Java Nd4j.zeros方法代碼示例

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
private static List<TrainingData> readTestData(String fn) {
    int[] shape = { 3, 1 };
    List<TrainingData> trainingDataSet = new ArrayList<>();
    try {
        CSVReader reader = new CSVReader(new FileReader(fn));
        String[] row;
        while ((row = reader.readNext()) != null) {
            int type = Integer.parseInt(row[0]);
            double f1 = Double.parseDouble(row[1]);
            double f2 = Double.parseDouble(row[2]);
            double f3 = Double.parseDouble(row[3]);
            TrainingData trainingData = new TrainingData();
            trainingData.input = Nd4j.create(new double[] { f1, f2, f3 }, shape);
            trainingData.output = Nd4j.zeros(shape);
            trainingData.output.putScalar(type, (double) 1);
            trainingDataSet.add(trainingData);
        }
    } catch (java.io.IOException e) {
    }
    return trainingDataSet;
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
/**
 * Constructor. Fill the matrix with given value.
 */
public NDMatrix(int rows, int cols, double value) {
    if (value == 0.0)
        A = Nd4j.zeros(rows, cols);
    else if (value == 1.0)
        A = Nd4j.ones(rows, cols);
    else
        A = Nd4j.zeros(rows, cols).addi(value);
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
/**
 * Store the mean/std values for width/height of each (populated) omr shape.
 *
 * @throws IOException on IO error
 */
private void storeDims ()
        throws IOException
{
    INDArray dimStats = Nd4j.zeros(4, SHAPE_COUNT);
    logger.info("Symbol dimensions for populated shapes:");

    for (Entry<OmrShape, DistributionStats.Builder> entry : dimMap.entrySet()) {
        OmrShape shape = entry.getKey();
        DistributionStats.Builder builder = entry.getValue();
        DistributionStats stats = builder.build();
        INDArray means = stats.getMean();
        INDArray stds = stats.getStd();
        int index = shape.ordinal();
        double meanWidth = means.getDouble(0);
        double stdWidth = stds.getDouble(0);
        double meanHeight = means.getDouble(1);
        double stdHeight = stds.getDouble(1);
        dimStats.putScalar(new int[]{0, index}, meanWidth);
        dimStats.putScalar(new int[]{1, index}, stdWidth);
        dimStats.putScalar(new int[]{2, index}, meanHeight);
        dimStats.putScalar(new int[]{3, index}, stdHeight);
        logger.info(
                String.format(
                        "%27s width{mean:%.2f std:%.2f} height{mean:%.2f std:%.2f}",
                        shape,
                        meanWidth,
                        stdWidth,
                        meanHeight,
                        stdHeight));
    }

    Nd4j.saveBinary(dimStats, DIMS_PATH.toFile());
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
/** Generate a sample from the network, given an (optional, possibly null) initialization. Initialization
 * can be used to 'prime' the RNN with a sequence you want to extend/continue.<br>
 * Note that the initalization is used for all samples
 * @param initialization String, may be null. If null, select a random character as initialization for all samples
 * @param charactersToSample Number of characters to sample from network (excluding initialization)
 * @param net MultiLayerNetwork with one or more GravesLSTM/RNN layers and a softmax output layer
 * @param iter CharacterIterator. Used for going from indexes back to characters
 */
private static String[] sampleCharactersFromNetwork(String initialization, MultiLayerNetwork net,
                                                    CharacterIterator iter, Random rng, int charactersToSample, int numSamples ){
    //Set up initialization. If no initialization: use a random character
    if( initialization == null ){
        initialization = String.valueOf(iter.getRandomCharacter());
    }

    //Create input for initialization
    INDArray initializationInput = Nd4j.zeros(numSamples, iter.inputColumns(), initialization.length());
    char[] init = initialization.toCharArray();
    for( int i=0; i<init.length; i++ ){
        int idx = iter.convertCharacterToIndex(init[i]);
        for( int j=0; j<numSamples; j++ ){
            initializationInput.putScalar(new int[]{j,idx,i}, 1.0f);
        }
    }

    StringBuilder[] sb = new StringBuilder[numSamples];
    for( int i=0; i<numSamples; i++ ) sb[i] = new StringBuilder(initialization);

    //Sample from network (and feed samples back into input) one character at a time (for all samples)
    //Sampling is done in parallel here
    net.rnnClearPreviousState();
    INDArray output = net.rnnTimeStep(initializationInput);
    output = output.tensorAlongDimension(output.size(2)-1,1,0);	//Gets the last time step output

    for( int i=0; i<charactersToSample; i++ ){
        //Set up next input (single time step) by sampling from previous output
        INDArray nextInput = Nd4j.zeros(numSamples,iter.inputColumns());
        //Output is a probability distribution. Sample from this for each example we want to generate, and add it to the new input
        for( int s=0; s<numSamples; s++ ){
            double[] outputProbDistribution = new double[iter.totalOutcomes()];
            for( int j=0; j<outputProbDistribution.length; j++ ) outputProbDistribution[j] = output.getDouble(s,j);
            int sampledCharacterIdx = sampleFromDistribution(outputProbDistribution,rng);

            nextInput.putScalar(new int[]{s,sampledCharacterIdx}, 1.0f);		//Prepare next time step input
            sb[s].append(iter.convertIndexToCharacter(sampledCharacterIdx));	//Add sampled character to StringBuilder (human readable output)
        }

        output = net.rnnTimeStep(nextInput);	//Do one time step of forward pass
    }

    String[] out = new String[numSamples];
    for( int i=0; i<numSamples; i++ ) out[i] = sb[i].toString();
    return out;
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
/**
     * The main method.
     * @param args No arguments are used.
     */
    public static void main(final String args[]) throws IOException, InterruptedException {

        MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
                .seed(123)
                .iterations(1)
                .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
                .learningRate(0.1)
//                 .useDropConnect(false)
//                .biasInit(0)
                .miniBatch(false)
//                .updater(Updater.SGD)
                .list()
                .layer(0, new DenseLayer.Builder().nIn(2).nOut(2)
//                        .weightInit(WeightInit.DISTRIBUTION).dist(new UniformDistribution(0,1))
                        .activation(Activation.SIGMOID)
                        .build())
                .layer(1, new OutputLayer.Builder(LossFunctions.LossFunction.MSE)
//                        .weightInit(WeightInit.DISTRIBUTION).dist(new UniformDistribution(0,1))
                        .activation(Activation.SIGMOID)
                        .nIn(2).nOut(1).build())
                .pretrain(false).backprop(true).build();


        MultiLayerNetwork model = new MultiLayerNetwork(conf);
        model.init();
        model.setListeners(new ScoreIterationListener(1));  //Print score every 10 parameter updates


        DoublesDataSetIterator iterator = new DoublesDataSetIterator(Arrays.asList(
                makePair(XOR_INPUT[0],XOR_IDEAL[0]),
                makePair(XOR_INPUT[1],XOR_IDEAL[1]),
                makePair(XOR_INPUT[2],XOR_IDEAL[2]),
                makePair(XOR_INPUT[3],XOR_IDEAL[3]))
                ,1);

        for (int n = 0; n < 10000; n++) {
            model.fit(iterator);
        }

        Evaluation eval = model.evaluate(iterator);
        List<Prediction> predictionErrors = eval.getPredictionErrors();
        System.out.println("\n\n+++++ Prediction Errors +++++");
        if (predictionErrors != null) {
            for (Prediction p : predictionErrors) {
                System.out.printf("Predicted class: %d, Actual class: %d\t%s%n", p.getPredictedClass(), p.getActualClass(), p.getRecordMetaData(RecordMetaData.class));
            }
        }
        //Print the evaluation statistics
        System.out.println(eval.stats());

        INDArray data = Nd4j.zeros(2, 2);
        data.putScalar(0,0,0d);
        data.putScalar(0,1,1d);
        data.putScalar(1,0,1d);
        data.putScalar(1,1,1d);
        INDArray output = model.output(data);

        for (int i=0;i<data.rows();i++) {
            System.out.println(data.getRow(i) +" -> "+ output.getRow(i));
        }
    }
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
/**
 * base case: on a leaf hv = pv
 * on a non-leaf node with n children: hv = pv + k centroids of the n hv
 */
private static INDArray getPar2HierVector(WeightLookupTable<VocabWord> lookupTable, PatriciaTrie<String> trie, String node,
                                          int k, Map<String, INDArray> hvs, Method method) {
  if (hvs.containsKey(node)) {
    return hvs.get(node);
  }
  INDArray hv = lookupTable.vector(node);
  String[] split = node.split(REGEX);
  Collection<String> descendants = new HashSet<>();
  if (split.length == 2) {
    String separator = ".";
    String prefix = node.substring(0, node.indexOf(split[1])) + separator;

    SortedMap<String, String> sortedMap = trie.prefixMap(prefix);

    for (Map.Entry<String, String> entry : sortedMap.entrySet()) {
      if (prefix.lastIndexOf(separator) == entry.getKey().lastIndexOf(separator)) {
        descendants.add(entry.getValue());
      }
    }
  } else {
    descendants = Collections.emptyList();
  }
  if (descendants.size() == 0) {
    // just the pv
    hvs.put(node, hv);
    return hv;
  } else {
    INDArray chvs = Nd4j.zeros(descendants.size(), hv.columns());
    int i = 0;
    for (String desc : descendants) {
      // child hierarchical vector
      INDArray chv = getPar2HierVector(lookupTable, trie, desc, k, hvs, method);
      chvs.putRow(i, chv);
      i++;
    }

    double[][] centroids;
    if (chvs.rows() > k) {
      centroids = Par2HierUtils.getTruncatedVT(chvs, k);
    } else if (chvs.rows() == 1) {
      centroids = Par2HierUtils.getDoubles(chvs.getRow(0));
    } else {
      centroids = Par2HierUtils.getTruncatedVT(chvs, 1);
    }
    switch (method) {
      case CLUSTER:
        INDArray matrix = Nd4j.zeros(centroids.length + 1, hv.columns());
        matrix.putRow(0, hv);
        for (int c = 0; c < centroids.length; c++) {
          matrix.putRow(c + 1, Nd4j.create(centroids[c]));
        }
        hv = Nd4j.create(Par2HierUtils.getTruncatedVT(matrix, 1));
        break;
      case SUM:
        for (double[] centroid : centroids) {
          hv.addi(Nd4j.create(centroid));
        }
        break;
    }

    hvs.put(node, hv);
    return hv;
  }
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
private Goal createGoal(
    final TeamColor teamColor,
    final CardinalDirection direction,
    final GeometryFieldSize geometryData
) {
  switch (direction) {
    case NORTH:
      return new Goal.State(
          Nd4j.create(
              new float[]{
                  geometryData.getGoalWidth(),
                  geometryData.getGoalDepth(),
                  ((geometryData.getFieldLength() + geometryData.getGoalDepth()) / 2f),
                  0
              }),
          teamColor,
          direction);
    case SOUTH:
      return new Goal.State(
          Nd4j.create(
              new float[]{
                  geometryData.getGoalWidth(),
                  geometryData.getGoalDepth(),
                  -1f * ((geometryData.getFieldLength() + geometryData.getGoalDepth()) / 2f),
                  0
              }),
          teamColor,
          direction);
    case EAST:
      return new Goal.State(
          Nd4j.create(
              new float[]{
                  geometryData.getGoalWidth(),
                  geometryData.getGoalDepth(),
                  0,
                  ((geometryData.getFieldWidth() + geometryData.getGoalWidth()) / 2f)
              }),
          teamColor,
          direction);
    case WEST:
      return new Goal.State(
          Nd4j.create(
              new float[]{
                  geometryData.getGoalWidth(),
                  geometryData.getGoalDepth(),
                  0,
                  -1f * ((geometryData.getFieldWidth() + geometryData.getGoalWidth()) / 2f)
              }),
          teamColor,
          direction);
    default:
      return new Goal.State(Nd4j.zeros(1, 4), TeamColor.NONE, CardinalDirection.NORTH);
  }
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
@Override
public SeldonMessage aggregate(List<SeldonMessage> outputs, PredictiveUnitState state){
	
	if (outputs.size()==0){
		throw new APIException(APIException.ApiExceptionType.ENGINE_INVALID_COMBINER_RESPONSE, String.format("Combiner received no inputs"));
	}
	
	int[] shape = PredictorUtils.getShape(outputs.get(0).getData());
	
	if (shape == null){
		throw new APIException(APIException.ApiExceptionType.ENGINE_INVALID_COMBINER_RESPONSE, String.format("Combiner cannot extract data shape"));
	}
	
	if (shape.length!=2){
		throw new APIException(APIException.ApiExceptionType.ENGINE_INVALID_COMBINER_RESPONSE, String.format("Combiner received data that is not 2 dimensional"));
	}
	
	INDArray currentSum = Nd4j.zeros(shape[0],shape[1]);
	SeldonMessage.Builder respBuilder = SeldonMessage.newBuilder();
	
	for (Iterator<SeldonMessage> i = outputs.iterator(); i.hasNext();)
	{
		DefaultData inputData = i.next().getData();
		int[] inputShape = PredictorUtils.getShape(inputData);
		if (inputShape == null){
			throw new APIException(APIException.ApiExceptionType.ENGINE_INVALID_COMBINER_RESPONSE, String.format("Combiner cannot extract data shape"));
		}
		if (inputShape.length!=2){
			throw new APIException(APIException.ApiExceptionType.ENGINE_INVALID_COMBINER_RESPONSE, String.format("Combiner received data that is not 2 dimensional"));
		}
		if (inputShape[0] != shape[0]){
			throw new APIException(APIException.ApiExceptionType.ENGINE_INVALID_COMBINER_RESPONSE, String.format("Expected batch length %d but found %d",shape[0],inputShape[0]));
		}
		if (inputShape[1] != shape[1]){
			throw new APIException(APIException.ApiExceptionType.ENGINE_INVALID_COMBINER_RESPONSE, String.format("Expected batch length %d but found %d",shape[1],inputShape[1]));
		}
		INDArray inputArr = PredictorUtils.getINDArray(inputData);
		currentSum = currentSum.add(inputArr);
	}
	currentSum = currentSum.div((float)outputs.size());
	
	DefaultData newData = PredictorUtils.updateData(outputs.get(0).getData(), currentSum);
	respBuilder.setData(newData);
	respBuilder.setMeta(outputs.get(0).getMeta());
	respBuilder.setStatus(outputs.get(0).getStatus());
	
	return respBuilder.build();
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
/** @return {@link INDArray} */
public INDArray zeros () {
	return Nd4j.zeros(1, layerSize);
}
 

import org.nd4j.linalg.factory.Nd4j; //導入方法依賴的package包/類
@Override
public MultiDataSet next(int num) {
    int i = currentBatch * batchSize;
    int currentBatchSize = Math.min(batchSize, corpus.size() - i - 1);
    INDArray input = Nd4j.zeros(currentBatchSize, 1, rowSize);
    INDArray prediction = Nd4j.zeros(currentBatchSize, dictSize, rowSize);
    INDArray decode = Nd4j.zeros(currentBatchSize, dictSize, rowSize);
    INDArray inputMask = Nd4j.zeros(currentBatchSize, rowSize);
    // this mask is also used for the decoder input, the length is the same
    INDArray predictionMask = Nd4j.zeros(currentBatchSize, rowSize);
    for (int j = 0; j < currentBatchSize; j++) {
        List<Double> rowIn = new ArrayList<>(corpus.get(i));
        Collections.reverse(rowIn);
        List<Double> rowPred = new ArrayList<>(corpus.get(i + 1));
        rowPred.add(1.0); // add <eos> token
        // replace the entire row in "input" using NDArrayIndex, it's faster than putScalar(); input is NOT made of
        // one-hot vectors because of the embedding layer that accepts token indexes directly
        input.put(new INDArrayIndex[] { NDArrayIndex.point(j), NDArrayIndex.point(0), NDArrayIndex.interval(0, rowIn.size()) },
                Nd4j.create(ArrayUtils.toPrimitive(rowIn.toArray(new Double[0]))));
        inputMask.put(new INDArrayIndex[] { NDArrayIndex.point(j), NDArrayIndex.interval(0, rowIn.size()) },
                Nd4j.ones(rowIn.size()));
        predictionMask.put(new INDArrayIndex[] { NDArrayIndex.point(j), NDArrayIndex.interval(0, rowPred.size()) },
                Nd4j.ones(rowPred.size()));
        // prediction (output) and decode ARE one-hots though, I couldn't add an embedding layer on top of the decoder
        // and I'm not sure it's a good idea either
        double predOneHot[][] = new double[dictSize][rowPred.size()];
        double decodeOneHot[][] = new double[dictSize][rowPred.size()];
        decodeOneHot[2][0] = 1; // <go> token
        int predIdx = 0;
        for (Double pred : rowPred) {
            predOneHot[pred.intValue()][predIdx] = 1;
            // put the same vals to decode with +1 offset except the last token that is <eos>
            if (predIdx < rowPred.size() - 1) { decodeOneHot[pred.intValue()][predIdx + 1] = 1; }
            ++predIdx;
        }
        prediction.put(new INDArrayIndex[] { NDArrayIndex.point(j), NDArrayIndex.interval(0, dictSize),
                NDArrayIndex.interval(0, rowPred.size()) }, Nd4j.create(predOneHot));
        decode.put(new INDArrayIndex[] { NDArrayIndex.point(j), NDArrayIndex.interval(0, dictSize),
                NDArrayIndex.interval(0, rowPred.size()) }, Nd4j.create(decodeOneHot));
        ++i;
    }
    ++currentBatch;
    return new org.nd4j.linalg.dataset.MultiDataSet(new INDArray[] { input, decode }, new INDArray[] { prediction },
            new INDArray[] { inputMask, predictionMask }, new INDArray[] { predictionMask });
}