正常卷积

也就是我们平常用的比较普遍的卷积：

它的参数量是：112，即：$(卷积核大小）\ast 输入通道\ast 输出通道+偏置$

深度可分离卷积

等于：$深度卷积+逐点卷积$
还是上面的例子，由于输入通道为3，所以我们设置group=3。

深度卷积

它的参数量是：30，即：$(卷积核大小）\ast 输入通道\ast 输出通道\ast group数+偏置$

逐点卷积

就是用卷积核大小为1的卷积核，做正常卷积，此时卷积核个数为输出通道数，即4。

它的参数量是：16，即：$(卷积核大小）\ast 输入通道\ast 输出通道+偏置$

对比

正常卷积参数量：$112$
深度可分离卷积参数量：$30+16=46$
很明显，这么一个简单的卷积，就减少了一半多的参数，所以会很快。

代码实现查看（torch实现）

from torch import nn
import torch
import torchsummary
import torchviz

class Model_1(nn.Module):
    def __init__(self,in_channels, out_channels, groups) 
        super(Model_1, self).__init__()

        self.conv = nn.Conv2d(in_channels=in_channels,out_channels=out_channels,groups=groups,kernel_size=3, padding=1)

    def forward(self, x):
        y = self.conv(x)
        return y

class Model_2(nn.Module):
    def __init__(self, in_channels, out_channels, groups):
        super(Model_2, self).__init__()

        self.conv_1 = nn.Conv2d(in_channels=in_channels,out_channels=in_channels,groups=in_channels,kernel_size=3, padding=1)
        self.conv_2 = nn.Conv2d(in_channels=in_channels,out_channels=out_channels,kernel_size=1)
    
    def forward(self, x):
        x = self.conv_1(x)
        y = self.conv_2(x)
        return y


if __name__ == '__main__':    

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = Model_1(3,4,1).to(device)
    torchsummary.summary(model,input_size=(3,224,224))

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = Model_2(3,4,1).to(device)
    torchsummary.summary(model,input_size=(3,224,224))

结果

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1          [-1, 4, 224, 224]             112
================================================================
Total params: 112
Trainable params: 112
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 1.53
Params size (MB): 0.00
Estimated Total Size (MB): 2.11
----------------------------------------------------------------
----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1          [-1, 3, 224, 224]              30
            Conv2d-2          [-1, 4, 224, 224]              16
================================================================
Total params: 46
Trainable params: 46
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 2.68
Params size (MB): 0.00
Estimated Total Size (MB): 3.25
----------------------------------------------------------------