基于经典网络组装自己的网络架构

笔记来源：【深度学习直播课精选01】深度学习拒绝调包！从0构建自己的神经网络架构哔哩哔哩bilibili

感谢Up主：TsaiTsai

直接使用经典网络可能存在的问题

1. 经典网络结构太深

2. 实际自己的数据样本较小，不适合经典网络

查看Pytorch中包含的经典网络

import torch
import torch.nn as nn
from torchvision import models as m
​
#查看所包含的模型，模型名称大写的为父类，小写的为子类
dir(m)

视频课程中的网络搭建

基于vgg16_bn和resnet18，前者主要使用其卷积层，后者主要使用其残差层

# 实例化两个网络模型
vgg16_bn_ = m.vgg16_bn()
resnet18_ = m.resnet18()

选择vgg16_bn的7-13层

vgg16_bn_.features[7:14]

输出：

Sequential(
  (7): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (8): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (9): ReLU(inplace=True)
  (10): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (11): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (12): ReLU(inplace=True)
  (13): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)

选择resnet18的layer3

resnet18_.layer3

输出：

Sequential(
  (0): BasicBlock(
    (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
    (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (relu): ReLU(inplace=True)
    (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (downsample): Sequential(
      (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
      (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (1): BasicBlock(
    (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (relu): ReLU(inplace=True)
    (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  )
)

网络搭建

# 学习能力越强的部分越放在后面
​
# 输入32*32，通道数为1
​
class MyNet(nn.Module):
    def __init__(self):
        super().__init__()
        # block1
        self.conv1 = nn.Sequential(nn.Conv2d(1, 64, kernel_size=3, stride=1, padding=1),
                                  nn.BatchNorm2d(64),
                                  nn.ReLU(inplace=True))
        # block2+block3
        self.block2 = vgg16_bn_.features[7:14]  # 16*16
        self.block3 = resnet18_.layer3 # 8*8
        self.avgpool = resnet18_.avgpool #所有特征图尺寸缩小为1*1
        self.fc = nn.Linear(in_features=256, out_features=100, bias=True)  #参考resnet18
        
    
    def forward(self, x):
        x = self.conv1(x)
        x = self.block2(x)
        x = self.block3(x)
        x = self.avgpool(x) # (batch_size, 256, 1, 1)
        x = x.view(x.shape[0], 256)
        x = self.fc(x)
        return x

检查模型的正确性

使用库torchinfo · PyPI

from torchinfo import summary
net = MyNet()
summary(net, input_size=(10, 1, 32, 32), depth=3, device="cpu")

输出：