深度可分离卷积vs标准卷积

本文只比较深度可分离卷积与标准卷积的参数量、运算量、时间，模型性能请看Mobilenet系列论文。

1. Model

• 深度可分离卷积

import torch
import torch.nn as nn
import torch.nn.functional as F
import time
from torchstat import stat


class Depthpoint_conv(nn.Module):
    def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
        super(Depthpoint_conv, self).__init__()
        self.op = nn.Sequential(
            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, affine=affine),
            nn.LeakyReLU()
        )

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


class Net1(nn.Module):
    def __init__(self):
        super(Net1, self).__init__()
        self.op = nn.Sequential(
            Depthpoint_conv(3, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1),
            Depthpoint_conv(128, 128, 3, 1, 1)
        )

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


model_1 = Net1()
model_1.eval()
x = torch.randn((1, 3, 256, 256)).cuda()
stat(model_1, (3, 256, 256))

• 标准卷积

import torch
import torch.nn as nn
import torch.nn.functional as F
import time
from torchstat import stat


class Conv2d_block(nn.Module):
    def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
        super(Conv2d_block, self).__init__()
        self.op = nn.Sequential(
            nn.Conv2d(C_in, C_out, kernel_size=kernel_size, stride=stride, padding=padding, bias=False),
            nn.BatchNorm2d(C_out, affine=affine),
            nn.LeakyReLU()
        )

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


class Net2(nn.Module):
    def __init__(self):
        super(Net2, self).__init__()
        self.op = nn.Sequential(
            Conv2d_block(3, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1),
            Conv2d_block(128, 128, 3, 1, 1)
        )

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


model_2 = Net2()
model_2.eval()
x = torch.randn((1, 3, 256, 256)).cuda()
stat(model_2, (3, 256, 256))

2. 结果

避坑指南：不能简单地用深度可分离卷积代替标准卷积，模型性能会变得很差