std::vector lltm_backward()

in cpp/lltm.cpp [50:85]

• 30 lines of code
• 1 McCabe index (conditional complexity)

std::vector<torch::Tensor> lltm_backward(
    torch::Tensor grad_h,
    torch::Tensor grad_cell,
    torch::Tensor new_cell,
    torch::Tensor input_gate,
    torch::Tensor output_gate,
    torch::Tensor candidate_cell,
    torch::Tensor X,
    torch::Tensor gate_weights,
    torch::Tensor weights) {
  auto d_output_gate = torch::tanh(new_cell) * grad_h;
  auto d_tanh_new_cell = output_gate * grad_h;
  auto d_new_cell = d_tanh(new_cell) * d_tanh_new_cell + grad_cell;

  auto d_old_cell = d_new_cell;
  auto d_candidate_cell = input_gate * d_new_cell;
  auto d_input_gate = candidate_cell * d_new_cell;

  auto gates = gate_weights.chunk(3, /*dim=*/1);
  d_input_gate *= d_sigmoid(gates[0]);
  d_output_gate *= d_sigmoid(gates[1]);
  d_candidate_cell *= d_elu(gates[2]);

  auto d_gates =
      torch::cat({d_input_gate, d_output_gate, d_candidate_cell}, /*dim=*/1);

  auto d_weights = d_gates.t().mm(X);
  auto d_bias = d_gates.sum(/*dim=*/0, /*keepdim=*/true);

  auto d_X = d_gates.mm(weights);
  const auto state_size = grad_h.size(1);
  auto d_old_h = d_X.slice(/*dim=*/1, 0, state_size);
  auto d_input = d_X.slice(/*dim=*/1, state_size);

  return {d_old_h, d_input, d_weights, d_bias, d_old_cell};
}