Dear friend,
I check function dcn_v2_cpu_forward, think the following code is wrong:
auto ones_T = at::transpose(ones.contiguous(), 2, 0);
ones_T = at::mul(ones_T, bias.contiguous());
ones_T = at::transpose(ones_T, 2, 0);
output_n = at::add(output_n, ones_T);
Because: ones is 2 dimension data with auto ones = at::ones({ height_out, width_out }, input.options()); So cpu code does not work.
Would you please kindly check it ?
Dear friend, I check function dcn_v2_cpu_forward, think the following code is wrong: auto ones_T = at::transpose(ones.contiguous(), 2, 0); ones_T = at::mul(ones_T, bias.contiguous()); ones_T = at::transpose(ones_T, 2, 0); output_n = at::add(output_n, ones_T); Because: ones is 2 dimension data with auto ones = at::ones({ height_out, width_out }, input.options()); So cpu code does not work. Would you please kindly check it ?