Pytorch实现各种2d卷积示例

普通卷积

使用nn.Conv2d()，一般还会接上BN和ReLu

参数量NNCin*Cout+Cout(如果有bias，相对来说表示对参数量影响很小，所以后面不考虑)

class ConvBNReLU(nn.Module):

def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
 super(ConvBNReLU, self).__init__()
 self.op = nn.Sequential(
  nn.Conv2d(C_in, C_out, kernel_size, stride=stride, padding=padding, bias=False),
  nn.BatchNorm2d(C_out, eps=1e-3, affine=affine),
  nn.ReLU(inplace=False)
 )

def forward(self, x):
 return self.op(x)

深度可分离卷积depthwise separable convolution

卷积操作可以分为NN 的Depthwise卷积（不改变通道数）和11的Pointwise卷积（改变为输出通道数），同样后接BN，ReLU。参数量明显减少

参数量：

NNCin+Cin11*Cout

class SepConv(nn.Module):

def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
 super(SepConv, self).__init__()
 self.op = nn.Sequential(
  nn.ReLU(inplace=False),
  nn.Conv2d(C_in, C_in, kernel_size=kernel_size, stride=stride, padding=padding, groups=C_in, bias=False),
  nn.Conv2d(C_in, C_out, kernel_size=1, padding=0, bias=False),
  nn.BatchNorm2d(C_out, eps=1e-3, affine=affine)
  )
def forward(self, x):
 return self.op(x)

空洞卷积dilated convolution

空洞卷积（dilated convolution）是针对图像语义分割问题中下采样会降低图像分辨率、丢失信息而提出的一种卷积思路。利用添加空洞扩大感受野。

参数量不变，但感受野增大（可结合深度可分离卷积实现）

class DilConv(nn.Module):

def __init__(self, C_in, C_out, kernel_size, stride, padding, dilation, affine=True):
 super(DilConv, self).__init__()
 self.op = nn.Sequential(
  nn.ReLU(inplace=False),
  nn.Conv2d(C_in, C_in, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=C_in, bias=False),
  nn.Conv2d(C_in, C_out, kernel_size=1, padding=0, bias=False),
  nn.BatchNorm2d(C_out, eps=1e-3, affine=affine),
  )

def forward(self, x):
 return self.op(x)

Identity

这个其实不算卷积操作，但是在实现跨层传递捷径

class Identity(nn.Module):

def __init__(self):
 super(Identity, self).__init__()

def forward(self, x):
 return x

来源：https://blog.csdn.net/u014767662/article/details/88574643