对于代码有一些疑惑

[heiheihei-ops]

您好，请问以下代码段是什么意思呢？它的作用是什么？

D3Dnet/code/dcn/modules/deform_conv.py

Lines 150 to 167 in f3ca248

	if dimension_T == False:
	for i in range(
	self.deformable_groups * self.kernel_size[0] * self.kernel_size[1] * self.kernel_size[2]):
	offset[:, i * 3, :, :, :] = 0 # T
	offset[:, i * 3 + 1, :, :, :] = temp[:, i * 2, :, :, :]
	offset[:, i * 3 + 2, :, :, :] = temp[:, i * 2 + 1, :, :, :]
	if dimension_H == False:
	for i in range(
	self.deformable_groups * self.kernel_size[0] * self.kernel_size[1] * self.kernel_size[2]):
	offset[:, i * 3, :, :, :] = temp[:, i * 2, :, :, :]
	offset[:, i * 3 + 1, :, :, :] = 0 # H
	offset[:, i * 3 + 2, :, :, :] = temp[:, i * 2 + 1, :, :, :]
	if dimension_W == False:
	for i in range(
	self.deformable_groups * self.kernel_size[0] * self.kernel_size[1] * self.kernel_size[2]):
	offset[:, i * 3, :, :, :] = temp[:, i * 2, :, :, :]
	offset[:, i * 3 + 1, :, :, :] = temp[:, i * 2 + 1, :, :, :]
	offset[:, i * 3 + 2, :, :, :] = 0 # W

[XinyiYing]

这段代码的目的是降低网络的参数量，即当在三维卷积的空域进行变形，只需要生成通道维度为2N的offset（其中N为卷积核的尺寸）。然而，在可变形三维卷积的底层C和cuda代码中，输入的offset的通道维数需要为3N。因此，需要将通道维度为2N的offset补零成通道维度为3N，再进行运算。
This code aims to reduce the number of parameters. That is, when only perform deformation in the spatial domain, only an offset with 2N channels is generated (N is the size of the convolution kernel). However, the basic D3Dnet codes on “C” and “cuda” require that the input offset should have 3N channels. Therefore, we first pad the 2N channels offset to 3N channels offset, and then perform the subsequent operations.

[heiheihei-ops]

谢谢您的答复，请问您这边的dcn的底层C代码是在DCNv2的基础上改动的吗？

[XinyiYing]

是的，代码是基于[DCNv2]进行改进的。

[heiheihei-ops]

好的 谢谢您的答复