Group convolution

0、 Preface

First, let's look at ordinary convolution , use pytorch Defined as ：

nn.Conv2d(in_channels=3,
          out_channels=4,
          kernel_size=3)

It can be abstractly expressed as , A stack is a convolution kernel , Corresponding to the above figure ：

1、 Group convolution

Grouping convolution is well-known , I don't know much about , First of all, from the The figure below Let's look at the difference between ordinary convolution and block convolution （ The left figure is a general convolution , On the right is block convolution ）

No matter what kind of convolution , Input characteristic graphs are 12 passageway , Input channel diagrams are 6 passageway , But the intermediate operation process is different .

Each of the small stacked blocks in the figure is a convolution kernel , For example, ordinary convolution , Input is 12 passageway , So the size of a convolution kernel is (12,k_size,k_size), Such a convolution kernel operates with input to obtain a characteristic graph . What we want is 6 Output channel , So there are 6 A convolution kernel like this , The overall size is (6,12,k_size,k_size).

In fact, no matter ordinary convolution or grouping convolution , The number of convolution kernels does not change , But the size of the convolution kernel of the block convolution becomes smaller .

It can be considered that block convolution is to cut off each convolution kernel of ordinary convolution , Cut into small convolution kernel .

For example, in deep separable convolution depthwise Convolution , It's the number of groups = Input channel = Packet convolution of output channels , The original (3,3,3) Convolution kernel of (1,3,3), That is, put each 3 The convolution kernel of the channel is cut into 1 Convolution kernel of channel , But the number of convolution kernels is still 3 individual .

2、Pytorch The implementation of the

pytorch in conv2 It's provided by group Parameters , It should be noted that , It can be seen from the picture above , Block convolution is to divide the input characteristic graph and output characteristic graph n Group , That is, the number of output channels and the number of input channels can be divided into n Group , Otherwise, an error will be reported .

For the above grouping, we use pytorch Realization ：

nn.Conv2d(in_channels=12,
          out_channels=6,
          kernel_size=3,
          groups=3)

For the picture above （depthwise Convolution ）, We use pytorch Realization ：

nn.Conv2d(in_channels=3,
          out_channels=3,
          kernel_size=3,
          groups=3)

actually depthwise Convolution is the number of groups = Enter the number of channels = Special grouping convolution of the number of output channels

Reference resources

Grouping convolution Group Converlution
Depthwise Convolution and Pointwise Convolution