Examples with INDArray org.nd4j.linalg.api.ndarray.INDArray used on opensource projects

Example 56 with INDArray

use of org.nd4j.linalg.api.ndarray.INDArray in project deeplearning4j by deeplearning4j.

the class GloVe method update.

private void update(T element1, INDArray wordVector, INDArray contextVector, double gradient) {
    //gradient for word vectors
    INDArray grad1 = contextVector.mul(gradient);
    INDArray update = weightAdaGrad.getGradient(grad1, element1.getIndex(), syn0.shape());
    //update vector
    wordVector.subi(update);
    double w1Bias = bias.getDouble(element1.getIndex());
    double biasGradient = biasAdaGrad.getGradient(gradient, element1.getIndex(), bias.shape());
    double update2 = w1Bias - biasGradient;
    bias.putScalar(element1.getIndex(), update2);
}

Example 57 with INDArray

use of org.nd4j.linalg.api.ndarray.INDArray in project deeplearning4j by deeplearning4j.

the class DBOW method inferSequence.

/**
     * This method does training on previously unseen paragraph, and returns inferred vector
     *
     * @param sequence
     * @param nextRandom
     * @param learningRate
     * @return
     */
@Override
public INDArray inferSequence(Sequence<T> sequence, long nextRandom, double learningRate, double minLearningRate, int iterations) {
    AtomicLong nr = new AtomicLong(nextRandom);
    if (sequence.isEmpty())
        return null;
    Random random = Nd4j.getRandomFactory().getNewRandomInstance(configuration.getSeed() * sequence.hashCode(), lookupTable.layerSize() + 1);
    INDArray ret = Nd4j.rand(new int[] { 1, lookupTable.layerSize() }, random).subi(0.5).divi(lookupTable.layerSize());
    for (int iter = 0; iter < iterations; iter++) {
        nr.set(Math.abs(nr.get() * 25214903917L + 11));
        dbow(0, sequence, (int) nr.get() % window, nr, learningRate, true, ret);
        learningRate = ((learningRate - minLearningRate) / (iterations - iter)) + minLearningRate;
    }
    finish();
    return ret;
}

Example 58 with INDArray

use of org.nd4j.linalg.api.ndarray.INDArray in project deeplearning4j by deeplearning4j.

the class WordVectorSerializerTest method testWriteWordVectorsFromWord2Vec.

@Test
@Ignore
public void testWriteWordVectorsFromWord2Vec() throws IOException {
    WordVectors vec = WordVectorSerializer.loadGoogleModel(binaryFile, true);
    WordVectorSerializer.writeWordVectors((Word2Vec) vec, pathToWriteto);
    WordVectors wordVectors = WordVectorSerializer.loadTxtVectors(new File(pathToWriteto));
    INDArray wordVector1 = wordVectors.getWordVectorMatrix("Morgan_Freeman");
    INDArray wordVector2 = wordVectors.getWordVectorMatrix("JA_Montalbano");
    assertEquals(vec.getWordVectorMatrix("Morgan_Freeman"), wordVector1);
    assertEquals(vec.getWordVectorMatrix("JA_Montalbano"), wordVector2);
    assertTrue(wordVector1.length() == 300);
    assertTrue(wordVector2.length() == 300);
    assertEquals(wordVector1.getDouble(0), 0.044423, 1e-3);
    assertEquals(wordVector2.getDouble(0), 0.051964, 1e-3);
}

Example 59 with INDArray

use of org.nd4j.linalg.api.ndarray.INDArray in project deeplearning4j by deeplearning4j.

the class Word2VecTests method testUnknown1.

@Test
public void testUnknown1() throws Exception {
    // Strip white space before and after for each line
    SentenceIterator iter = new BasicLineIterator(inputFile.getAbsolutePath());
    // Split on white spaces in the line to get words
    TokenizerFactory t = new DefaultTokenizerFactory();
    t.setTokenPreProcessor(new CommonPreprocessor());
    Word2Vec vec = new Word2Vec.Builder().minWordFrequency(10).useUnknown(true).unknownElement(new VocabWord(1.0, "PEWPEW")).iterations(1).layerSize(100).stopWords(new ArrayList<String>()).seed(42).learningRate(0.025).minLearningRate(0.001).sampling(0).elementsLearningAlgorithm(new CBOW<VocabWord>()).epochs(1).windowSize(5).useHierarchicSoftmax(true).allowParallelTokenization(true).modelUtils(new FlatModelUtils<VocabWord>()).iterate(iter).tokenizerFactory(t).build();
    vec.fit();
    assertTrue(vec.hasWord("PEWPEW"));
    assertTrue(vec.getVocab().containsWord("PEWPEW"));
    INDArray unk = vec.getWordVectorMatrix("PEWPEW");
    assertNotEquals(null, unk);
    File tempFile = File.createTempFile("temp", "file");
    tempFile.deleteOnExit();
    WordVectorSerializer.writeWord2VecModel(vec, tempFile);
    log.info("Original configuration: {}", vec.getConfiguration());
    Word2Vec restored = WordVectorSerializer.readWord2VecModel(tempFile);
    assertTrue(restored.hasWord("PEWPEW"));
    assertTrue(restored.getVocab().containsWord("PEWPEW"));
    INDArray unk_restored = restored.getWordVectorMatrix("PEWPEW");
    assertEquals(unk, unk_restored);
    // now we're getting some junk word
    INDArray random = vec.getWordVectorMatrix("hhsd7d7sdnnmxc_SDsda");
    INDArray randomRestored = restored.getWordVectorMatrix("hhsd7d7sdnnmxc_SDsda");
    log.info("Restored configuration: {}", restored.getConfiguration());
    assertEquals(unk, random);
    assertEquals(unk, randomRestored);
}

Example 60 with INDArray

use of org.nd4j.linalg.api.ndarray.INDArray in project deeplearning4j by deeplearning4j.

the class GloveWeightLookupTable method iterateSample.

/**
     * glove iteration
     * @param w1 the first word
     * @param w2 the second word
     * @param score the weight learned for the particular co occurrences
     */
public double iterateSample(T w1, T w2, double score) {
    INDArray w1Vector = syn0.slice(w1.getIndex());
    INDArray w2Vector = syn0.slice(w2.getIndex());
    //prediction: input + bias
    if (w1.getIndex() < 0 || w1.getIndex() >= syn0.rows())
        throw new IllegalArgumentException("Illegal index for word " + w1.getLabel());
    if (w2.getIndex() < 0 || w2.getIndex() >= syn0.rows())
        throw new IllegalArgumentException("Illegal index for word " + w2.getLabel());
    //w1 * w2 + bias
    double prediction = Nd4j.getBlasWrapper().dot(w1Vector, w2Vector);
    prediction += bias.getDouble(w1.getIndex()) + bias.getDouble(w2.getIndex());
    double weight = Math.pow(Math.min(1.0, (score / maxCount)), xMax);
    double fDiff = score > xMax ? prediction : weight * (prediction - Math.log(score));
    if (Double.isNaN(fDiff))
        fDiff = Nd4j.EPS_THRESHOLD;
    //amount of change
    double gradient = fDiff;
    //note the update step here: the gradient is
    //the gradient of the OPPOSITE word
    //for adagrad we will use the index of the word passed in
    //for the gradient calculation we will use the context vector
    update(w1, w1Vector, w2Vector, gradient);
    update(w2, w2Vector, w1Vector, gradient);
    return fDiff;
}