd4j.linalg.api.ops.impl.transforms.SoftMax; +import org.nd4j.linalg.factory.Nd4j; +import org.nd4j.linalg.ops.transforms.Transforms; + +import java.io.BufferedWriter; +import java.io.File; +import java.io.FileWriter; +import java.io.IOException; +import java.util.Date; +import java.util.LinkedList; +import java.util.List; + +/** + * A basic char/word-level stacked RNN model (2 hidden recurrent layers), based on Stacked RNN architecture from ICLR 2014's + * "How to Construct Deep Recurrent Neural Networks" by Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho and Yoshua Bengio + * and Andrej Karpathy's notes on RNNs. + * See also: + * + * @see <a href="http://karpathy.github.io/2015/05/21/rnn-effectiveness";>The Unreasonable Effectiveness of Recurrent Neural Networks</a> + * @see <a href="https://arxiv.org/abs/1312.6026";>How to Construct Deep Recurrent Neural Networks</a> + */ +public class StackedRNN extends RNN { + + // model parameters + private final INDArray wxh; // input to hidden + private final INDArray whh; // hidden to hidden + private final INDArray whh2; // hidden to hidden2 + private final INDArray wh2y; // hidden2 to output + private final INDArray wxh2; + private final INDArray bh; // hidden bias + private final INDArray bh2; // hidden2 bias + private final INDArray by; // output bias + + private final double reg = 1e-8; + + private INDArray hPrev = null; // memory state + private INDArray hPrev2 = null; // memory state + + public StackedRNN(float learningRate, int seqLength, int hiddenLayerSize, int epochs, String text) { + this(learningRate, seqLength, hiddenLayerSize, epochs, text, true); + } + + public StackedRNN(float learningRate, int seqLength, int hiddenLayerSize, int epochs, String text, boolean useChars) { + super(learningRate, seqLength, hiddenLayerSize, epochs, text, useChars); + + wxh = Nd4j.randn(hiddenLayerSize, vocabSize).div(Math.sqrt(hiddenLayerSize)); + whh = Nd4j.randn(hiddenLayerSize, hiddenLayerSize).div(Math.sqrt(hiddenLayerSize)); + whh2 = Nd4j.randn(hiddenLayerSize, hiddenLayerSize).div(Math.sqrt(hiddenLayerSize)); + wxh2 = Nd4j.randn(hiddenLayerSize, hiddenLayerSize).div(Math.sqrt(hiddenLayerSize)); + wh2y = Nd4j.randn(vocabSize, hiddenLayerSize).div(Math.sqrt(vocabSize)); + bh = Nd4j.zeros(hiddenLayerSize, 1); + bh2 = Nd4j.zeros(hiddenLayerSize, 1); + by = Nd4j.zeros(vocabSize, 1); + } + + public void learn() { + + int currentEpoch = -1; + + int n = 0; + int p = 0; + + // memory variables for Adagrad + INDArray mWxh = Nd4j.zerosLike(wxh); + INDArray mWxh2 = Nd4j.zerosLike(wxh2); + INDArray mWhh = Nd4j.zerosLike(whh); + INDArray mWhh2 = Nd4j.zerosLike(whh2); + INDArray mWh2y = Nd4j.zerosLike(wh2y); + + INDArray mbh = Nd4j.zerosLike(bh); + INDArray mbh2 = Nd4j.zerosLike(bh2); + INDArray mby = Nd4j.zerosLike(by); + + // loss at iteration 0 + double smoothLoss = -Math.log(1.0 / vocabSize) * seqLength; + + while (true) { + // prepare inputs (we're sweeping from left to right in steps seqLength long) + if (p + seqLength + 1 >= data.size() || n == 0) { + hPrev = Nd4j.zeros(hiddenLayerSize, 1); // reset RNN memory + hPrev2 = Nd4j.zeros(hiddenLayerSize, 1); // reset RNN memory + p = 0; // go from start of data + currentEpoch++; + if (currentEpoch == epochs) { + System.out.println("training finished: e:" + epochs + ", l: " + smoothLoss + ", h:(" + learningRate + ", " + seqLength + ", " + hiddenLayerSize + ")"); + break; + } + } + + INDArray inputs = getSequence(p); + INDArray targets = getSequence(p + 1); + + // sample from the model every now and then + if (n % 1000 == 0 && n > 0) { + String txt = sample(inputs.getInt(0)); + System.out.printf("\n---\n %s \n----\n", txt); + } + + INDArray dWxh = Nd4j.zerosLike(wxh); + INDArray dWxh2 = Nd4j.zerosLike(wxh2); + INDArray dWhh = Nd4j.zerosLike(whh); + INDArray dWhh2 = Nd4j.zerosLike(whh2); + INDArray dWh2y = Nd4j.zerosLike(wh2y); + + INDArray dbh = Nd4j.zerosLike(bh); + INDArray dbh2 = Nd4j.zerosLike(bh); + INDArray dby = Nd4j.zerosLike(by); + + // forward seqLength characters through the net and fetch gradient + double loss = lossFun(inputs, targets, dWxh, dWhh, dWxh2, dWhh2, dWh2y, dbh, dbh2, dby); + smoothLoss = smoothLoss * 0.999 + loss * 0.001; + if (Double.isNaN(smoothLoss) || Double.isInfinite(smoothLoss)) { + System.out.println("loss is " + smoothLoss + " (over/underflow occured, try adjusting hyperparameters)"); + break; + } + if (n % 100 == 0) { + System.out.printf("iter %d, loss: %f\n", n, smoothLoss); // print progress + } + + // perform parameter update with Adagrad + mWxh.addi(dWxh.mul(dWxh)); + wxh.subi(dWxh.mul(learningRate).div(Transforms.sqrt(mWxh.add(reg)))); + + mWxh2.addi(dWxh2.mul(dWxh2)); + wxh2.subi(dWxh2.mul(learningRate).div(Transforms.sqrt(mWxh2.add(reg)))); + + mWhh.addi(dWhh.mul(dWhh)); + whh.subi(dWhh.mul(learningRate).div(Transforms.sqrt(mWhh.add(reg)))); + + mWhh2.addi(dWhh2.mul(dWhh2)); + whh2.subi(dWhh2.mul(learningRate).div(Transforms.sqrt(mWhh2.add(reg)))); + + mbh2.addi(dbh2.mul(dbh2)); + bh2.subi(dbh2.mul(learningRate).div(Transforms.sqrt(mbh2.add(reg)))); + + mWh2y.addi(dWh2y.mul(dWh2y)); + wh2y.subi(dWh2y.mul(learningRate).div(Transforms.sqrt(mWh2y.add(reg)))); + + mbh.addi(dbh.mul(dbh)); + bh.subi(dbh.mul(learningRate).div(Transforms.sqrt(mbh.add(reg)))); + + mby.addi(dby.mul(dby)); + by.subi(dby.mul(learningRate).div(Transforms.sqrt(mby.add(reg)))); + + p += seqLength; // move data pointer + n++; // iteration counter + } + } + + /** + * inputs, targets are both list of integers + * hprev is Hx1 array of initial hidden state + * returns the loss, gradients on model parameters and last hidden state + */ + private double lossFun(INDArray inputs, INDArray targets, INDArray dWxh, INDArray dWhh, INDArray dWxh2, INDArray dWhh2, INDArray dWh2y, + INDArray dbh, INDArray dbh2, INDArray dby) { + + INDArray xs = Nd4j.zeros(seqLength, vocabSize); + INDArray hs = null; + INDArray hs2 = null; + INDArray ys = null; + INDArray ps = null; + + double loss = 0; + + // forward pass + for (int t = 0; t < seqLength; t++) { + int tIndex = inputs.getScalar(t).getInt(0); + xs.putScalar(t, tIndex, 1); // encode in 1-of-k representation + + INDArray xst = xs.getRow(t); + + hPrev = Transforms.tanh((wxh.mmul(xst.transpose()).add(whh.mmul(hPrev)).add(bh))); // hidden state + if (hs == null) { + hs = init(seqLength, hPrev.shape()); + } + hs.putRow(t, hPrev.dup()); + + hPrev2 = Transforms.tanh((wxh2.mmul(hPrev).add(whh2.mmul(hPrev2)).add(bh2))); // hidden state 2 + if (hs2 == null) { + hs2 = init(seqLength, hPrev2.shape()); + } + hs2.putRow(t, hPrev2.dup()); + + INDArray yst = wh2y.mmul(hPrev2).add(by); // unnormalized log probabilities for next chars + if (ys == null) { + ys = init(seqLength, yst.shape()); + } + ys.putRow(t, yst); + + INDArray pst = Nd4j.getExecutioner().execAndReturn(new SoftMax(yst)); // probabilities for next chars + if (ps == null) { + ps = init(seqLength, pst.shape()); + } + ps.putRow(t, pst); + + loss += -Math.log(pst.getDouble(targets.getInt(t))); // softmax (cross-entropy loss) + } + + // backward pass: compute gradients going backwards + INDArray dhNext = Nd4j.zerosLike(hs.getRow(0)); + INDArray dh2Next = Nd4j.zerosLike(hs2.getRow(0)); + for (int t = seqLength - 1; t >= 0; t--) { + INDArray dy = ps.getRow(t); + dy.getRow(targets.getInt(t)).subi(1); // backprop into y + + INDArray hs2t = hs2.getRow(t); + INDArray hs2tm1 = t == 0 ? hPrev2 : hs2.getRow(t - 1); + + dWh2y.addi(dy.mmul(hs2t.transpose())); + dby.addi(dy); + + INDArray dh2 = wh2y.transpose().mmul(dy).add(dh2Next); // backprop into h2 + INDArray dhraw2 = (Nd4j.ones(hs2t.shape()).sub(hs2t.mul(hs2t))).mul(dh2); // backprop through tanh nonlinearity + dbh2.addi(dhraw2); + INDArray hst = hs.getRow(t); + dWxh2.addi(dhraw2.mmul(hst.transpose())); + dWhh2.addi(dhraw2.mmul(hs2tm1.transpose())); + dh2Next = whh2.transpose().mmul(dhraw2); + + INDArray dh = wxh2.transpose().mmul(dhraw2).add(dhNext); // backprop into h + INDArray dhraw = (Nd4j.ones(hst.shape()).sub(hst.mul(hst))).mul(dh); // backprop through tanh nonlinearity + dbh.addi(dhraw); + dWxh.addi(dhraw.mmul(xs.getRow(t))); + INDArray hsRow = t == 0 ? hPrev : hs.getRow(t - 1); + dWhh.addi(dhraw.mmul(hsRow.transpose())); + dhNext = whh.transpose().mmul(dhraw); + + } + + this.hPrev = hs.getRow(seqLength - 1); + this.hPrev2 = hs2.getRow(seqLength - 1); + + return loss; + } + + /** + * sample a sequence of integers from the model, using current (hPrev) memory state, seedIx is seed letter for first time step + */ + @Override + public String sample(int seedIx) { + + INDArray x = Nd4j.zeros(vocabSize, 1); + x.putScalar(seedIx, 1); + int sampleSize = seqLength * 2; + INDArray ixes = Nd4j.create(sampleSize); + + INDArray h = hPrev.dup(); + INDArray h2 = hPrev2.dup(); + + for (int t = 0; t < sampleSize; t++) { + h = Transforms.tanh((wxh.mmul(x)).add(whh.mmul(h)).add(bh)); + h2 = Transforms.tanh((wxh2.mmul(h)).add(whh2.mmul(h2)).add(bh2)); + INDArray y = wh2y.mmul(h2).add(by); + INDArray pm = Nd4j.getExecutioner().execAndReturn(new SoftMax(y)).ravel(); + + List<Pair<Integer, Double>> d = new LinkedList<>(); + for (int pi = 0; pi < vocabSize; pi++) { + d.add(new Pair<>(pi, pm.getDouble(0, pi))); + } + try { + EnumeratedDistribution<Integer> distribution = new EnumeratedDistribution<>(d); + + int ix = distribution.sample(); + + x = Nd4j.zeros(vocabSize, 1); + x.putScalar(ix, 1); + ixes.putScalar(t, ix); + } catch (Exception e) { + } + } + + return getSampleString(ixes); + } + + @Override + public void serialize(String prefix) throws IOException { + BufferedWriter bufferedWriter = new BufferedWriter(new FileWriter(new File(prefix + new Date().toString() + ".txt"))); + bufferedWriter.write("wxh"); + bufferedWriter.write(wxh.toString()); + bufferedWriter.write("whh"); + bufferedWriter.write(whh.toString()); + bufferedWriter.write("wxh2"); + bufferedWriter.write(wxh2.toString()); + bufferedWriter.write("whh2"); + bufferedWriter.write(whh2.toString()); + bufferedWriter.write("wh2y"); + bufferedWriter.write(wh2y.toString()); + bufferedWriter.write("bh"); + bufferedWriter.write(bh.toString()); + bufferedWriter.write("bh2"); + bufferedWriter.write(bh2.toString()); + bufferedWriter.write("by"); + bufferedWriter.write(by.toString()); + bufferedWriter.flush(); + bufferedWriter.close(); + } + +} \ No newline at end of file http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/test/java/opennlp/tools/dl/RNNTest.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/test/java/opennlp/tools/dl/RNNTest.java b/opennlp-dl/src/test/java/opennlp/tools/dl/RNNTest.java new file mode 100644 index 0000000..2808f4d --- /dev/null +++ b/opennlp-dl/src/test/java/opennlp/tools/dl/RNNTest.java @@ -0,0 +1,110 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package opennlp.tools.dl; + +import java.io.FileInputStream; +import java.io.InputStream; +import java.util.Arrays; +import java.util.Collection; +import java.util.List; +import java.util.Random; + +import org.apache.commons.io.IOUtils; +import org.junit.Before; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.junit.runners.Parameterized; + +/** + * CV tests for {@link RNN} + */ +@RunWith(Parameterized.class) +public class RNNTest { + + private float learningRate; + private int seqLength; + private int hiddenLayerSize; + private Random r = new Random(); + private String text; + private final int epochs = 20; + private List<String> words; + + public RNNTest(float learningRate, int seqLength, int hiddenLayerSize) { + this.learningRate = learningRate; + this.seqLength = seqLength; + this.hiddenLayerSize = hiddenLayerSize; + } + + @Before + public void setUp() throws Exception { + InputStream stream = getClass().getResourceAsStream("/text/sentences.txt"); + text = IOUtils.toString(stream); + words = Arrays.asList(text.split(" ")); + stream.close(); + } + + @Parameterized.Parameters + public static Collection<Object[]> data() { + return Arrays.asList(new Object[][] { + {1e-1f, 25, 20}, + {1e-1f, 25, 40}, + {1e-1f, 25, 60}, + }); + } + + @Test + public void testVanillaCharRNNLearn() throws Exception { + RNN rnn = new RNN(learningRate, seqLength, hiddenLayerSize, epochs, text); + evaluate(rnn, true); + rnn.serialize("target/crnn-weights-"); + } + + @Test + public void testVanillaWordRNNLearn() throws Exception { + RNN rnn = new RNN(learningRate, seqLength, hiddenLayerSize, epochs * 2, text, false); + evaluate(rnn, true); + rnn.serialize("target/wrnn-weights-"); + } + + private void evaluate(RNN rnn, boolean checkRatio) { + System.out.println(rnn); + rnn.learn(); + double c = 0; + for (int i = 0; i < 2; i++) { + int seed = r.nextInt(rnn.getVocabSize()); + String sample = rnn.sample(seed); + System.out.println(sample); + if (checkRatio && rnn.useChars) { + String[] sampleWords = sample.split(" "); + for (String sw : sampleWords) { + if (words.contains(sw)) { + c++; + } + } + if (c > 0) { + c /= sampleWords.length; + } + } + } + if (checkRatio && rnn.useChars) { + System.out.println("average correct word ratio: " + (c / 10d)); + } + } + +} \ No newline at end of file http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/test/java/opennlp/tools/dl/StackedRNNTest.java ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/test/java/opennlp/tools/dl/StackedRNNTest.java b/opennlp-dl/src/test/java/opennlp/tools/dl/StackedRNNTest.java new file mode 100644 index 0000000..ac0434c --- /dev/null +++ b/opennlp-dl/src/test/java/opennlp/tools/dl/StackedRNNTest.java @@ -0,0 +1,109 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package opennlp.tools.dl; + +import java.io.InputStream; +import java.util.Arrays; +import java.util.Collection; +import java.util.List; +import java.util.Random; + +import org.apache.commons.io.IOUtils; +import org.junit.Before; +import org.junit.Test; +import org.junit.runner.RunWith; +import org.junit.runners.Parameterized; + +/** + * Tests for {@link StackedRNN} + */ +@RunWith(Parameterized.class) +public class StackedRNNTest { + + private float learningRate; + private int seqLength; + private int hiddenLayerSize; + private Random r = new Random(); + private String text; + private final int epochs = 20; + private List<String> words; + + public StackedRNNTest(float learningRate, int seqLength, int hiddenLayerSize) { + this.learningRate = learningRate; + this.seqLength = seqLength; + this.hiddenLayerSize = hiddenLayerSize; + } + + @Before + public void setUp() throws Exception { + InputStream stream = getClass().getResourceAsStream("/text/sentences.txt"); + text = IOUtils.toString(stream); + words = Arrays.asList(text.split(" ")); + stream.close(); + } + + @Parameterized.Parameters + public static Collection<Object[]> data() { + return Arrays.asList(new Object[][] { + {1e-1f, 25, 20}, + {1e-1f, 25, 40}, + {1e-1f, 25, 60}, + }); + } + + @Test + public void testStackedCharRNNLearn() throws Exception { + RNN rnn = new StackedRNN(learningRate, seqLength, hiddenLayerSize, epochs, text); + evaluate(rnn, true); + rnn.serialize("target/scrnn-weights-"); + } + + @Test + public void testStackedWordRNNLearn() throws Exception { + RNN rnn = new StackedRNN(learningRate, seqLength, hiddenLayerSize, epochs, text, false); + evaluate(rnn, true); + rnn.serialize("target/swrnn-weights-"); + } + + private void evaluate(RNN rnn, boolean checkRatio) { + System.out.println(rnn); + rnn.learn(); + double c = 0; + for (int i = 0; i < 2; i++) { + int seed = r.nextInt(rnn.getVocabSize()); + String sample = rnn.sample(seed); + System.out.println(sample); + if (checkRatio && rnn.useChars) { + String[] sampleWords = sample.split(" "); + for (String sw : sampleWords) { + if (words.contains(sw)) { + c++; + } + } + if (c > 0) { + c /= sampleWords.length; + } + } + } + if (checkRatio && rnn.useChars) { + System.out.println("average correct word ratio: " + (c / 10d)); + } + } + +} \ No newline at end of file http://git-wip-us.apache.org/repos/asf/opennlp-sandbox/blob/a63ec16c/opennlp-dl/src/test/resources/text/sentences.txt ---------------------------------------------------------------------- diff --git a/opennlp-dl/src/test/resources/text/sentences.txt b/opennlp-dl/src/test/resources/text/sentences.txt new file mode 100644 index 0000000..3079de1 --- /dev/null +++ b/opennlp-dl/src/test/resources/text/sentences.txt @@ -0,0 +1,52 @@ +The word2vec software of Tomas Mikolov and colleagues has gained a lot of traction lately and provides state-of-the-art word embeddings . +The learning models behind the software are described in two research papers . +We found the description of the models in these papers to be somewhat cryptic and hard to follow . +While the motivations and presentation may be obvious to the neural-networks language-modeling crowd we had to struggle quite a bit to figure out the rationale behind the equations . +This note is an attempt to explain the negative sampling equation in Distributed Representations of Words and Phrases and their Compositionality by Tomas Mikolov Ilya Sutskever Kai Chen Greg Corrado and Jeffrey Dean . +The departure point of the paper is the skip-gram model . +In this model we are given a corpus of words w and their contexts c . +We consider the conditional probabilities p(c|w) and given a corpus Text the goal is to set the parameters θ of p(c|w;θ) so as to maximize the corpus probability . +The recently introduced continuous Skip-gram model is an efficient method for learning high-quality distributed vector representations that capture a large number of precise syntactic and semantic word relationships . +In this paper we present several extensions that improve both the quality of the vectors and the training speed . +By subsampling of the frequent words we obtain significant speedup and also learn more regular word representations . +We also describe a simple alternative to the hierarchical softmax called negative sampling . +An inherent limitation of word representations is their indifference to word order and their inability to represent idiomatic phrases . +For example the meanings of Canada and Air cannot be easily combined to obtain Air Canada . +Motivated by this example we present a simple method for finding phrases in text and show that learning good vector representations for millions of phrases is possible . +The similarity metrics used for nearest neighbor evaluations produce a single scalar that quantifies the relatedness of two words . +This simplicity can be problematic since two given words almost always exhibit more intricate relationships than can be captured by a single number . +For example man may be regarded as similar to woman in that both words describe human beings on the other hand the two words are often considered opposites since they highlight a primary axis along which humans differ from one another . +In order to capture in a quantitative way the nuance necessary to distinguish man from woman it is necessary for a model to associate more than a single number to the word pair . +A natural and simple candidate for an enlarged set of discriminative numbers is the vector difference between the two word vectors . +GloVe is designed in order that such vector differences capture as much as possible the meaning specified by the juxtaposition of two words . +Unsupervised word representations are very useful in NLP tasks both as inputs to learning algorithms and as extra word features in NLP systems . +However most of these models are built with only local context and one representation per word . +This is problematic because words are often polysemous and global context can also provide useful information for learning word meanings . +We present a new neural network architecture which 1) learns word embeddings that better capture the semantics of words by incorporating both local and global document context and 2) accounts for homonymy and polysemy by learning multiple embeddings per word . +We introduce a new dataset with human judgments on pairs of words in sentential context and evaluate our model on it showing that our model outperforms competitive baselines and other neural language models . +Information Retrieval ( IR ) models need to deal with two difficult issues vocabulary mismatch and term dependencies . +Vocabulary mismatch corresponds to the difficulty of retrieving relevant documents that do not contain exact query terms but semantically related terms . +Term dependencies refers to the need of considering the relationship between the words of the query when estimating the relevance of a document . +A multitude of solutions has been proposed to solve each of these two problems but no principled model solve both . +In parallel in the last few years language models based on neural networks have been used to cope with complex natural language processing tasks like emotion and paraphrase detection . +Although they present good abilities to cope with both term dependencies and vocabulary mismatch problems thanks to the distributed representation of words they are based upon such models could not be used readily in IR where the estimation of one language model per document ( or query ) is required . +This is both computationally unfeasible and prone to over-fitting . +Based on a recent work that proposed to learn a generic language model that can be modified through a set of document-specific parameters we explore use of new neural network models that are adapted to ad-hoc IR tasks . +Within the language model IR framework we propose and study the use of a generic language model as well as a document-specific language model . +Both can be used as a smoothing component but the latter is more adapted to the document at hand and has the potential of being used as a full document language model . +We experiment with such models and analyze their results on TREC-1 to 8 datasets . +The word2vec model and application by Mikolov et al have attracted a great amount of attention in recent two years . +The vector representations of words learned by word2vec models have been proven to be able to carry semantic meanings and are useful in various NLP tasks . +As an increasing number of researchers would like to experiment with word2vec I notice that there lacks a material that comprehensively explains the parameter learning process of word2vec in details thus preventing many people with less neural network experience from understanding how exactly word2vec works . +This note provides detailed derivations and explanations of the parameter update equations for the word2vec models including the original continuous bag-of-word ( CBOW ) and skip-gram models as well as advanced tricks hierarchical soft-max and negative sampling . +In the appendix a review is given on the basics of neuron network models and backpropagation . +To avoid the inaccuracy caused by classifying the example into several categories given by TREC manually we take the word2vec to represent all attractions and user contexts in the continuous vector space learnt by neural network language models . +The base of NNML is using neural networks for the probability function . +The model learns simultaneously a distributed representation for each word along with the probability function for word sequences expressed in terms of these representations . +Training such large models we propose continuous bag of words as our framework and soft-max as the active function . +So we use the word2vec to train wikitravel corpus and got the word vector . +To avoid the curse of dimensionality by learning a distributed representation for words as our word vector we define a test set that compare different dimensionality of vectors for our task using the same training data and using the same model architecture . +We extend the word2vec framework to capture meaning across languages . +The input consists of a source text and a word-aligned parallel text in a second language . +The joint word2vec tool then represents words in both languages within a common âsemanticâ vector space . +The result can be used to enrich lexicons of under-resourced languages to identify ambiguities and to perform clustering and classification . \ No newline at end of file
