|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
from __future__ import absolute_import |
|
from __future__ import print_function |
|
from __future__ import division |
|
|
|
import time |
|
import torch |
|
import torch.nn as nn |
|
from torch.autograd import gradcheck |
|
|
|
from functions.ms_deform_attn_func import MSDeformAttnFunction, ms_deform_attn_core_pytorch |
|
|
|
|
|
N, M, D = 1, 2, 2 |
|
Lq, L, P = 2, 2, 2 |
|
shapes = torch.as_tensor([(6, 4), (3, 2)], dtype=torch.long).cuda() |
|
level_start_index = torch.cat((shapes.new_zeros((1, )), shapes.prod(1).cumsum(0)[:-1])) |
|
S = sum([(H*W).item() for H, W in shapes]) |
|
|
|
|
|
torch.manual_seed(3) |
|
|
|
|
|
@torch.no_grad() |
|
def check_forward_equal_with_pytorch_double(): |
|
value = torch.rand(N, S, M, D).cuda() * 0.01 |
|
sampling_locations = torch.rand(N, Lq, M, L, P, 2).cuda() |
|
attention_weights = torch.rand(N, Lq, M, L, P).cuda() + 1e-5 |
|
attention_weights /= attention_weights.sum(-1, keepdim=True).sum(-2, keepdim=True) |
|
im2col_step = 2 |
|
output_pytorch = ms_deform_attn_core_pytorch(value.double(), shapes, sampling_locations.double(), attention_weights.double()).detach().cpu() |
|
output_cuda = MSDeformAttnFunction.apply(value.double(), shapes, level_start_index, sampling_locations.double(), attention_weights.double(), im2col_step).detach().cpu() |
|
fwdok = torch.allclose(output_cuda, output_pytorch) |
|
max_abs_err = (output_cuda - output_pytorch).abs().max() |
|
max_rel_err = ((output_cuda - output_pytorch).abs() / output_pytorch.abs()).max() |
|
|
|
print(f'* {fwdok} check_forward_equal_with_pytorch_double: max_abs_err {max_abs_err:.2e} max_rel_err {max_rel_err:.2e}') |
|
|
|
|
|
@torch.no_grad() |
|
def check_forward_equal_with_pytorch_float(): |
|
value = torch.rand(N, S, M, D).cuda() * 0.01 |
|
sampling_locations = torch.rand(N, Lq, M, L, P, 2).cuda() |
|
attention_weights = torch.rand(N, Lq, M, L, P).cuda() + 1e-5 |
|
attention_weights /= attention_weights.sum(-1, keepdim=True).sum(-2, keepdim=True) |
|
im2col_step = 2 |
|
output_pytorch = ms_deform_attn_core_pytorch(value, shapes, sampling_locations, attention_weights).detach().cpu() |
|
output_cuda = MSDeformAttnFunction.apply(value, shapes, level_start_index, sampling_locations, attention_weights, im2col_step).detach().cpu() |
|
fwdok = torch.allclose(output_cuda, output_pytorch, rtol=1e-2, atol=1e-3) |
|
max_abs_err = (output_cuda - output_pytorch).abs().max() |
|
max_rel_err = ((output_cuda - output_pytorch).abs() / output_pytorch.abs()).max() |
|
|
|
print(f'* {fwdok} check_forward_equal_with_pytorch_float: max_abs_err {max_abs_err:.2e} max_rel_err {max_rel_err:.2e}') |
|
|
|
|
|
def check_gradient_numerical(channels=4, grad_value=True, grad_sampling_loc=True, grad_attn_weight=True): |
|
|
|
value = torch.rand(N, S, M, channels).cuda() * 0.01 |
|
sampling_locations = torch.rand(N, Lq, M, L, P, 2).cuda() |
|
attention_weights = torch.rand(N, Lq, M, L, P).cuda() + 1e-5 |
|
attention_weights /= attention_weights.sum(-1, keepdim=True).sum(-2, keepdim=True) |
|
im2col_step = 2 |
|
func = MSDeformAttnFunction.apply |
|
|
|
value.requires_grad = grad_value |
|
sampling_locations.requires_grad = grad_sampling_loc |
|
attention_weights.requires_grad = grad_attn_weight |
|
|
|
gradok = gradcheck(func, (value.double(), shapes, level_start_index, sampling_locations.double(), attention_weights.double(), im2col_step)) |
|
|
|
print(f'* {gradok} check_gradient_numerical(D={channels})') |
|
|
|
|
|
if __name__ == '__main__': |
|
check_forward_equal_with_pytorch_double() |
|
check_forward_equal_with_pytorch_float() |
|
|
|
for channels in [30, 32, 64, 71]: |
|
check_gradient_numerical(channels, True, True, True) |
|
|
