刘二大人 PyTorch深度学习实践 笔记 P11 卷积神经网络（高级篇）
- 1、GoogleNet
* - I 网络结构
- II 减少代码冗余思想(减少代码重复)
- 2、Inception Module
* - I 基本概念
- III 代码实现
- II Stack Layer
- 3、residual net
* - I 普通网络与残差网络的区别
- II Residual block
- III 代码实现
- 4、作业
* - 作业1：阅读论文 Identity Mappings in Deep Residual Networks
- 实现 constant scaling
- 实现conv shortcut
- 5、建议学习流程
1、GoogleNet
I 网络结构
神经网络当中还有许多更为复杂的网络结构，那么它们如何来实现？用什么样的方法？GoogleNet网络结构如图所示：
GoogleNet常被用作基础主干网络，图中红色圈出的一个部分称为Inception块。
; II 减少代码冗余思想(减少代码重复)
- 在c语言中 使用函数
- 面向对象过程中时 构造类
在GoogleNet中把相同的块封装成一个类来减少代码冗余。
2、Inception Module
I 基本概念
问题： 构造神经网络时，超参数比较难选，比如kernel。
解决办法： 把几种卷积都用一下，效果更好的卷积被赋予的权重会更大，自动找到最优卷积的组合，针对每一个卷积结果再进行求和。
- concarenate： 把张量拼接起来，必须保证图像的宽度和高度是一致的。
- 均值池化： 最大池化会导致图像变为原来的一半，均值池化可以人为指定padding 和 stride 来保证输入和输出的图像是一样的。
- 信息融合： 本质就是得到的值通过三个值通过某种运算得到的信息。考试对各科分数求总分进行比较分数高低，在多个维度下不太好比较。
- 1*1卷积： 也是相同大小的卷积核，其个数取决于输入张量的通道，最主要目的就是改变通道的数量，减少运算量。
此处就是在做一个通道的变换，原通道数为3，新的通道数是卷积核的个数，高度和宽度不变。
运算量变成了原来的十分之一，大大提高了计算效率。
; III 代码实现
import torch
from torch import 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
import matplotlib.pyplot as plt
batch_size = 64
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])
train_dataset = datasets.MNIST(root='dataset/mnist',
train=True,
download=True,
transform=transform)
train_loader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_dataset = datasets.MNIST(root='dataset/mnist',
train=False,
download=True,
transform=transform)
test_loader = DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
class InceptionA(nn.Module):
def __init__(self, in_channels):
super(InceptionA, self).__init__()
self.branch1X1 = nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch5X5_1 = nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch5X5_2 = nn.Conv2d(16, 24, kernel_size=5, padding=2)
self.branch3X3_1 = nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch3X3_2 = nn.Conv2d(16, 24, kernel_size=3, padding=1)
self.branch3X3_3 = nn.Conv2d(24, 24, kernel_size=3, padding=1)
self.branch_pool = nn.Conv2d(in_channels, 24, kernel_size=1)
def forward(self, x):
branch1X1 = self.branch1X1(x)
branch5X5 = self.branch5X5_1(x)
branch5X5 = self.branch5X5_2(branch5X5)
branch3X3 = self.branch3X3_1(x)
branch3X3 = self.branch3X3_2(branch3X3)
branch3X3 = self.branch3X3_3(branch3X3)
branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)
branch_pool = self.branch_pool(branch_pool)
outputs = [branch1X1, branch5X5, branch3X3, branch_pool]
return torch.cat(outputs, dim=1)
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
self.conv2 = nn.Conv2d(88, 20, kernel_size=5)
self.incep1 = InceptionA(in_channels=10)
self.incep2 = InceptionA(in_channels=20)
self.mp = nn.MaxPool2d(2)
self.fc = nn.Linear(1408, 10)
def forward(self, x):
in_size = x.size(0)
x = F.relu(self.mp(self.conv1(x)))
x = self.incep1(x)
x = F.relu(self.mp(self.conv2(x)))
x = self.incep2(x)
x = x.view(in_size, -1)
x = self.fc(x)
return x
model = Net()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
def train(epoch):
running_loss = 0
for batch_idx, data in enumerate(train_loader, 0):
inputs, target = data
inputs, target = inputs.to(device), target.to(device)
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
accuracy = []
def test():
correct = 0
total = 0
with torch.no_grad():
for data in test_loader:
images, labels = data
images, labels = images.to(device), labels.to(device)
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))
accuracy.append(100 * correct / total)
if __name__ == '__main__':
for epoch in range(10):
train(epoch)
test()
print(accuracy)
plt.plot(range(10), accuracy)
plt.xlabel("epoch")
plt.ylabel("Accuracy")
plt.show()
输出：
[1, 300] loss: 0.767
[1, 600] loss: 0.186
[1, 900] loss: 0.141
Accuracy on test set: 96 %
[2, 300] loss: 0.109
[2, 600] loss: 0.098
[2, 900] loss: 0.096
Accuracy on test set: 97 %
[3, 300] loss: 0.083
[3, 600] loss: 0.076
[3, 900] loss: 0.076
Accuracy on test set: 97 %
[4, 300] loss: 0.066
[4, 600] loss: 0.066
[4, 900] loss: 0.064
Accuracy on test set: 98 %
[5, 300] loss: 0.054
[5, 600] loss: 0.057
[5, 900] loss: 0.054
Accuracy on test set: 98 %
[6, 300] loss: 0.049
[6, 600] loss: 0.052
[6, 900] loss: 0.049
Accuracy on test set: 98 %
[7, 300] loss: 0.044
[7, 600] loss: 0.047
[7, 900] loss: 0.042
Accuracy on test set: 98 %
[8, 300] loss: 0.043
[8, 600] loss: 0.039
[8, 900] loss: 0.041
Accuracy on test set: 98 %
[9, 300] loss: 0.034
[9, 600] loss: 0.041
[9, 900] loss: 0.038
Accuracy on test set: 98 %
[10, 300] loss: 0.034
[10, 600] loss: 0.035
[10, 900] loss: 0.033
Accuracy on test set: 98 %
[96.51, 97.37, 97.94, 98.45, 98.31, 98.58, 98.59, 98.8, 98.73, 98.9]
性能提高不多，可能是最好全连接层太少，训练次数不一定越多越好，当前网络参数可以进行存盘，存储训练效果最好的结果。
II Stack Layer
问题： 为什么网络层数更深反而准确率会下降，训练效果更差？
梯度消失： 在反向传播时需要根据链式法则把一连串的梯度乘起来，若每个梯度都小于1，则乘起来的结果会接近于0，导致权重在更新时得不到什么更新，进而导致最开始的这些块（离输入近的块）没办法得到充分的训练。
解决办法： 逐层训练，每一层加锁，但是深度学习中层数太多了，难以实现。
; 3、residual net
I 普通网络与残差网络的区别
残差网络多一个跳连接，在做完卷积激活之前，将该层的输入加上输出一起作为整个的输出来激活。
; II Residual block
偏导数+1一定大于等于1，所以不会出现梯度消失的问题。
III 代码实现
import torch
from torch import 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
import matplotlib.pyplot as plt
batch_size = 64
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])
train_dataset = datasets.MNIST(root='dataset/mnist',
train=True,
download=True,
transform=transform)
train_loader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_dataset = datasets.MNIST(root='dataset/mnist',
train=False,
download=True,
transform=transform)
test_loader = DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
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)
self.mp = nn.MaxPool2d(2)
self.rblock1 = ResidualBlock(16)
self.rblock2 = ResidualBlock(32)
self.fc = nn.Linear(512, 10)
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()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
def train(epoch):
running_loss = 0
for batch_idx, data in enumerate(train_loader, 0):
inputs, target = data
inputs, target = inputs.to(device), target.to(device)
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
accuracy = []
def test():
correct = 0
total = 0
with torch.no_grad():
for data in test_loader:
images, labels = data
images, labels = images.to(device), labels.to(device)
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))
accuracy.append(100 * correct / total)
if __name__ == '__main__':
for epoch in range(10):
train(epoch)
test()
print(accuracy)
plt.plot(range(10), accuracy)
plt.xlabel("epoch")
plt.ylabel("Accuracy")
plt.show()
输出：
[1, 300] loss: 0.520
[1, 600] loss: 0.159
[1, 900] loss: 0.118
Accuracy on test set: 97 %
[2, 300] loss: 0.090
[2, 600] loss: 0.081
[2, 900] loss: 0.074
Accuracy on test set: 98 %
[3, 300] loss: 0.063
[3, 600] loss: 0.058
[3, 900] loss: 0.055
Accuracy on test set: 98 %
[4, 300] loss: 0.046
[4, 600] loss: 0.050
[4, 900] loss: 0.048
Accuracy on test set: 98 %
[5, 300] loss: 0.044
[5, 600] loss: 0.038
[5, 900] loss: 0.038
Accuracy on test set: 98 %
[6, 300] loss: 0.035
[6, 600] loss: 0.033
[6, 900] loss: 0.034
Accuracy on test set: 98 %
[7, 300] loss: 0.028
[7, 600] loss: 0.029
[7, 900] loss: 0.032
Accuracy on test set: 98 %
[8, 300] loss: 0.027
[8, 600] loss: 0.028
[8, 900] loss: 0.026
Accuracy on test set: 98 %
[9, 300] loss: 0.021
[9, 600] loss: 0.026
[9, 900] loss: 0.022
Accuracy on test set: 98 %
[10, 300] loss: 0.021
[10, 600] loss: 0.023
[10, 900] loss: 0.021
Accuracy on test set: 98 %
[97.03, 98.21, 98.47, 98.8, 98.52, 98.88, 98.88, 98.98, 98.95, 98.98]
4、作业
作业1：阅读论文 Identity Mappings in Deep Residual Networks
给出了很多residual block实现的方式。
; 实现 constant scaling
返回结果为原来的一半
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(x)
z = 0.5 * (x + y)
return F.relu(z)
输出：
[1, 300] loss: 0.947
[1, 600] loss: 0.252
[1, 900] loss: 0.173
Accuracy on test set: 96 %
[2, 300] loss: 0.126
[2, 600] loss: 0.113
[2, 900] loss: 0.107
Accuracy on test set: 97 %
[3, 300] loss: 0.085
[3, 600] loss: 0.084
[3, 900] loss: 0.077
Accuracy on test set: 98 %
[4, 300] loss: 0.064
[4, 600] loss: 0.066
[4, 900] loss: 0.068
Accuracy on test set: 98 %
[5, 300] loss: 0.057
[5, 600] loss: 0.058
[5, 900] loss: 0.055
Accuracy on test set: 98 %
[6, 300] loss: 0.051
[6, 600] loss: 0.051
[6, 900] loss: 0.047
Accuracy on test set: 98 %
[7, 300] loss: 0.042
[7, 600] loss: 0.044
[7, 900] loss: 0.048
Accuracy on test set: 98 %
[8, 300] loss: 0.041
[8, 600] loss: 0.040
[8, 900] loss: 0.040
Accuracy on test set: 98 %
[9, 300] loss: 0.035
[9, 600] loss: 0.037
[9, 900] loss: 0.037
Accuracy on test set: 98 %
[10, 300] loss: 0.031
[10, 600] loss: 0.038
[10, 900] loss: 0.031
Accuracy on test set: 98 %
[96.09, 97.78, 98.07, 98.29, 98.41, 98.67, 98.03, 98.86, 98.75, 98.81]
实现conv shortcut
多进行一次卷积
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)
self.conv3 = nn.Conv2d(channels, channels,
kernel_size=1)
def forward(self, x):
y = F.relu(self.conv1(x))
y = self.conv2(x)
z = self.conv3(x) + y
return F.relu(z)
输出：
[1, 300] loss: 0.686
[1, 600] loss: 0.192
[1, 900] loss: 0.137
Accuracy on test set: 96 %
[2, 300] loss: 0.105
[2, 600] loss: 0.093
[2, 900] loss: 0.078
Accuracy on test set: 98 %
[3, 300] loss: 0.073
[3, 600] loss: 0.065
[3, 900] loss: 0.060
Accuracy on test set: 98 %
[4, 300] loss: 0.054
[4, 600] loss: 0.049
[4, 900] loss: 0.056
Accuracy on test set: 98 %
[5, 300] loss: 0.042
[5, 600] loss: 0.048
[5, 900] loss: 0.040
Accuracy on test set: 98 %
[6, 300] loss: 0.041
[6, 600] loss: 0.039
[6, 900] loss: 0.037
Accuracy on test set: 98 %
[7, 300] loss: 0.034
[7, 600] loss: 0.033
[7, 900] loss: 0.035
Accuracy on test set: 98 %
[8, 300] loss: 0.029
[8, 600] loss: 0.030
[8, 900] loss: 0.031
Accuracy on test set: 98 %
[9, 300] loss: 0.025
[9, 600] loss: 0.027
[9, 900] loss: 0.028
Accuracy on test set: 98 %
[10, 300] loss: 0.023
[10, 600] loss: 0.026
[10, 900] loss: 0.026
Accuracy on test set: 98 %
[96.42, 98.2, 98.48, 98.7, 98.9, 98.89, 98.92, 98.99, 98.68, 98.97]
作业2：阅读论文 Densely Connected Convolutional Networks
怎么实现？
5、建议学习流程
- 理解网络模型理论 看花书 《动手学深度学习》。
- 阅读pytorch文档(至少通读一遍)，知道提供了什么功能以及文档结构。
- 复现经典工作，不是跑通代码，是先去读代码，学习架构；然后尝试自己来写，如此往复。
- 选特定研究领域，融会贯通，扩充视野，广泛阅读(前提是拥有前面的能力，看到论文，可以反映出代码怎么写，需要慢慢地积累)。
Original: https://blog.csdn.net/qq_44948213/article/details/126820242
Author: 小白*进阶ing
Title: 刘二大人 PyTorch深度学习实践 笔记 P11 卷积神经网络（高级篇）
原创文章受到原创版权保护。转载请注明出处：https://www.johngo689.com/708185/
转载文章受原作者版权保护。转载请注明原作者出处！