视频：https://www.bilibili.com/video/BV1Y7411d7Ys?p=11
博客

主要内容：各种CNN优化模型

语雀内容

GoogLeNet

GoogLeNet：14年ImageNet分类第一名。引入Inception模块，采用不同大小的卷积核意味着不同大小的感受野，最后拼接意味着不同尺度特征的融合；采用了average pooling来代替全连接层；避免梯度消失，网络额外增加了2个辅助的softmax用于向前传导梯度。

image.png

说明：Inception Moudel

  1. 卷积核超参数选择困难，自动**找到卷积的最佳组合，相应的权重大。**
  2. Concatenate：把张量拼接在一块,但要保证宽度和高度是一致的（average pooling宽高一致）
  3. 1x1卷积核，减少计算量，不同通道的信息融合。数量取决于输入张量的通道
    1. 卷积过程：各通道卷积再求和叠加为一个通道（cxwxh —> 1xwxh）
    2. 卷积后的每个通道信息包含了卷积前的各通道信息（融合）
    3. 多少个卷积核，卷积后多少个通道

image.pngimage.png

image.png

Inception Moudel代码说明

下图中，括号是输出channel
image.png

拼接在一起cat

沿着通道的维度拼接在一起 dim=1

  1. outputs = [branch1x1, branch5x5, branch3x3, branch_pool]
  2. return torch.cat(outputs, dim=1)

image.png

初始通道注意多少，没有写死
Inception后都是88通道

image.png
image.png

代码实现

代码说明：
1、先是1个卷积层(conv,maxpooling,relu)，然后inceptionA模块，接下来又是一个卷积层(conv,mp,relu),然后inceptionA模块，最后一个全连接层(fc)。
2、1408这个数据可以通过x = x.view(in_size, -1)后调用x.shape得到。
1408 = channelswidthheight = 8844

widthheight 44来源：**

原始照片size：28x*28 nn.Conv2d(1, 10, kernel_size=5)（-4）：24*24 nn.Conv2d(88, 20, kernel_size=5)（-4） ：20*20 第一次InceptionA（-8）： 第二次InceptionA（-8）：**44*

Inception后都是88通道
image.png

  1. '''
  2. Description: Inception Moudel
  3. 视频：https://www.bilibili.com/video/BV1Y7411d7Ys?p=11
  4. 博客
  5. • https://blog.csdn.net/bit452/article/details/109693790
  6. •  https://blog.csdn.net/weixin_44841652/article/details/105256034
  7. Author: HCQ
  8. Company(School): UCAS
  9. Email: 1756260160@qq.com
  10. Date: 2020-12-12 11:32:16
  11. LastEditTime: 2020-12-12 15:24:18
  12. FilePath: /pytorch/PyTorch深度学习实践/11.1AdvancedCNN-Inception.py
  13. '''
  14. """ 
  15. 1408 = channels*width*height = 88*4*4
  18. width*height 4*4来源：
  19. 原始照片size：28x*28
  20. nn.Conv2d(1, 10, kernel_size=5)（-4）：24*24
  21.  nn.Conv2d(88, 20, kernel_size=5)（-4） ：20*20
  22. 第一次InceptionA（-8）：
  23. 第二次InceptionA（-8）：4*4
  24.  """
  28. import torch
  29. import torch.nn as nn
  30. from torchvision import transforms
  31. from torchvision import datasets
  32. from torch.utils.data import DataLoader
  33. import torch.nn.functional as F
  34. import torch.optim as optim
  36. # prepare dataset
  38. batch_size = 64
  39. transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]) # 归一化,均值和方差
  41. train_dataset = datasets.MNIST(root='./data/mnist/', train=True, download=True, transform=transform)
  42. train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)
  43. test_dataset = datasets.MNIST(root='./data/mnist/', train=False, download=True, transform=transform)
  44. test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size)
  46. # design model using class
  47. class InceptionA(nn.Module):
  48.     def __init__(self, in_channels):
  49.         super(InceptionA, self).__init__()
  50.         self.branch1x1 = nn.Conv2d(in_channels, 16, kernel_size=1)
  52.         self.branch5x5_1 = nn.Conv2d(in_channels, 16, kernel_size=1)
  53.         self.branch5x5_2 = nn.Conv2d(16, 24, kernel_size=5, padding=2)
  55.         self.branch3x3_1 = nn.Conv2d(in_channels, 16, kernel_size=1)
  56.         self.branch3x3_2 = nn.Conv2d(16, 24, kernel_size=3, padding=1)
  57.         self.branch3x3_3 = nn.Conv2d(24, 24, kernel_size=3, padding=1)
  59.         self.branch_pool = nn.Conv2d(in_channels, 24, kernel_size=1)
  61.     def forward(self, x):
  62.         branch1x1 = self.branch1x1(x)
  64.         branch5x5 = self.branch5x5_1(x)
  65.         branch5x5 = self.branch5x5_2(branch5x5)
  67.         branch3x3 = self.branch3x3_1(x)
  68.         branch3x3 = self.branch3x3_2(branch3x3)
  69.         branch3x3 = self.branch3x3_3(branch3x3)
  71.         branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)
  72.         branch_pool = self.branch_pool(branch_pool)
  74.         outputs = [branch1x1, branch5x5, branch3x3, branch_pool]
  75.         return torch.cat(outputs, dim=1) # b,c,w,h  c对应的是dim=1
  77. class Net(nn.Module):
  78.     def __init__(self):
  79.         super(Net, self).__init__()
  80.         self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
  81.         self.conv2 = nn.Conv2d(88, 20, kernel_size=5) # 88 = 24x3 + 16
  83.         self.incep1 = InceptionA(in_channels=10) # 与conv1 中的10对应
  84.         self.incep2 = InceptionA(in_channels=20) # 与conv2 中的20对应
  86.         self.mp = nn.MaxPool2d(2)
  87.         self.fc = nn.Linear(1408, 10) # 暂时不知道1408咋能自动出来的  1408 = channels*width*height = 88*4*4
  90.     def forward(self, x):
  91.         in_size = x.size(0) # 64        x.size:  torch.Size([64, 10])
  92.         x = F.relu(self.mp(self.conv1(x)))
  93.         x = self.incep1(x)
  94.         x = F.relu(self.mp(self.conv2(x)))
  95.         x = self.incep2(x)
  96.         x = x.view(in_size, -1)  # torch.Size([64, 1408])    Channels:64  每一层channels1408
  97.         x = self.fc(x)
  99.         return x
  101. model = Net()
  103. # construct loss and optimizer
  104. criterion = torch.nn.CrossEntropyLoss()
  105. optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
  107. # training cycle forward, backward, update
  110. def train(epoch):
  111.     running_loss = 0.0
  112.     for batch_idx, data in enumerate(train_loader, 0):
  113.         inputs, target = data
  114.         optimizer.zero_grad()
  116.         outputs = model(inputs)
  117.         loss = criterion(outputs, target)
  118.         loss.backward()
  119.         optimizer.step()
  121.         running_loss += loss.item()
  122.         if batch_idx % 300 == 299:
  123.             print('[%d, %5d] loss: %.3f' % (epoch+1, batch_idx+1, running_loss/300))
  124.             running_loss = 0.0
  127. def test():
  128.     correct = 0
  129.     total = 0
  130.     with torch.no_grad():
  131.         for data in test_loader:
  132.             images, labels = data
  133.             outputs = model(images)
  134.             _, predicted = torch.max(outputs.data, dim=1)
  135.             total += labels.size(0)
  136.             correct += (predicted == labels).sum().item()
  137.     print('accuracy on test set: %d %% ' % (100*correct/total))
  140. if __name__ == '__main__':
  141.     for epoch in range(10):
  142.         train(epoch)
  143.         test()

结果展示

image.png

ResNet：引入残差单元，简化学习目标和难度，加快训练速度，模型加深时，不会产生退化问题；能够有效解决训练过程中梯度消失和梯度爆炸问题。ResNet

bert思路：多任务轮流训练

视频中截图： 说明：1、要解决的问题：梯度消失 2、跳连接，H(x) = F(x) + x,张量维度必须一样，加完后再激活。不要做pooling，张量的维度会发生变化。 代码说明： 1、先是1个卷积层(conv,maxpooling,relu)，然后ResidualBlock模块，接下来又是一个卷积层(conv,mp,relu),然后esidualBlock模块模块，最后一个全连接层(fc)。

层数越深越好？

针对CIFAR-10，20层比56层loss要小
看上去网络变复杂了，但其实性能降低了。可能原因：梯度消失

说明

1、要解决的问题：梯度消失
2、跳连接，H(x) = F(x) + x,张量维度必须一样，加完后再激活。不要做pooling，张量的维度会发生变化
最小梯度也是1左右

image.png
image.png

代码实现

代码说明：

  • 先是1个卷积层(conv,maxpooling,relu)，然后ResidualBlock模块，接下来又是一个卷积层(conv,mp,relu),然后ResidualBlock模块模块，最后一个全连接层
  • ResidualBlock模块的输出张量维度和输入张量维度要保持一致

image.png

image.png

输出通道和输入通道一致，代码中直接设置同一个channels

  • 卷积是层中的事情，res是层间的事情

image.png

总体代码

写网络，增量式开发的思想：一层写好测试下，再写下一层

image.png

'''
Description: ResNet
视频：https://www.bilibili.com/video/BV1Y7411d7Ys?p=11
博客
• https://blog.csdn.net/bit452/article/details/109693790
•  https://blog.csdn.net/weixin_44841652/article/details/105256034
Author: HCQ
Company(School): UCAS
Email: 1756260160@qq.com
Date: 2020-12-12 15:25:16
LastEditTime: 2020-12-12 15:29:18
FilePath: /pytorch/PyTorch深度学习实践/11.2AdvancedCNN-ResNet.py
'''


import torch
import torch.nn as nn
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader
import torch.nn.functional as F
import torch.optim as optim

# prepare dataset

batch_size = 64
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]) # 归一化,均值和方差

train_dataset = datasets.MNIST(root='./data/mnist/', train=True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)
test_dataset = datasets.MNIST(root='./data/mnist/', train=False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size)

# design model using class
class ResidualBlock(nn.Module):
    def __init__(self, channels):
        super(ResidualBlock, self).__init__()
        self.channels = channels
        self.conv1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
        self.conv2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)

    def forward(self, x):
        y = F.relu(self.conv1(x))
        y = self.conv2(y)
        return F.relu(x + y)

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 16, kernel_size=5)
        self.conv2 = nn.Conv2d(16, 32, kernel_size=5) # 88 = 24x3 + 16

        self.rblock1 = ResidualBlock(16)
        self.rblock2 = ResidualBlock(32)

        self.mp = nn.MaxPool2d(2)
        self.fc = nn.Linear(512, 10) # 512咋能自动出来的  512 = channels*width*height 


    def forward(self, x):
        in_size = x.size(0)

        x = self.mp(F.relu(self.conv1(x)))
        x = self.rblock1(x)
        x = self.mp(F.relu(self.conv2(x)))
        x = self.rblock2(x)

        x = x.view(in_size, -1)
        x = self.fc(x)
        return x

model = Net()

# construct loss and optimizer
criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)

# training cycle forward, backward, update


def train(epoch):
    running_loss = 0.0
    for batch_idx, data in enumerate(train_loader, 0):
        inputs, target = data
        optimizer.zero_grad()

        outputs = model(inputs)
        loss = criterion(outputs, target)
        loss.backward()
        optimizer.step()

        running_loss += loss.item()
        if batch_idx % 300 == 299:
            print('[%d, %5d] loss: %.3f' % (epoch+1, batch_idx+1, running_loss/300))
            running_loss = 0.0


def test():
    correct = 0
    total = 0
    with torch.no_grad():
        for data in test_loader:
            images, labels = data
            outputs = model(images)
            _, predicted = torch.max(outputs.data, dim=1)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()
    print('accuracy on test set: %d %% ' % (100*correct/total))


if __name__ == '__main__':
    for epoch in range(10):
        train(epoch)
        test()

结果展示

image.png

课后Exercise

He K, Zhang X, Ren S, et al. Identity Mappings in Deep Residual Networks[C]

更多Residual相关设计

image.png

Huang G, Liu Z, Laurens V D M, et al. Densely Connected Convolutional Networks[J]. 2016:2261-2269.

DenseNet：密集连接；加强特征传播，鼓励特征复用，极大的减少了参数量。DenseNet
多了好多跳
**

image.png