1 groups = 1（默认值）

就如同普通的卷积

import torch
import torch.nn as nn
N,C_in,H,W,C_out =  10,4,16,16,6
x = torch.randn(N,C_in,H,W).float()
conv = nn.Conv2d(C_in,C_out,kernel_size=3,stride=3,padding=1)
y = conv(x)
print(y.size())
'''torch.Size([10, 6, 6, 6])'''

2 groups = 2 分组卷积

简单翻译一下，当groups = 2时，相当于有两个卷积层，每一层的输入为原来输入通道数的一半，输出也为原来输出的一半，然后再连接：

import torch
import torch.nn as nn
N,C_in,H,W,C_out =  10,4,16,16,6
x = torch.randn(N,C_in,H,W).float()
conv = nn.Conv2d(C_in,C_out,kernel_size=3,stride=3,padding=1)
conv_group = nn.Conv2d(C_in,C_out,kernel_size=3,stride=3,padding=1,groups = 2)
y = conv(x)
y_group = conv_group(x)
print(y.size())
print(y_group.size())
'''torch.Size([10, 6, 6, 6])'''
'''torch.Size([10, 6, 6, 6])'''

3 深度卷积depthwise

此时group=in_channels，也就是对每一个输入的通道分别用不同的卷积核卷积。out_channels必须是in_channels的整数倍，即
$\frac{out\_channels} {in\_channels} = k$

当k = 1时，out_channels = in_channels,每一个卷积核分别和每一个输入的通道进行卷积，最后在concat起来。如图

[2].png

例如：

import torch
import torch.nn as nn
N,C_in,H,W,C_out =  10,4,16,16,4
x = torch.randn(N,C_in,H,W).float()
conv = nn.Conv2d(C_in,C_out,kernel_size=3,stride=3,padding=1,bias = False)
conv_group = nn.Conv2d(C_in,C_out,kernel_size=3,stride=3,padding=1,groups = C_in,bias=False)
y = conv(x)
y_group = conv_group(x)
conv_1x1 = nn.Conv2d(C_in,C_out,kernel_size=1)
print("groups=1时参数大小：%d"%sum(param.numel() for param in conv.parameters()))
print("groups=in_channels时参数大小：%d"%sum(param.numel() for param in conv_group.parameters()))
'''
groups=1时参数大小：144
groups=in_channels时参数大小：36
'''

在上面的例子中，in_channels = out_channels = 4,kernal_size = 3
当groups=1时，参数大小为 $out\_channels*kernal\_size*kernal\_size*in\_channels$ 即4*3*3*4 = 144
当groups = in_channels时，参数大小为
$out\_channels*kernal\_size*kernal\_size$
即4*3*3 = 36