build method

@override void build(Tensor input)

Initializes the W_xh, W_hh, and b_h parameter tensors.

This method infers the inputSize from the input sequence and creates the weight matrices with the correct shapes. It uses Xavier/Glorot initialization, which is a good practice for layers with Tanh activations.

Implementation

@override
void build(Tensor<dynamic> input) {
  Matrix inputMatrix = input.value as Matrix;
  int inputSize = inputMatrix.isNotEmpty ? inputMatrix[0].length : 0;
  Random random = Random();

  double xavierStdDev(int fanIn, int fanOut) => sqrt(2.0 / (fanIn + fanOut));

  double inputToHiddenStdDev = xavierStdDev(inputSize, hiddenSize);
  Matrix wXhValues = [];
  for (int i = 0; i < hiddenSize; i++) {
    Vector row = [];
    for (int j = 0; j < inputSize; j++) {
      row.add((random.nextDouble() * 2 - 1) * inputToHiddenStdDev);
    }
    wXhValues.add(row);
  }

  double hiddenToHiddenStdDev = xavierStdDev(hiddenSize, hiddenSize);
  Matrix wHhValues = [];
  for (int i = 0; i < hiddenSize; i++) {
    Vector row = [];
    for (int j = 0; j < hiddenSize; j++) {
      row.add((random.nextDouble() * 2 - 1) * hiddenToHiddenStdDev);
    }
    wHhValues.add(row);
  }

  W_xh = Tensor<Matrix>(wXhValues);
  W_hh = Tensor<Matrix>(wHhValues);
  b_h = Tensor<Vector>(List<double>.filled(hiddenSize, 0.0));

  super.build(input);
}