恰当新手搭建ResNet50残差网络的架构图（最全）

恰当新手搭建ResNet50残差网络的架构图+代码（最全）
网上的教程大多复杂难明，不恰当新手，本来神经网络就难，这些教程本身更难，对新手极度不友好，因此本身做的这个架构图和写的代码，面向新手，大神跳过

1. from torch import nn
2. import torch
3. from torchviz import make_dot
5. class box(nn.Module):
6.     def __init__(self, in_channels, index=999, stride=1, downsample=False):
7.         super(box, self).__init__()
9.         last_stride = 2  # 虚残差中卷积核的步距
11.         if downsample:  # 虚残差结构
12.             f_out_channnels = in_channels * 2
13.             out_channels = int(in_channels / 2)
14.             if index == 0:  # here is first core
15.                 in_channels = int(in_channels / 2)  # 第一层设置为128，是方便了后面的统一处理
16.                 out_channels = in_channels
17.                 f_out_channnels = in_channels * 4
18.                 last_stride = 1
20.                 stride = 1
22.         else:  # 实残差
23.             f_out_channnels = in_channels * 1
24.             out_channels = int(in_channels / 4)
28.         self.downsample = downsample
29.         self.conv1 = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0, bias=False)
30.         self.relu = nn.ReLU(inplace=True)
31.         self.bn1 = nn.BatchNorm2d(out_channels)
33.         self.conv2 = nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=1, bias=False)
34.         self.relu = nn.ReLU(inplace=True)
35.         self.bn2 = nn.BatchNorm2d(out_channels)
37.         self.conv3 = nn.Conv2d(in_channels=out_channels, out_channels=f_out_channnels, kernel_size=1, stride=1, padding=0)
38.         self.bn3 = nn.BatchNorm2d(f_out_channnels)
40.         self.fe = nn.Sequential(
41.             nn.Conv2d(in_channels=in_channels, out_channels=f_out_channnels, kernel_size=1, stride=last_stride,  padding=0, bias=False),
42.             nn.BatchNorm2d(f_out_channnels),
43.         )
47.     def forward(self, x):
48.         identity = x
49.         if self.downsample:
50.             identity = self.fe(x)
51.         x = self.conv1(x)
52.         x = self.bn1(x)
53.         x = self.relu(x)
55.         x = self.conv2(x)
56.         x = self.bn2(x)
57.         x = self.relu(x)
59.         x = self.conv3(x)
60.         x = self.bn3(x)
62.         out = x + identity
63.         out = self.relu(out)
64.         return out
67. class New50(nn.Module):
68.     def __init__(self,in_out, num_classes=4):
69.         super(New50, self).__init__()
72.         self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2,padding=3, bias=False)
73.         self.bn1 = nn.BatchNorm2d(64)
74.         self.relu = nn.ReLU(inplace=True)
76.         self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
78.         self.avgpool = nn.AdaptiveAvgPool2d((1, 1))  # output size = (1, 1)
79.         self.fc = nn.Linear(512 * 4 , num_classes)
81.         layers = []
82.         for index, z in enumerate(in_out):
83.             in_ch = z[0]                                      # 这里通道/2
84.             layers.append(box(in_channels=in_ch, stride=2, downsample=z[2], index=index))  # 这里处理第一层
86.             for i in range(1, z[1]):
87.                 layers.append(box(in_channels=z[3]))  # 这里处理其他两层
88.         # print(layers)
89.         self.fes = nn.Sequential(*layers)
91.     def forward(self, x):
92.         x = self.conv1(x)
93.         x = self.bn1(x)
94.         x = self.relu(x)
95.         x = self.maxpool(x)
96.         x = self.fes(x)
97.         x = self.avgpool(x)
98.         x = torch.flatten(x, 1)
99.         x = self.fc(x)
100.         return x
105. in_out = [(128, 3, True, 256), (256, 4, True, 512), (512, 6, True, 1024), (1024, 3, True, 2048)]
107. s = New50(in_out=in_out)
108. def resnet500():
109.     return New50(in_out=in_out)
111. '''每层的第一层输入
112.    每层重复的次数
113.    是否走虚残差
114.    每层的第二个卷积核的输入'''

复制代码

后续还会上传ResNet30，FCN，UNet等架构图和代码。

来源：https://blog.csdn.net/qq_44697987/article/details/128178998
免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！