PyTorch程序练习（一）：PyTorch实现CIFAR-10多分类
一、准备数据
代码
import torchvision
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
# ========================================准备数据========================================
# 定义预处理函数
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
"""
①transfom.Compose可以把一些转换函数组合在一起。
②transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))对张量进行归一化：图像有三个通道，每个通道均值0.5，方差0.5。)
"""
# 下载CIFAR10数据，并对数据进行预处理
trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=False, transform=transform)
testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=False, transform=transform)
# 得到一个生成器
trainloader = DataLoader(trainset, batch_size=4, shuffle=True)
testloader = DataLoader(testset, batch_size=4, shuffle=False) # 数据分批
"""
①dataloader是一个可迭代对象，可以使用迭代器一样使用。
②用DataLoader得到生成器，这可节省内存。
"""
# 分类类别
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
# ========================================可视化源数据========================================
import matplotlib.pyplot as plt
import numpy as np
# 显示图像
def imshow(img):
img = img / 2 + 0.5 # unnormalize反归一化：img = img * std + mu
npimg = img.numpy() # 将图片转换为数组
plt.imshow(np.transpose(npimg, (1, 2, 0))) # transpose对高维矩阵进行轴对换：把图片中表示颜色顺序的RGB改为GBR
# 显示图像
plt.xticks([])
plt.yticks([])
plt.show()
# 随机获取部分训练数据
dataiter = iter(trainloader)
images, labels = dataiter.next()
# 打印标签
print(' '.join('%5s' % classes[labels[j]] for j in range(4)))
# 显示图像
imshow(torchvision.utils.make_grid(images))运行结果
horse ship truck plane二、全连接层FC
代码
from 准备数据 import trainloader,images,classes,testloader
# ========================================构建网络========================================
import torch
import torch.nn as nn
import torch.nn.functional as F
# 检测是否有GPU，有则使用，否则使用CPU
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") # cuda:0表示使用第一块GPU；若有显卡torch.cuda.is_available()返回True
"""
#若有多块显卡可并行化训练：
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs")
dim = 0 [20, xxx] -> [10, ...], [10, ...] on 2GPUs
model = nn.DataParallel(model)
"""
# --------------------构建卷积网络--------------------
class CNNNet(nn.Module):
def __init__(self):
super(CNNNet, self).__init__()
self.conv1 = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=(5,5), stride=(1,1)) #卷积层1
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2) #池化层1
self.conv2 = nn.Conv2d(in_channels=16, out_channels=36, kernel_size=(3,3), stride=(1,1)) #卷积层2
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2) #池化层2
self.fc1 = nn.Linear(1296, 128) #全连接层1
self.fc2 = nn.Linear(128, 10) #全连接层2
def forward(self, x):
x = self.pool1(F.relu(self.conv1(x)))
x = self.pool2(F.relu(self.conv2(x)))
x = x.view(-1, 36 * 6 * 6)
x = F.relu(self.fc2(F.relu(self.fc1(x))))
return x
# --------------------实例化网络--------------------
net = CNNNet() #实例化网络
net = net.to(device) #使用GPU训练
# 定义损失函数和优化器
import torch.optim as optim
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
#optimizer = optim.Adam(net.parameters(), lr=0.001)
"""
model.parameters()：传入模型参数；
lr：学习率；
momentum：动量，更新缩减和平移参数的频率和幅度，结合当前梯度与上一次更新信息用于当前更新
"""
# ========================================训练模型========================================
import time
for epoch in range(10):
start = time.time()
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
# 获取训练数据
inputs, labels = data #解包元组
inputs, labels = inputs.to(device), labels.to(device) #将数据放在GPU上
# 前向传播
outputs = net(inputs)
loss = criterion(outputs, labels)
# 反向传播
optimizer.zero_grad() #清空上一步的残余更新参数值，将神经网络参数梯度降为0
loss.backward() #误差反向传递，计算参数更新值
optimizer.step() #优化梯度，将参数更新值施加到net的parameters上
# 记录误差
running_loss += loss.item()
# 每2000步打印损失
if i % 2000 == 1999:
print('[%d, %5d] loss: %.3f, time:%.2f' % (epoch + 1, i + 1, running_loss / 2000, time.time()-start))
running_loss = 0.0
print('训练完成')
# 预测结果
images, labels = images.to(device), labels.to(device)
outputs = net(images)
_, predicted = torch.max(outputs, 1) #输出值的最大值作为预测值
print('预测结果: ', ' '.join('%5s' % classes[predicted[j]]for j in range(4)))
# ========================================测试模型========================================
# --------------------整体预测准确率--------------------
correct = 0
total = 0
with torch.no_grad():
# 从测试数据中取出数据
for data in testloader:
images, labels = data
images, labels = images.to(device), labels.to(device)
outputs = net(images)
_, predicted = torch.max(outputs.data, 1) #模型预测的最大值作为预测值
total += labels.size(0) #统计测试图片个数
correct += (predicted == labels).sum().item() #统计正确预测的图片数
print('10000张测试图片的准确率：%d%%' % (100 * correct / total))
# --------------------每个类别预测准确率--------------------
class_correct = list(0. for i in range(10)) #正确的类别个数
class_total = list(0. for i in range(10)) #一共10个类别
with torch.no_grad():
# 从测试数据中取出数据
for data in testloader:
images, labels = data
images, labels = images.to(device), labels.to(device)
outputs = net(images)
_, predicted = torch.max(outputs, 1)
c = (predicted == labels).squeeze() #预测正确的返回True，预测错误的返回False；squeeze将数据转换为一维数据
for i in range(4):
label = labels[i] #提取标签
class_correct[label] += c[i].item() #预测正确个数
class_total[label] += 1 #总数
for i in range(10):
print('%5s的准确率:%2d%%' % (classes[i], 100 * class_correct[i] / class_total[i]))运行结果
plane car deer horse[1, 2000] loss: 2.167, time:5.35
[1, 4000] loss: 1.979, time:9.37
[1, 6000] loss: 1.842, time:13.38
[1, 8000] loss: 1.732, time:17.44
[1, 10000] loss: 1.474, time:21.50
[1, 12000] loss: 1.380, time:25.49
[2, 2000] loss: 1.246, time:4.10
[2, 4000] loss: 1.210, time:8.21
[2, 6000] loss: 1.180, time:12.33
[2, 8000] loss: 1.156, time:16.47
[2, 10000] loss: 1.088, time:20.49
[2, 12000] loss: 1.063, time:24.45
[3, 2000] loss: 0.987, time:3.99
[3, 4000] loss: 0.975, time:7.97
[3, 6000] loss: 0.970, time:11.93
[3, 8000] loss: 0.947, time:15.89
[3, 10000] loss: 0.964, time:19.85
[3, 12000] loss: 0.930, time:23.81
[4, 2000] loss: 0.814, time:3.98
[4, 4000] loss: 0.831, time:7.93
[4, 6000] loss: 0.833, time:11.91
[4, 8000] loss: 0.829, time:15.86
[4, 10000] loss: 0.841, time:19.81
[4, 12000] loss: 0.826, time:23.78
[5, 2000] loss: 0.701, time:3.98
[5, 4000] loss: 0.708, time:7.93
[5, 6000] loss: 0.710, time:11.88
[5, 8000] loss: 0.746, time:15.84
[5, 10000] loss: 0.734, time:19.81
[5, 12000] loss: 0.756, time:23.78
[6, 2000] loss: 0.582, time:4.23
[6, 4000] loss: 0.597, time:8.23
[6, 6000] loss: 0.628, time:12.22
[6, 8000] loss: 0.661, time:16.20
[6, 10000] loss: 0.667, time:20.18
[6, 12000] loss: 0.678, time:24.14
[7, 2000] loss: 0.494, time:3.98
[7, 4000] loss: 0.543, time:7.94
[7, 6000] loss: 0.576, time:11.92
[7, 8000] loss: 0.577, time:15.89
[7, 10000] loss: 0.586, time:19.85
[7, 12000] loss: 0.604, time:23.83
[8, 2000] loss: 0.442, time:4.02
[8, 4000] loss: 0.457, time:7.97
[8, 6000] loss: 0.494, time:11.96
[8, 8000] loss: 0.524, time:15.93
[8, 10000] loss: 0.521, time:19.92
[8, 12000] loss: 0.558, time:23.90
[9, 2000] loss: 0.349, time:3.99
[9, 4000] loss: 0.389, time:7.96
[9, 6000] loss: 0.408, time:11.92
[9, 8000] loss: 0.464, time:15.88
[9, 10000] loss: 0.500, time:19.85
[9, 12000] loss: 0.490, time:23.82
[10, 2000] loss: 0.319, time:4.00
[10, 4000] loss: 0.354, time:7.98
[10, 6000] loss: 0.356, time:12.02
[10, 8000] loss: 0.428, time:16.11
[10, 10000] loss: 0.422, time:20.06
[10, 12000] loss: 0.447, time:24.25
训练完成
预测结果: plane car dog horse
10000张测试图片的准确率：66%
plane的准确率:72%
car的准确率:74%
bird的准确率:60%
cat的准确率:50%
deer的准确率:54%
dog的准确率:60%
frog的准确率:72%
horse的准确率:68%
ship的准确率:80%
truck的准确率:73%三、全局平均池化层GAP
代码
from 准备数据 import trainloader, testloader, classes, images
# ========================================构建网络========================================
import torch
import torch.nn as nn
import torch.nn.functional as F
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# --------------------构建卷积网络--------------------
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(3, 16, 5) #卷积层1
self.pool1 = nn.MaxPool2d(2, 2) #池化层1
self.conv2 = nn.Conv2d(16, 36, 5) #卷积层2
self.pool2 = nn.MaxPool2d(2, 2) #池化层2
self.aap=nn.AdaptiveAvgPool2d(1) #全局平均池化层
self.fc3 = nn.Linear(36, 10) #全连接层
def forward(self, x):
x = self.pool1(F.relu(self.conv1(x))) #卷积层1->激励函数层->池化层1
x = self.pool2(F.relu(self.conv2(x))) #卷积层2->激励函数层->池化层2
x = self.aap(x)
x = x.view(x.shape[0], -1)
x = self.fc3(x)
return x
# --------------------实例化网络--------------------
net = Net()
net=net.to(device)
# 定义损失函数和优化器
import torch.optim as optim
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
# ========================================训练模型========================================
import time
for epoch in range(10):
start=time.time()
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
# 获取训练数据
inputs, labels = data #解包元组
inputs, labels = inputs.to(device), labels.to(device) #将数据放在GPU上
# 前向传播
outputs = net(inputs)
loss = criterion(outputs, labels)
# 反向传播
optimizer.zero_grad() # 清空上一步的残余更新参数值，将神经网络参数梯度降为0
loss.backward() # 误差反向传递，计算参数更新值
optimizer.step() # 优化梯度，将参数更新值施加到net的parameters上
# 记录误差
running_loss += loss.item()
# 每2000步打印损失
if i % 2000 == 1999:
print('[%d, %5d] loss: %.3f, time:%.2f' % (epoch + 1, i + 1, running_loss / 2000, time.time()-start))
running_loss = 0.0
print('训练完成')
# 预测结果
images, labels = images.to(device), labels.to(device)
outputs = net(images)
_, predicted = torch.max(outputs, 1) #输出值的最大值作为预测值
print('预测结果: ', ' '.join('%5s' % classes[predicted[j]]for j in range(4)))
# ========================================测试模型========================================
# --------------------整体预测准确率--------------------
correct = 0
total = 0
with torch.no_grad():
# 从测试数据中取出数据
for data in testloader:
images, labels = data
images, labels = images.to(device), labels.to(device)
outputs = net(images)
_, predicted = torch.max(outputs.data, 1) #模型预测的最大值作为预测值
total += labels.size(0) #统计测试图片个数
correct += (predicted == labels).sum().item() #统计正确预测的图片数
print('10000张测试图片的准确率：%d%%' % (100 * correct / total))
# --------------------每个类别预测准确率--------------------
class_correct = list(0. for i in range(10)) #正确的类别个数
class_total = list(0. for i in range(10)) #一共10个类别
with torch.no_grad():
# 从测试数据中取出数据
for data in testloader:
images, labels = data
images, labels = images.to(device), labels.to(device)
outputs = net(images)
_, predicted = torch.max(outputs, 1)
c = (predicted == labels).squeeze() #预测正确的返回True，预测错误的返回False；squeeze将数据转换为一维数据
for i in range(4):
label = labels[i] #提取标签
class_correct[label] += c[i].item() #预测正确个数
class_total[label] += 1 #总数
for i in range(10):
print('%5s的准确率:%2d%%' % (classes[i], 100 * class_correct[i] / class_total[i]))
运行结果
deer ship horse plane[1, 2000] loss: 2.129, time:5.14
[1, 4000] loss: 1.875, time:8.86
[1, 6000] loss: 1.741, time:12.61
[1, 8000] loss: 1.693, time:16.29
[1, 10000] loss: 1.638, time:20.04
[1, 12000] loss: 1.598, time:23.71
[2, 2000] loss: 1.549, time:3.71
[2, 4000] loss: 1.504, time:7.43
[2, 6000] loss: 1.484, time:11.13
[2, 8000] loss: 1.461, time:14.89
[2, 10000] loss: 1.431, time:18.68
[2, 12000] loss: 1.407, time:22.37
[3, 2000] loss: 1.376, time:3.69
[3, 4000] loss: 1.343, time:7.37
[3, 6000] loss: 1.349, time:11.18
[3, 8000] loss: 1.324, time:14.89
[3, 10000] loss: 1.303, time:18.60
[3, 12000] loss: 1.297, time:22.36
[4, 2000] loss: 1.262, time:3.69
[4, 4000] loss: 1.250, time:7.36
[4, 6000] loss: 1.246, time:11.01
[4, 8000] loss: 1.240, time:14.68
[4, 10000] loss: 1.239, time:18.35
[4, 12000] loss: 1.232, time:22.00
[5, 2000] loss: 1.185, time:3.69
[5, 4000] loss: 1.214, time:7.34
[5, 6000] loss: 1.194, time:11.01
[5, 8000] loss: 1.175, time:14.67
[5, 10000] loss: 1.148, time:18.36
[5, 12000] loss: 1.174, time:22.03
[6, 2000] loss: 1.113, time:3.70
[6, 4000] loss: 1.160, time:7.36
[6, 6000] loss: 1.129, time:11.03
[6, 8000] loss: 1.144, time:14.70
[6, 10000] loss: 1.125, time:18.35
[6, 12000] loss: 1.110, time:22.01
[7, 2000] loss: 1.103, time:3.70
[7, 4000] loss: 1.100, time:7.34
[7, 6000] loss: 1.098, time:11.02
[7, 8000] loss: 1.091, time:14.67
[7, 10000] loss: 1.077, time:18.32
[7, 12000] loss: 1.076, time:21.97
[8, 2000] loss: 1.065, time:3.68
[8, 4000] loss: 1.066, time:7.33
[8, 6000] loss: 1.037, time:10.99
[8, 8000] loss: 1.058, time:14.65
[8, 10000] loss: 1.054, time:18.31
[8, 12000] loss: 1.064, time:21.96
[9, 2000] loss: 1.060, time:3.68
[9, 4000] loss: 1.024, time:7.33
[9, 6000] loss: 1.032, time:10.99
[9, 8000] loss: 1.029, time:14.66
[9, 10000] loss: 1.029, time:18.30
[9, 12000] loss: 1.011, time:21.96
[10, 2000] loss: 0.995, time:3.67
[10, 4000] loss: 1.023, time:7.33
[10, 6000] loss: 1.007, time:11.00
[10, 8000] loss: 1.020, time:14.64
[10, 10000] loss: 0.996, time:18.29
[10, 12000] loss: 1.014, time:21.96
训练完成
预测结果: frog car horse plane
10000张测试图片的准确率：63%
plane的准确率:66%
car的准确率:73%
bird的准确率:51%
cat的准确率:57%
deer的准确率:43%
dog的准确率:32%
frog的准确率:76%
horse的准确率:74%
ship的准确率:81%
truck的准确率:79%四、程序分析
本程序用了两种方法构建卷积神经网络，即全局平均池化和不采用全局平均池化。通过程序的运行结果可以看出，不采用全局平均池化相较采用全局平均池化准确率更高，但训练时间更长。
两种方法构建卷积神经网络时，同样采用两个卷积层和两个池化层（MaxPool2d），但不采用全局平均池化的神经网络使用了两个全连接层，采用全局平均池化的神经网络只使用了一个全连接测。
相较之下，使用全局平均池化层能减少很多参数，可以减轻过拟合的发生，而且在减少参数的同时，其泛化能力也比较好。