demoNet

demo

class NetworkInit(vocabSize: Int) {
  private val embeddingWidth = DatasetTools.getTomlConfTable("romain").getLong("inputsize").toInt
  private val hiddenSize = DatasetTools.getTomlConfTable("romain").getLong("hiddensize").toInt
  private val numberOfFeats = DatasetTools.getTomlConfTable("romain").getLong("feats").toInt
  private val numberOfClasses = DatasetTools.getTomlConfTable("romain").getLong("classes").toInt

  val config: ComputationGraphConfiguration = new NeuralNetConfiguration.Builder()
    .learningRate(DatasetTools.getTomlConfTable("romain").getDouble("minlr"))
    .graphBuilder()
    .addInputs("wordIndeces")
    .addInputs("features")

    .addLayer("wordVectorizer",
      new EmbeddingLayer.Builder()
        .nIn(vocabSize)
        .nOut(embeddingWidth)
        .build(),
      "wordIndeces")

    .addLayer("linear1",
      new DenseLayer.Builder()
        .nIn(numberOfFeats)
        .nOut(embeddingWidth)
        .build(),
      "features")

    .addVertex("sum", new ElementWiseVertex(ElementWiseVertex.Op.Add), "wordVectorizer", "linear1")

    .addLayer("hidden",
      new GravesLSTM.Builder()
        .activation(Activation.TANH)
        .nIn(hiddenSize)
        .nOut(hiddenSize)
        .build(),
      "sum")

    .addLayer("linear2",
      new DenseLayer.Builder()
        .activation(Activation.SIGMOID) // only if numberOfClasses == 1
        //.activation(Activation.SOFTMAX) // use if numberOfClasses > 1
        .nIn(hiddenSize)
        .nOut(numberOfClasses)
        .build(),
      "hidden")

    .setOutputs("linear2")

  .build()
}