def check_pooling_zero_offset():
input = torch.randn(2, 16, 64, 64).cuda().zero_()
input[0, :, 16:26, 16:26] = 1.
input[1, :, 10:20, 20:30] = 2.
rois = torch.tensor([
[0, 65, 65, 103, 103],
[1, 81, 41, 119, 79],
]).cuda().float()
pooling = DCNv2Pooling(spatial_scale=1.0 / 4,
pooled_size=7,
output_dim=16,
no_trans=True,
group_size=1,
trans_std=0.0).cuda()
out = pooling(input, rois, input.new())
s = ', '.join(
['%f' % out[i, :, :, :].mean().item() for i in range(rois.shape[0])])
print(s)
dpooling = DCNv2Pooling(spatial_scale=1.0 / 4,
pooled_size=7,
output_dim=16,
no_trans=False,
group_size=1,
trans_std=0.0).cuda()
offset = torch.randn(20, 2, 7, 7).cuda().zero_()
dout = dpooling(input, rois, offset)
s = ', '.join(
['%f' % dout[i, :, :, :].mean().item() for i in range(rois.shape[0])])
print(s)
def example_dpooling():
from dcn_v2 import DCNv2Pooling
input = torch.randn(2, 32, 64, 64).cuda()
batch_inds = torch.randint(2, (20, 1)).cuda().float()
x = torch.randint(256, (20, 1)).cuda().float()
y = torch.randint(256, (20, 1)).cuda().float()
w = torch.randint(64, (20, 1)).cuda().float()
h = torch.randint(64, (20, 1)).cuda().float()
rois = torch.cat((batch_inds, x, y, x + w, y + h), dim=1)
offset = torch.randn(20, 2, 7, 7).cuda()
# normal roi_align
pooling = DCNv2Pooling(spatial_scale=1.0 / 4,
pooled_size=7,
output_dim=32,
no_trans=True,
group_size=1,
trans_std=0.1).cuda()
# deformable pooling
dpooling = DCNv2Pooling(spatial_scale=1.0 / 4,
pooled_size=7,
output_dim=32,
no_trans=False,
group_size=1,
trans_std=0.1).cuda()
out = pooling(input, rois, offset)
dout = dpooling(input, rois, offset)
print(out.shape)
print(dout.shape)
def example_dpooling():
input = torch.randn(2, 32, 64, 64).cuda()
batch_inds = torch.randint(2, (20, 1)).cuda().float()
x = torch.randint(256, (20, 1)).cuda().float()
y = torch.randint(256, (20, 1)).cuda().float()
w = torch.randint(64, (20, 1)).cuda().float()
h = torch.randint(64, (20, 1)).cuda().float()
rois = torch.cat((batch_inds, x, y, x + w, y + h), dim=1)
offset = torch.randn(20, 2, 7, 7).cuda()
input.requires_grad = True
offset.requires_grad = True
# normal roi_align
pooling = DCNv2Pooling(
spatial_scale=1.0 / 4, pooled_size=7, output_dim=32, no_trans=True, group_size=1, trans_std=0.1
).cuda()
# deformable pooling
dpooling = DCNv2Pooling(
spatial_scale=1.0 / 4, pooled_size=7, output_dim=32, no_trans=False, group_size=1, trans_std=0.1
).cuda()
out = pooling(input, rois, offset)
dout = dpooling(input, rois, offset)
print(out.shape)
print(dout.shape)
target_out = out.new(*out.size())
target_out.data.uniform_(-0.01, 0.01)
target_dout = dout.new(*dout.size())
target_dout.data.uniform_(-0.01, 0.01)
e = (target_out - out).mean()
e.backward()
e = (target_dout - dout).mean()
e.backward()
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()
dpooling = DCNv2Pooling(spatial_scale=1.0 / 4,
pooled_size=3,
output_dim=3,
no_trans=False,
group_size=1,
trans_std=0.0).cuda()
input.requires_grad = True
offset.requires_grad = True
print('check_gradient_dpooling', gradcheck(dpooling, (input, rois, offset), eps=1e-4))