用torchvision库的主干替换网络

import torchvision.models as models
from torchinfo import summary

model = models.mobilenet_v3_small(pretrained=True,progress=True) 
summary(model.features,input_size=(1,3,640,640))

type(model.features)  # sequential支持切片操作

summary(model.features[:4],input_size=(1,3,640,640))# 可以取出对应的下采样特征图

分为三部分：features、avgpool、classifier

结果显示：

  可以定义主干网路类：

class MobilenetV3(nn.Module):
    def __init__(self, slice) :
        super(MobilenetV3, self).__init__()
        self.model = None
        if silce == 1:
            self.model = models.mobilenet_v3_small(pretrained=True).features[:4]
        if silce == 2:
            self.model = models.mobilenet_v3_small(pretrained=True).features[4:9]
        if silce == 3:
            self.model = models.mobilenet_v3_small(pretrained=True).features[9:]
    def forward (self, x):
        return self.model(x)

注册：

elif m is MobilenetV3:
    C2 = args[0]
    args = args[1：]