Group convolution

PyTorch If you want to use block convolution , Only need nn.Conv2d When defining the network structure, specify groups that will do .

Ordinary convolution

Ordinary convolution ： Suppose that channel by $C_{in}$, The convolution kernel size is kxk, Output channel by n, At this time, the number of parameters of ordinary convolution is :$k\ast k\ast C_{in}\ast n$

Group Convolution

Grouping convolution ： Suppose the of the input characteristic matrix channel by $C_{in}$, Of input features channel Divided into g A set of , For each group channel by $C_{in}/g$, Each group corresponds to a convolution . Output per group channel by $n/g$, Finally, g Of each group channel Conduct concat after , Of the output characteristic matrix channel Still for n.

Therefore, the following convolution parameter calculation ：
$(k\ast k\ast c_{in}/g\ast n/g)\ast g$=$k\ast k\ast C_{in}\ast n\ast 1/g$

Contrast is known ：

• use group The number of parameters of convolution is ordinary convolution $1/g$
• Number of groups g=C_{in}, That is, the number of groups follows our input characteristic matrix channel bring into correspondence with , And output channel $n=C_{in}$, At this point DW Convolution