1

我正在处理一些固定长度的时间序列数据。此数据的格式为 headers #0-->time series trace#0-->header#1-->time series trace #1 -->..->header#P-->time series trace #P。通过使用一些旨在处理此类数据的 API，我终于能够将数据转换为形状为 [P, length of time series] 的 INDArray。如果我的目标是从数据中去除噪音，我应该使用什么作为 DNN？到目前为止，我一直在使用堆叠去噪自动编码器，但不确定我实现它的方式是否正确。

  INDArray sinput=Nd4j.zeros(n1,n2);           //creating a 2D matrix of rows=samples, columns=traceNumbers


    float[] data = new float[n1];           //for the length of n1 samples


    for(int j=0;j<n2;j++){                  //for each trace
        input.read(data);

        for(int k=0;k< n1;k++){
            System.out.println("reading "+k+","+j+" value: "+data[k] );
            sinput.putScalar(new int[]{k,j},data[k]);


        }
    }



    Activation act=Activation.RELU;
    LearningRatePolicy lrPol=LearningRatePolicy.None;
    MultiLayerConfiguration configuration= new NeuralNetConfiguration.Builder()
            .seed(13)
            .iterations(20000)
            .learningRate(0.003)
            .learningRateDecayPolicy(lrPol)
            .lrPolicyDecayRate(1e-5)
            .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
            .l1(1e-5)
            .regularization(true)
            .l2(1e-3)
            .list()

            .layer(0,new AutoEncoder.Builder().nIn(n2).nOut(FIRST_HIDDEN_LAYER_WIDTH).activation(act).corruptionLevel(0.3).weightInit(WeightInit.XAVIER).build())
            .layer(1,new AutoEncoder.Builder().nIn(FIRST_HIDDEN_LAYER_WIDTH).nOut(SECOND_HIDDEN_LAYER_WIDTH).activation(act).weightInit(WeightInit.XAVIER).build())
            .layer(2,new AutoEncoder.Builder().nIn(SECOND_HIDDEN_LAYER_WIDTH).nOut(THIRD_HIDDEN_LAYER_WIDTH).activation(act).weightInit(WeightInit.XAVIER).build())
            .layer(3,new AutoEncoder.Builder().nIn(THIRD_HIDDEN_LAYER_WIDTH).nOut(FOURTH_HIDDEN_LAYER_WIDTH).activation(act).weightInit(WeightInit.XAVIER).build())



            .layer(4,new AutoEncoder.Builder().nIn(FOURTH_HIDDEN_LAYER_WIDTH).nOut(THIRD_HIDDEN_LAYER_WIDTH).activation(act).weightInit(WeightInit.XAVIER).corruptionLevel(0.3).build())
            .layer(5,new AutoEncoder.Builder().nIn(THIRD_HIDDEN_LAYER_WIDTH).nOut(SECOND_HIDDEN_LAYER_WIDTH).activation(act).weightInit(WeightInit.XAVIER).corruptionLevel(0.3).build())
            .layer(6,new AutoEncoder.Builder().nIn(SECOND_HIDDEN_LAYER_WIDTH).nOut(FIRST_HIDDEN_LAYER_WIDTH).activation(act).weightInit(WeightInit.XAVIER).corruptionLevel(0.3).build())
            .layer(7,new AutoEncoder.Builder().nIn(FIRST_HIDDEN_LAYER_WIDTH).nOut(n2).activation(act).weightInit(WeightInit.XAVIER).corruptionLevel(0.3).build())
            //  .layer(1,new AutoEncoder.Builder().nIn(SECOND_HIDDEN_LAYER_WIDTH).nOut(SECOND_HIDDEN_LAYER_WIDTH).activation(Activation.SIGMOID).corruptionLevel(0.3)momentum(0.3)..build())
            .layer(8,new OutputLayer.Builder(LossFunctions.LossFunction.SQUARED_LOSS).activation(Activation.IDENTITY).nIn(n2).nOut(n2).build())
            .pretrain(false).backprop(true)
            .build();








    MultiLayerNetwork network = new MultiLayerNetwork(configuration);
    network.init();

    DataSet dataSet=new DataSet(sinput,sinput);



    network.init();
    network.fit(new DataSet(dataSet.getFeatureMatrix(), dataSet.getFeatureMatrix()));             //training
    //network.fit(seisInput,seisInput);
    System.out.println("Layer 0 has "+network.getLayer(0).numParams()+" and Params: ");
    System.out.println(network.getLayer(0).params());
    System.out.println();
 //DataSet testDataSet=new DataSet(sinput,sinput);
   // DataSetIterator iterator=new INDArrayDataSetIterator()

    INDArray outputFromNet=network.output(testDataSet.getFeatureMatrix());
    INDArray reconstructedFromNet=network.reconstruct(outputFromNet,1);
4

0 回答 0