org.deeplearning4j.models.embeddings.learning.impl.sequence.DBOW Maven / Gradle / Ivy

/*******************************************************************************
 * Copyright (c) 2015-2018 Skymind, Inc.
 *
 * This program and the accompanying materials are made available under the
 * terms of the Apache License, Version 2.0 which is available at
 * https://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.
 *
 * SPDX-License-Identifier: Apache-2.0
 ******************************************************************************/

package org.deeplearning4j.models.embeddings.learning.impl.sequence;

import lombok.NonNull;
import org.deeplearning4j.models.embeddings.WeightLookupTable;
import org.deeplearning4j.models.embeddings.learning.ElementsLearningAlgorithm;
import org.deeplearning4j.models.embeddings.learning.SequenceLearningAlgorithm;
import org.deeplearning4j.models.embeddings.learning.impl.elements.BatchSequences;
import org.deeplearning4j.models.embeddings.learning.impl.elements.SkipGram;
import org.deeplearning4j.models.embeddings.loader.VectorsConfiguration;
import org.deeplearning4j.models.sequencevectors.interfaces.SequenceIterator;
import org.deeplearning4j.models.sequencevectors.sequence.Sequence;
import org.deeplearning4j.models.sequencevectors.sequence.SequenceElement;
import org.deeplearning4j.models.word2vec.wordstore.VocabCache;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.rng.Random;
import org.nd4j.linalg.factory.Nd4j;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import java.util.ArrayList;
import java.util.List;
import java.util.concurrent.atomic.AtomicLong;

/**
 * @author [email protected]
 */
public class DBOW implements SequenceLearningAlgorithm {
    protected VocabCache vocabCache;
    protected WeightLookupTable lookupTable;
    protected VectorsConfiguration configuration;


    protected int window;
    protected boolean useAdaGrad;
    protected double negative;

    protected SkipGram skipGram = new SkipGram<>();

    private static final Logger log = LoggerFactory.getLogger(DBOW.class);

    @Override
    public ElementsLearningAlgorithm getElementsLearningAlgorithm() {
        return skipGram;
    }

    public DBOW() {

    }

    @Override
    public String getCodeName() {
        return "PV-DBOW";
    }

    @Override
    public void configure(@NonNull VocabCache vocabCache, @NonNull WeightLookupTable lookupTable,
                    @NonNull VectorsConfiguration configuration) {
        this.vocabCache = vocabCache;
        this.lookupTable = lookupTable;

        this.window = configuration.getWindow();
        this.useAdaGrad = configuration.isUseAdaGrad();
        this.negative = configuration.getNegative();
        this.configuration = configuration;

        skipGram.configure(vocabCache, lookupTable, configuration);
    }

    /**
     * DBOW doesn't involves any pretraining
     *
     * @param iterator
     */
    @Override
    public void pretrain(SequenceIterator iterator) {

    }

    @Override
    public double learnSequence(@NonNull Sequence sequence, @NonNull AtomicLong nextRandom, double learningRate,
                                BatchSequences batchSequences) {

        // we just pass data to dbow, and loop over sequence there
        dbow(0, sequence, (int) nextRandom.get() % window, nextRandom, learningRate, false, null,
                batchSequences);


        return 0;
    }

    /**
     * DBOW has no reasons for early termination
     * @return
     */
    @Override
    public boolean isEarlyTerminationHit() {
        return false;
    }

    protected void dbow(int i, Sequence sequence, int b, AtomicLong nextRandom, double alpha, boolean isInference,
                    INDArray inferenceVector, BatchSequences batchSequences) {

        //final T word = sequence.getElements().get(i);
        List sentence = skipGram.applySubsampling(sequence, nextRandom).getElements();


        if (sequence.getSequenceLabel() == null)
            return;

        List labels = new ArrayList<>();
        labels.addAll(sequence.getSequenceLabels());

        if (sentence.isEmpty() || labels.isEmpty())
            return;

        int batchSize = configuration.getBatchSize();

        for (T lastWord : labels) {
            for (T word : sentence) {
                if (word == null)
                    continue;

                if (batchSize == 1 || batchSequences == null || isInference)
                    skipGram.iterateSample(word, lastWord, nextRandom, alpha, isInference, inferenceVector);
                else
                    batchSequences.put(word, lastWord, nextRandom.get(), alpha);
            }
        }

        if (skipGram != null && skipGram.getBatch() != null && skipGram.getBatch() != null
                        && skipGram.getBatch().size() >= configuration.getBatchSize()) {
            Nd4j.getExecutioner().exec(skipGram.getBatch());
            skipGram.getBatch().clear();
        }
    }

    /**
     * This method does training on previously unseen paragraph, and returns inferred vector
     *
     * @param sequence
     * @param nextRandom
     * @param learningRate
     * @return
     */
    @Override
    public INDArray inferSequence(Sequence sequence, long nextRandom, double learningRate, double minLearningRate,
                    int iterations) {
        AtomicLong nr = new AtomicLong(nextRandom);

        // we probably don't want subsampling here
        // Sequence seq = cbow.applySubsampling(sequence, nextRandom);
        // if (sequence.getSequenceLabel() == null) throw new IllegalStateException("Label is NULL");

        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, null);

            learningRate = ((learningRate - minLearningRate) / (iterations - iter)) + minLearningRate;
        }

        finish();

        return ret;
    }

    @Override
    public void finish() {
        if (skipGram != null && skipGram.getBatch() != null && !skipGram.getBatch().isEmpty()) {
            Nd4j.getExecutioner().exec(skipGram.getBatch());
            skipGram.getBatch().clear();
        }
    }
}