()
| 28 | |
| 29 | @torch.no_grad() |
| 30 | def check_forward_equal_with_pytorch_double(): |
| 31 | value = torch.rand(N, S, M, D).cuda() * 0.01 |
| 32 | sampling_locations = torch.rand(N, Lq, M, L, P, 2).cuda() |
| 33 | attention_weights = torch.rand(N, Lq, M, L, P).cuda() + 1e-5 |
| 34 | attention_weights /= attention_weights.sum(-1, |
| 35 | keepdim=True).sum(-2, |
| 36 | keepdim=True) |
| 37 | im2col_step = 2 |
| 38 | output_pytorch = ms_deform_attn_core_pytorch( |
| 39 | value.double(), shapes, sampling_locations.double(), |
| 40 | attention_weights.double()).detach().cpu() |
| 41 | output_cuda = MSDeformAttnFunction.apply(value.double(), shapes, |
| 42 | level_start_index, |
| 43 | sampling_locations.double(), |
| 44 | attention_weights.double(), |
| 45 | im2col_step).detach().cpu() |
| 46 | fwdok = torch.allclose(output_cuda, output_pytorch) |
| 47 | max_abs_err = (output_cuda - output_pytorch).abs().max() |
| 48 | max_rel_err = ((output_cuda - output_pytorch).abs() / |
| 49 | output_pytorch.abs()).max() |
| 50 | |
| 51 | print( |
| 52 | f'* {fwdok} check_forward_equal_with_pytorch_double: max_abs_err {max_abs_err:.2e} max_rel_err {max_rel_err:.2e}' |
| 53 | ) |
| 54 | |
| 55 | |
| 56 | @torch.no_grad() |
no test coverage detected