🤵 Author ：Horizon Max
✨ 编程技巧篇：各种操作小结
🎇 机器视觉篇：会变魔术 OpenCV
💥 深度学习篇：简单入门 PyTorch
🏆 神经网络篇：经典网络模型
💻 算法篇：再忙也别忘了 LeetCode
[ 图像分类 ] 经典网络模型4——ResNet 详解与复现
🚀 Residual Network
Residual Network 简称 ResNet (残差网络)，何凯明团队于2015年提出的一种网络；
在2015年 ImageNet 挑战赛(ILSVRC) classification 任务中获得了 冠军；
目前在检测，分割，识别等领域里得到了广泛的应用；
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/69b902adfe414d12ab926083f5f81832.gif#pic_center)
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://www.johngo689.com/wp-content/themes/justnews/themer/assets/images/lazy.png)
🔗 论文地址：Deep Residual Learning for Image Recognition
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/9ec8b968fb88456980e137e889cf389c.png)
; 🚀 ResNet 详解
🎨 残差网络
对于一个网络，如果简单地增加深度，就会导致 梯度弥散 或 梯度爆炸，我们采取的解决方法是 正则化 ；
随着网络层数进一步增加，又会出现模型退化问题，在训练集上的 准确率出现饱和甚至下降 的现象 ；
特点：
通过利用内部的残差块实现跳跃连接，解决神经网络深度加深带来的 模型退化 问题：
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/6874056ec34349e2865d66ab92d1a784.png#pic_center)
residual block
在 传统网络 中采用的输入输出函数为：F(x) output1 = x input
在 残差网络 中利用残差模块使输入输出函数为：F(x) output2 = F(x) output1 + x input
x input 直接跳过多层 加入到最后的输出 F(x) output2 单元当中，解决 F(x) output1 可能带来的 模型退化问题
; 🎨 Residual Block
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/8060292fc262430cb6a18e85dd969f23.png)
浅层网络： 采用的是左侧的 residual block 结构（18-layer、34-layer）
深层网络： 采用的是右侧的 residual block 结构（50-layer、101-layer、152-layer）
很有意思的是，这两个设计具有参数量：
左侧： （3 X 3 X 64）X 64 +（3 X 3 X 64）X 64 = 73728
右侧： （1 X 1 X 64）X 64 +（3 X 3 X 64）X 64 +（1 X 1 X 64）X 256 +（1 X 1 X 64）X 256 = 73728
通过 1X1卷积 既能够改变通道数，又能大幅减少 计算量 和 参数量 ；
可以对比 34-layer 和 50-layer 发现它们的参数量分别为 3.6 X 109 和 3.8 X 109
🎨 ResNet50 详解
以下以 ResNet50 为代表进行介绍：
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/5772cb68a18049a8a1bbae9f1f572cc3.png#pic_center)
它是由 4 个 大block 组成 ；
每个 大block 分别由
[3, 4, 6, 3] 个 小block 组成 ；每个 小block 都有 三个 卷积操作 ；
在网络开始前还有 一个 卷积操作 ；
层数：（3+4+6+3）X 3 + 1 =
50 layer ；
其中每个 大的 block 里面都是由两部分组成： Conv Block 和 Identity Block
Conv Block ：输入和输出维度不相同，不能串联，主要用于 改变网络维度 ；
Identity Block ：输入和输出维度相同，可以串联，主要用于 加深网络层数 ；
Conv Block 结构框图：
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/16ec3041a0e94360831d12f1f440eba9.png)
Identity Block 结构框图：![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/955996f7721649cc866c840d600c696e.png)
ResNet50 [3, 4, 6, 3]可以表示为：
conv2_x： Conv Block + Identity Block + Identity Block
conv3_x： Conv Block + Identity Block + Identity Block + Identity Block
conv4_x： Conv Block + Identity Block + Identity Block + Identity Block + Identity Block + Identity Block
conv5_x： Conv Block + Identity Block + Identity Block
ResNet50 结构框图：
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/be5d67944200497a8ca5eb20c4bd166b.png#pic_center)
; 🚀 ResNet 复现
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchinfo import summary
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, in_planes, planes, stride=1):
super(BasicBlock, self).__init__()
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3,
stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion*planes,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion*planes)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
out = F.relu(out)
return out
class Bottleneck(nn.Module):
expansion = 4
def __init__(self, in_planes, planes, stride=1):
super(Bottleneck, self).__init__()
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
stride=stride, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.conv3 = nn.Conv2d(planes, self.expansion*planes,
kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(self.expansion*planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion*planes,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion*planes)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.bn2(self.conv2(out)))
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
out = F.relu(out)
return out
class ResNet(nn.Module):
def __init__(self, block, num_blocks, num_classes=1000):
super(ResNet, self).__init__()
self.in_planes = 64
self.conv1 = nn.Conv2d(3, 64, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.linear = nn.Linear(512 * block.expansion, num_classes)
def _make_layer(self, block, planes, num_blocks, stride):
strides = [stride] + [1]*(num_blocks-1)
layers = []
for stride in strides:
layers.append(block(self.in_planes, planes, stride))
self.in_planes = planes * block.expansion
return nn.Sequential(*layers)
def forward(self, x):
x = F.relu(self.bn1(self.conv1(x)))
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
x = torch.flatten(x, 1)
out = self.linear(x)
return out
def ResNet18():
return ResNet(BasicBlock, [2, 2, 2, 2])
def ResNet34():
return ResNet(BasicBlock, [3, 4, 6, 3])
def ResNet50():
return ResNet(Bottleneck, [3, 4, 6, 3])
def ResNet101():
return ResNet(Bottleneck, [3, 4, 23, 3])
def ResNet152():
return ResNet(Bottleneck, [3, 8, 36, 3])
def test():
net = ResNet50()
y = net(torch.randn(1, 3, 224, 224))
print(y.size())
summary(net, (1, 3, 224, 224))
if __name__ == '__main__':
test()
输出结果：
torch.Size([1, 1000])
==========================================================================================
Layer (type:depth-idx) Output Shape Param
==========================================================================================
ResNet -- --
├─Conv2d: 1-1 [1, 64, 224, 224] 1,728
├─BatchNorm2d: 1-2 [1, 64, 224, 224] 128
├─Sequential: 1-3 [1, 256, 224, 224] --
│ └─Bottleneck: 2-1 [1, 256, 224, 224] --
│ │ └─Conv2d: 3-1 [1, 64, 224, 224] 4,096
│ │ └─BatchNorm2d: 3-2 [1, 64, 224, 224] 128
│ │ └─Conv2d: 3-3 [1, 64, 224, 224] 36,864
│ │ └─BatchNorm2d: 3-4 [1, 64, 224, 224] 128
│ │ └─Conv2d: 3-5 [1, 256, 224, 224] 16,384
│ │ └─BatchNorm2d: 3-6 [1, 256, 224, 224] 512
│ │ └─Sequential: 3-7 [1, 256, 224, 224] 16,896
│ └─Bottleneck: 2-2 [1, 256, 224, 224] --
│ │ └─Conv2d: 3-8 [1, 64, 224, 224] 16,384
│ │ └─BatchNorm2d: 3-9 [1, 64, 224, 224] 128
│ │ └─Conv2d: 3-10 [1, 64, 224, 224] 36,864
│ │ └─BatchNorm2d: 3-11 [1, 64, 224, 224] 128
│ │ └─Conv2d: 3-12 [1, 256, 224, 224] 16,384
│ │ └─BatchNorm2d: 3-13 [1, 256, 224, 224] 512
│ │ └─Sequential: 3-14 [1, 256, 224, 224] --
│ └─Bottleneck: 2-3 [1, 256, 224, 224] --
│ │ └─Conv2d: 3-15 [1, 64, 224, 224] 16,384
│ │ └─BatchNorm2d: 3-16 [1, 64, 224, 224] 128
│ │ └─Conv2d: 3-17 [1, 64, 224, 224] 36,864
│ │ └─BatchNorm2d: 3-18 [1, 64, 224, 224] 128
│ │ └─Conv2d: 3-19 [1, 256, 224, 224] 16,384
│ │ └─BatchNorm2d: 3-20 [1, 256, 224, 224] 512
│ │ └─Sequential: 3-21 [1, 256, 224, 224] --
├─Sequential: 1-4 [1, 512, 112, 112] --
│ └─Bottleneck: 2-4 [1, 512, 112, 112] --
│ │ └─Conv2d: 3-22 [1, 128, 224, 224] 32,768
│ │ └─BatchNorm2d: 3-23 [1, 128, 224, 224] 256
│ │ └─Conv2d: 3-24 [1, 128, 112, 112] 147,456
│ │ └─BatchNorm2d: 3-25 [1, 128, 112, 112] 256
│ │ └─Conv2d: 3-26 [1, 512, 112, 112] 65,536
│ │ └─BatchNorm2d: 3-27 [1, 512, 112, 112] 1,024
│ │ └─Sequential: 3-28 [1, 512, 112, 112] 132,096
│ └─Bottleneck: 2-5 [1, 512, 112, 112] --
│ │ └─Conv2d: 3-29 [1, 128, 112, 112] 65,536
│ │ └─BatchNorm2d: 3-30 [1, 128, 112, 112] 256
│ │ └─Conv2d: 3-31 [1, 128, 112, 112] 147,456
│ │ └─BatchNorm2d: 3-32 [1, 128, 112, 112] 256
│ │ └─Conv2d: 3-33 [1, 512, 112, 112] 65,536
│ │ └─BatchNorm2d: 3-34 [1, 512, 112, 112] 1,024
│ │ └─Sequential: 3-35 [1, 512, 112, 112] --
│ └─Bottleneck: 2-6 [1, 512, 112, 112] --
│ │ └─Conv2d: 3-36 [1, 128, 112, 112] 65,536
│ │ └─BatchNorm2d: 3-37 [1, 128, 112, 112] 256
│ │ └─Conv2d: 3-38 [1, 128, 112, 112] 147,456
│ │ └─BatchNorm2d: 3-39 [1, 128, 112, 112] 256
│ │ └─Conv2d: 3-40 [1, 512, 112, 112] 65,536
│ │ └─BatchNorm2d: 3-41 [1, 512, 112, 112] 1,024
│ │ └─Sequential: 3-42 [1, 512, 112, 112] --
│ └─Bottleneck: 2-7 [1, 512, 112, 112] --
│ │ └─Conv2d: 3-43 [1, 128, 112, 112] 65,536
│ │ └─BatchNorm2d: 3-44 [1, 128, 112, 112] 256
│ │ └─Conv2d: 3-45 [1, 128, 112, 112] 147,456
│ │ └─BatchNorm2d: 3-46 [1, 128, 112, 112] 256
│ │ └─Conv2d: 3-47 [1, 512, 112, 112] 65,536
│ │ └─BatchNorm2d: 3-48 [1, 512, 112, 112] 1,024
│ │ └─Sequential: 3-49 [1, 512, 112, 112] --
├─Sequential: 1-5 [1, 1024, 56, 56] --
│ └─Bottleneck: 2-8 [1, 1024, 56, 56] --
│ │ └─Conv2d: 3-50 [1, 256, 112, 112] 131,072
│ │ └─BatchNorm2d: 3-51 [1, 256, 112, 112] 512
│ │ └─Conv2d: 3-52 [1, 256, 56, 56] 589,824
│ │ └─BatchNorm2d: 3-53 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-54 [1, 1024, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-55 [1, 1024, 56, 56] 2,048
│ │ └─Sequential: 3-56 [1, 1024, 56, 56] 526,336
│ └─Bottleneck: 2-9 [1, 1024, 56, 56] --
│ │ └─Conv2d: 3-57 [1, 256, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-58 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-59 [1, 256, 56, 56] 589,824
│ │ └─BatchNorm2d: 3-60 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-61 [1, 1024, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-62 [1, 1024, 56, 56] 2,048
│ │ └─Sequential: 3-63 [1, 1024, 56, 56] --
│ └─Bottleneck: 2-10 [1, 1024, 56, 56] --
│ │ └─Conv2d: 3-64 [1, 256, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-65 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-66 [1, 256, 56, 56] 589,824
│ │ └─BatchNorm2d: 3-67 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-68 [1, 1024, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-69 [1, 1024, 56, 56] 2,048
│ │ └─Sequential: 3-70 [1, 1024, 56, 56] --
│ └─Bottleneck: 2-11 [1, 1024, 56, 56] --
│ │ └─Conv2d: 3-71 [1, 256, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-72 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-73 [1, 256, 56, 56] 589,824
│ │ └─BatchNorm2d: 3-74 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-75 [1, 1024, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-76 [1, 1024, 56, 56] 2,048
│ │ └─Sequential: 3-77 [1, 1024, 56, 56] --
│ └─Bottleneck: 2-12 [1, 1024, 56, 56] --
│ │ └─Conv2d: 3-78 [1, 256, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-79 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-80 [1, 256, 56, 56] 589,824
│ │ └─BatchNorm2d: 3-81 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-82 [1, 1024, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-83 [1, 1024, 56, 56] 2,048
│ │ └─Sequential: 3-84 [1, 1024, 56, 56] --
│ └─Bottleneck: 2-13 [1, 1024, 56, 56] --
│ │ └─Conv2d: 3-85 [1, 256, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-86 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-87 [1, 256, 56, 56] 589,824
│ │ └─BatchNorm2d: 3-88 [1, 256, 56, 56] 512
│ │ └─Conv2d: 3-89 [1, 1024, 56, 56] 262,144
│ │ └─BatchNorm2d: 3-90 [1, 1024, 56, 56] 2,048
│ │ └─Sequential: 3-91 [1, 1024, 56, 56] --
├─Sequential: 1-6 [1, 2048, 28, 28] --
│ └─Bottleneck: 2-14 [1, 2048, 28, 28] --
│ │ └─Conv2d: 3-92 [1, 512, 56, 56] 524,288
│ │ └─BatchNorm2d: 3-93 [1, 512, 56, 56] 1,024
│ │ └─Conv2d: 3-94 [1, 512, 28, 28] 2,359,296
│ │ └─BatchNorm2d: 3-95 [1, 512, 28, 28] 1,024
│ │ └─Conv2d: 3-96 [1, 2048, 28, 28] 1,048,576
│ │ └─BatchNorm2d: 3-97 [1, 2048, 28, 28] 4,096
│ │ └─Sequential: 3-98 [1, 2048, 28, 28] 2,101,248
│ └─Bottleneck: 2-15 [1, 2048, 28, 28] --
│ │ └─Conv2d: 3-99 [1, 512, 28, 28] 1,048,576
│ │ └─BatchNorm2d: 3-100 [1, 512, 28, 28] 1,024
│ │ └─Conv2d: 3-101 [1, 512, 28, 28] 2,359,296
│ │ └─BatchNorm2d: 3-102 [1, 512, 28, 28] 1,024
│ │ └─Conv2d: 3-103 [1, 2048, 28, 28] 1,048,576
│ │ └─BatchNorm2d: 3-104 [1, 2048, 28, 28] 4,096
│ │ └─Sequential: 3-105 [1, 2048, 28, 28] --
│ └─Bottleneck: 2-16 [1, 2048, 28, 28] --
│ │ └─Conv2d: 3-106 [1, 512, 28, 28] 1,048,576
│ │ └─BatchNorm2d: 3-107 [1, 512, 28, 28] 1,024
│ │ └─Conv2d: 3-108 [1, 512, 28, 28] 2,359,296
│ │ └─BatchNorm2d: 3-109 [1, 512, 28, 28] 1,024
│ │ └─Conv2d: 3-110 [1, 2048, 28, 28] 1,048,576
│ │ └─BatchNorm2d: 3-111 [1, 2048, 28, 28] 4,096
│ │ └─Sequential: 3-112 [1, 2048, 28, 28] --
├─AdaptiveAvgPool2d: 1-7 [1, 2048, 1, 1] --
├─Linear: 1-8 [1, 1000] 2,049,000
==========================================================================================
Total params: 25,549,352
Trainable params: 25,549,352
Non-trainable params: 0
Total mult-adds (G): 63.59
==========================================================================================
Input size (MB): 0.60
Forward/backward pass size (MB): 2691.05
Params size (MB): 102.20
Estimated Total Size (MB): 2793.85
==========================================================================================
🚀 ResNet50 结构框图
![[ 图像分类 ] 经典网络模型4——ResNet 详解与复现](https://johngo-pic.oss-cn-beijing.aliyuncs.com/articles/20230619/d231078db2044bb78624ecaf787f0258.png)
Original: https://blog.csdn.net/weixin_45084253/article/details/124121400
Author: Horizon Max
Title: [ 图像分类 ] 经典网络模型4——ResNet 详解与复现
原创文章受到原创版权保护。转载请注明出处：https://www.johngo689.com/690271/
转载文章受原作者版权保护。转载请注明原作者出处！