forward method

List<Matrix2d> forward(List<Matrix2d> input)

Applies 2D convolution to multi-channel input

Implementation

List<Matrix2d> forward(List<Matrix2d> input) {
  if (input.length != inChannels) {
    throw ArgumentError(
        "Expected $inChannels input channels, got ${input.length}");
  }

  // Apply padding to each input channel if needed
  List<Matrix2d> paddedInputs =
      input.map((mat) => padding > 0 ? mat.pad(padding) : mat).toList();

  int outputRows = ((paddedInputs[0].rows() - kernelSize) ~/ stride) + 1;
  int outputCols = ((paddedInputs[0].cols() - kernelSize) ~/ stride) + 1;

  // Output feature maps for each output channel
  List<Matrix2d> outputFeatureMaps =
      List.generate(outChannels, (_) => Matrix2d(outputRows, outputCols));

  for (int outIdx = 0; outIdx < outChannels; outIdx++) {
    Matrix2d output = outputFeatureMaps[outIdx];
    Value bias = biases.values[outIdx];

    for (int i = 0; i < outputRows; i++) {
      for (int j = 0; j < outputCols; j++) {
        Value sum = Value(0);

        // Sum over all input channels
        for (int inIdx = 0; inIdx < inChannels; inIdx++) {
          Matrix2d inputChannel = paddedInputs[inIdx];
          Matrix2d kernel = kernels[outIdx][inIdx];

          for (int ki = 0; ki < kernelSize; ki++) {
            for (int kj = 0; kj < kernelSize; kj++) {
              int row = i * stride + ki;
              int col = j * stride + kj;
              sum += inputChannel.at(row, col) * kernel.at(ki, kj);
            }
          }
        }

        // Add bias and store result
        sum += bias;
        output.data!.values[i * outputCols + j] = sum;
      }
    }
  }

  return outputFeatureMaps;
}