Equivalence of slicing in ATen C++

In Python given a 2-D tensor, we can use tensor[:,:2] to slice the a 2x2 matrix of the first two elements in the top left of the matrix, e.g. :

x = torch.tensor([[-1.4673,  0.9980, -2.1427, -1.1798, -0.0646, -0.2635, -2.8930, -0.2563,
          0.4559, -0.7947, -0.4540,  3.3224,  0.2295,  5.5568, -8.0451, -2.4529,
          4.8724,  2.1640,  3.3255,  0.6693, -1.2362,  4.4713, -3.5547, -0.0528,
          0.1031, -1.2472, -1.6014,  1.8134],
        [ 2.1636, -1.1497, -5.0298,  2.8261, -0.5684,  0.6389,  2.9009, -5.1609,
          1.7358, -3.1819, -0.9877,  5.5598,  6.7142,  4.5704, -1.2683, -5.3046,
          3.0454,  3.2757, -3.2541,  3.6619, -3.6391, -0.2002,  5.7175,  5.7130,
          0.6632, -0.0744, -0.3502,  4.8116]])

y, z = x[:,:2].chunk(2,1)

print(y)

print(z)

[out]:

tensor([[-1.4673],
        [ 2.1636]])
tensor([[ 0.9980],
        [-1.1497]])

What is right way to do it in C++ for PyTorch’s ATen particularly?

For e.g. in the LSTM, there is the gate.chunk(4,1) function at https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/RNN.cpp#L253

If I want to do a gate[:,:2].chunk(2,1) to extract different parts of the gates, e.g. auto partial_gates = gates[:,:2].chunk(4, 1);, how can it be done?

	template <typename cell_params>
struct LSTMCell : Cell<std::tuple<Tensor, Tensor>, cell_params> {
  using hidden_type = std::tuple<Tensor, Tensor>;
  hidden_type operator()(const Tensor& input, const hidden_type& hidden, const cell_params& params) const override {
    auto hx = std::get<0>(hidden);
    auto cx = std::get<1>(hidden);

    if (input.is_cuda()) {
      auto igates = params.matmul_ih(input);
      auto hgates = params.matmul_hh(hx);
      auto result = at::_thnn_fused_lstm_cell(igates, hgates, cx, params.b_ih, params.b_hh);
      // Slice off the workspace argument (it's needed only for AD).
      return std::make_tuple(std::get<0>(result), std::get<1>(result));
    }

    auto gates = params.linear_ih(input) + params.linear_hh(hx);
    auto chunked_gates = gates.chunk(4, 1);

    auto partial_gates = gates[:,:2].chunk(4, 1);

    auto ingate = chunked_gates[0].sigmoid();
    auto forgetgate = chunked_gates[1].sigmoid();
    auto cellgate = chunked_gates[2].tanh();
    auto outgate = chunked_gates[3].sigmoid();

    auto cy = (forgetgate * cx) + (ingate * cellgate);
    auto hy = outgate * cy.tanh();

    return std::make_tuple(hy, cy);
  }
};
1 Like

Question also asked on https://stackoverflow.com/questions/56698047/equivalence-of-slicing-tensor-in-aten-c

The functionnal way to do [:,:2] is .narrow(1, 0, 2). This function should be available in cpp.
Note that selecting a single element is done with the .select() function.

1 Like

Does the .narrow(1, 0, 2) work for CUDA tensors too? Or just C++ CPU tensors?

Do we need to do anything special for CUDA tensors?

All functions (should) work the same way for cuda and cpu tensors.

Thanks Alban!! It works well =)