def check_gradient_dpooling():
input = torch.randn(2, 3, 5, 5).cuda() * 0.01
N = 4
batch_inds = torch.randint(2, (N, 1)).cuda().float()
x = torch.rand((N, 1)).cuda().float() * 15
y = torch.rand((N, 1)).cuda().float() * 15
w = torch.rand((N, 1)).cuda().float() * 10
h = torch.rand((N, 1)).cuda().float() * 10
rois = torch.cat((batch_inds, x, y, x + w, y + h), dim=1)
offset = torch.randn(N, 2, 3, 3).cuda()
input.requires_grad = True
offset.requires_grad = True
spatial_scale = 1.0 / 4
pooled_size = 3
output_dim = 3
no_trans = 0
group_size = 1
trans_std = 0.0
sample_per_part = 4
part_size = pooled_size
print(
'check_gradient_dpooling:',
gradcheck(
_DeformRoIPooling,
(input, rois, offset, spatial_scale, pooled_size, output_dim,
no_trans, group_size, part_size, sample_per_part, trans_std),
eps=1e-4))
def check_gradient_dconv():
stride = 1
padding = 1
groups = 2
dilation = 1
im2col_step = 1
input = torch.rand(N, inC, inH, inW).double().cuda()
print('max input:', input.max())
input.requires_grad = True
offset = torch.randn(N, deformable_groups * 2 * kW * kH, inH,
inW).double().cuda() * 2
# offset.data.zero_()
# offset.data -= 0.5
offset.requires_grad = True
weight = torch.randn(outC, int(inC // groups), kH, kW).double().cuda()
weight.requires_grad = True
bias = torch.rand(outC).double().cuda()
bias.requires_grad = True
# print('check_gradient_dconv: ',
# gradcheck(_DeformConv, (input, offset, weight, bias,
# stride, padding, dilation, groups, deformable_groups, im2col_step),
# eps=1e-3, atol=1e-3, rtol=1e-2, raise_exception=True))
print(
'check_gradient_dconv: ',
gradcheck(_DeformConv2d,
(input, offset, weight, bias, stride, padding, dilation,
groups, deformable_groups, im2col_step)))
def check_gradient_conv3d():
input = torch.rand(N, inC, inD, inH, inW).double().cuda() * 0.01
input.requires_grad = True
from torch.nn.functional import conv3d
weight = torch.randn(outC, inC, kD, kH, kW).double().cuda()
weight.requires_grad = True
bias = torch.rand(outC).double().cuda()
bias.requires_grad = True
stride = 1
padding = 1
dilation = 1
# print('check_gradient_conv: ',
# gradcheck(conv2d, (input, weight, bias,
# stride, padding, dilation, deformable_groups),
# eps=1e-3, atol=1e-2, rtol=1e-2, raise_exception=True))
print('check_gradient_conv: ',
gradcheck(conv3d, (input, weight, bias, stride, padding, dilation)))
def check_gradient_mdconv():
stride = 1
padding = 1
groups = 2
dilation = 1
im2col_step = 1
input = torch.rand(N, inC, inH, inW).cuda() * 0.01
input.requires_grad = True
offset = torch.randn(N, deformable_groups * 2 * kW * kH, inH,
inW).cuda() * 2
# offset.data.zero_()
# offset.data -= 0.5
offset.requires_grad = True
mask = torch.rand(N, deformable_groups * 1 * kW * kH, inH, inW).cuda()
# mask.data.zero_()
mask.requires_grad = True
mask = torch.sigmoid(mask)
weight = torch.randn(outC, int(inC // groups), kH, kW).cuda()
weight.requires_grad = True
bias = torch.rand(outC).cuda()
bias.requires_grad = True
print(
'check_gradient_mdconv: ',
gradcheck(_ModulatedDeformConv2d,
(input, offset, mask, weight, bias, stride, padding,
dilation, groups, deformable_groups, im2col_step),
eps=1e-3,
atol=1e-3,
rtol=1e-2,
raise_exception=True))