3

我是新手,没有 ML,我已经使用 Deeplearning4j 库进行了训练。我真的迷失在源代码中。我如何阅读具有多个标签的训练集,而不仅仅是 1 个?例如,我不想教 lstm 对 4 个类中的文本进行分类。我如何阅读训练数据集?谢谢

编辑: 这就是我的迭代器代码现在的样子。我有空缺的 POJO 类,它只包含技能 ID 和空缺文本的列表。在每个训练/测试集的每个文件中,有 2 行:一个带有 ids(逗号是分隔符)和文本。所有集合包含 4 个技能,所以 net 的输出等于 5。我训练了 word2vec 模型,所以我的迭代器也使用它。

我使用原始代码示例进行情感分析

我的迭代器:

package SkillsMiner;    
import SkillsMiner.Entities.VacancyLightEntity;
import SkillsMiner.Utils.Reader;
import org.apache.commons.io.FileUtils;
import org.apache.commons.io.FilenameUtils;
import org.deeplearning4j.datasets.iterator.DataSetIterator;
import org.deeplearning4j.models.embeddings.wordvectors.WordVectors;
import org.deeplearning4j.text.tokenization.tokenizer.preprocessor.CommonPreprocessor;
import org.deeplearning4j.text.tokenization.tokenizerfactory.DefaultTokenizerFactory;
import org.deeplearning4j.text.tokenization.tokenizerfactory.TokenizerFactory;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.dataset.DataSet;
import org.nd4j.linalg.dataset.api.DataSetPreProcessor;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.indexing.INDArrayIndex;
import org.nd4j.linalg.indexing.NDArrayIndex;

import java.io.File;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;
import java.util.NoSuchElementException;

/** This is a DataSetIterator that is specialized for the IMDB review dataset used in the Word2VecSentimentRNN example
 * It takes either the train or test set data from this data set, plus a WordVectors object (typically the Google News
 * 300 pretrained vectors from https://code.google.com/p/word2vec/) and generates training data sets.<br>
 * Inputs/features: variable-length time series, where each word (with unknown words removed) is represented by
 * its Word2Vec vector representation.<br>
 * Labels/target: a single class (negative or positive), predicted at the final time step (word) of each review
 *
 * @author Alex Black
 */
public class SentimentExampleIterator implements DataSetIterator {
    private final WordVectors wordVectors;
    private final int batchSize;
    private final int vectorSize;
    private final int truncateLength;

    private int cursor = 0;
    private final File[] filePathes;
    private final TokenizerFactory tokenizerFactory;
    private int labelsCount = 4;

    /**
     * @param dataDirectory the directory of the IMDB review data set
     * @param wordVectors WordVectors object
     * @param batchSize Size of each minibatch for training
     * @param truncateLength If reviews exceed
     * @param train If true: return the training data. If false: return the testing data.
     */
    public SentimentExampleIterator(String dataDirectory, WordVectors wordVectors, int batchSize, int truncateLength, boolean train) throws IOException {
        this.batchSize = batchSize;
        this.vectorSize = wordVectors.lookupTable().layerSize();

        File p = new File(FilenameUtils.concat(dataDirectory, "learning/" + (train ? "train" : "test")) + "/");
        filePathes = p.listFiles();

        this.wordVectors = wordVectors;
        this.truncateLength = truncateLength;

        tokenizerFactory = new DefaultTokenizerFactory();
        tokenizerFactory.setTokenPreProcessor(new CommonPreprocessor());
    }


    @Override
    public DataSet next(int num) {
        if (cursor >= filePathes.length) throw new NoSuchElementException();
        try{
            return nextDataSet(num);
        }catch(IOException e){
            throw new RuntimeException(e);
        }
    }

    private DataSet nextDataSet(int num) throws IOException {
        List<VacancyLightEntity> vacancies = new ArrayList<>(num);
        boolean[] positive = new boolean[num];
        for( int i=0; i<num && cursor<totalExamples(); i++ ){
            String path = filePathes[cursor].getAbsolutePath();
            vacancies.add(Reader.readVacancyFromFile(path));
            cursor++;
        }

        //Second: tokenize vacancies and filter out unknown words
        List<List<String>> allTokens = new ArrayList<>(vacancies.size());
        int maxLength = 0;
        for(VacancyLightEntity v : vacancies){
            List<String> tokens = tokenizerFactory.create(v.getText()).getTokens();
            List<String> tokensFiltered = new ArrayList<>();
            for(String t : tokens ){
                if(wordVectors.hasWord(t)) tokensFiltered.add(t);
            }
            allTokens.add(tokensFiltered);
            maxLength = Math.max(maxLength,tokensFiltered.size());
        }
        //If longest review exceeds 'truncateLength': only take the first 'truncateLength' words
        if(maxLength > truncateLength) maxLength = truncateLength;

        //Create data for training
        //Here: we have vacancies.size() examples of varying lengths
        INDArray features = Nd4j.create(vacancies.size(), vectorSize, maxLength);
        INDArray labels = Nd4j.create(vacancies.size(), labelsCount, maxLength);    //Two labels: positive or negative
        //Because we are dealing with vacancies of different lengths and only one output at the final time step: use padding arrays
        //Mask arrays contain 1 if data is present at that time step for that example, or 0 if data is just padding
        INDArray featuresMask = Nd4j.zeros(vacancies.size(), maxLength);
        INDArray labelsMask = Nd4j.zeros(vacancies.size(), maxLength);

        int[] temp = new int[2];
        for( int i=0; i<vacancies.size(); i++ ){
            List<String> tokens = allTokens.get(i);
            temp[0] = i;
            //Get word vectors for each word in review, and put them in the training data
            for( int j=0; j<tokens.size() && j<maxLength; j++ ){
                String token = tokens.get(j);
                INDArray vector = wordVectors.getWordVectorMatrix(token);
                features.put(new INDArrayIndex[]{NDArrayIndex.point(i), NDArrayIndex.all(), NDArrayIndex.point(j)}, vector);

                temp[1] = j;
                featuresMask.putScalar(temp, 1.0);  //Word is present (not padding) for this example + time step -> 1.0 in features mask
            }

            int idx = (positive[i] ? 0 : 1);
            int lastIdx = Math.min(tokens.size(),maxLength);
            labels.putScalar(new int[]{i,idx,lastIdx-1},1.0);   //Set label: [0,1] for negative, [1,0] for positive
            labelsMask.putScalar(new int[]{i,lastIdx-1},1.0);   //Specify that an output exists at the final time step for this example
        }

        return new DataSet(features,labels,featuresMask,labelsMask);
    }

    @Override
    public int totalExamples() {
        return filePathes.length;
    }

    @Override
    public int inputColumns() {
        return vectorSize;
    }

    @Override
    public int totalOutcomes() {
        return 2;
    }

    @Override
    public void reset() {
        cursor = 0;
    }

    @Override
    public int batch() {
        return batchSize;
    }

    @Override
    public int cursor() {
        return cursor;
    }

    @Override
    public int numExamples() {
        return totalExamples();
    }

    @Override
    public void setPreProcessor(DataSetPreProcessor preProcessor) {
        throw new UnsupportedOperationException();
    }

    @Override
    public List<String> getLabels() {
        return Arrays.asList("positive","negative");
    }

    @Override
    public boolean hasNext() {
        return cursor < numExamples();
    }

    @Override
    public DataSet next() {
        return next(batchSize);
    }

    @Override
    public void remove() {

    }
}
4

0 回答 0