最近在基于Pytorch框架补一些CNN的基础知识，学会自己写简单的卷积神经网络，从加载数据集到训练模型、测试模型、保存模型和输出测试结果，现在来总结一下。
首先基于Pytorch实现Vgg16网络，命名为model.py（可为其他任意名字，但是后续导入时要记得更改）
import torch
import torch.nn as nn
class VGG16(nn.Module):
def __init__(self):
super(VGG16, self).__init__()
self.layer1 = nn.Sequential(
nn.Conv2d(3, 64, 3, 1, 1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.Conv2d(64, 64, 3, 1, 1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(2, 2)
)
self.layer2 = nn.Sequential(
nn.Conv2d(64, 128, 3, 1, 1),
nn.BatchNorm2d(128),
nn.ReLU(),
nn.Conv2d(128, 128, 3, 1, 1),
nn.BatchNorm2d(128),
nn.ReLU(),
nn.MaxPool2d(2, 2)
)
self.layer3 = nn.Sequential(
nn.Conv2d(128, 256, 3, 1, 1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.Conv2d(256, 256, 3, 1, 1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.Conv2d(256, 256, 3, 1, 1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.MaxPool2d(2, 2)
)
self.layer4 = nn.Sequential(
nn.Conv2d(256, 512, 3, 1, 1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.Conv2d(512, 512, 3, 1, 1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.Conv2d(512, 512, 3, 1, 1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.MaxPool2d(2, 2)
)
self.layer5 = nn.Sequential(
nn.Conv2d(512, 512, 3, 1, 1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.Conv2d(512, 512, 3, 1, 1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.Conv2d(512, 512, 3, 1, 1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.MaxPool2d(2, 2)
)
self.fc1 = nn.Sequential(
nn.Flatten(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Dropout()
)
self.fc2 = nn.Sequential(
nn.Linear(512, 256),
nn.ReLU(),
nn.Dropout()
)
self.fc3 = nn.Linear(256, 10)
def forward(self, x):
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.layer5(x)
x = self.fc1(x)
x = self.fc2(x)
x = self.fc3(x)
return x
if __name__ == '__main__':
VGG16 = VGG16()
input = torch.ones((64, 3, 32, 32))
output = VGG16(input)
print(output.shape)
然后编写训练文件，导入需要的库（有些库没有的话需要提前安装）。
import torchvision
from torch import optim
from torch.utils.data import DataLoader
import torch.nn as nn
from model import *
import matplotlib.pyplot as plt
import time下载CIFAR10数据集，并设置batch_size，第一次运行时会自动下载，之后直接加载已经下载好的。当然也可以使用其他数据集，从官方文档https://pytorch.org/vision/stable/datasets.html可以看到有各种用于分类、分割、检测等其他视觉任务的数据集，但是为了方便学习，这里选择较小的CIFAR10数据集。
rain_data = torchvision.datasets.CIFAR10(root="data", train=True, transform=torchvision.transforms.ToTensor(),
download=True)
test_data = torchvision.datasets.CIFAR10(root="data", train=False, transform=torchvision.transforms.ToTensor(),
download=True)
train_dataloader = DataLoader(train_data, batch_size=64)
test_dataloader = DataLoader(test_data, batch_size=64)创建记录文档，名字为当下时间，将训练过程和结果保存并打印出来。
t = time.strftime('%Y-%m-%d_%H-%M-%S') # 切记中间符号不能用冒号，不然会报错
file = open('logs/{}.txt'.format(t), 'w')
train_data_size = len(train_data)
test_data_size = len(test_data)
print('The size of train_data:{}'.format(train_data_size))
file.write('The size of train_data:{}'.format(train_data_size)+'\n')
print('The size of train_data:{}'.format(test_data_size))
file.write('The size of train_data:{}'.format(test_data_size)+'\n')创建一些中间参数，配置损失函数和优化器，也可从官方文档中选择其他损失函数torch.nn.functional — PyTorch 1.11.0 documentation和优化器torch.optim — PyTorch 1.11.0 documentation
epoch = 10
train_step = 0
test_step = 0
vgg16 = VGG16() # 实例化网络
loss_fn = nn.CrossEntropyLoss()
learning_rate = 1e-2
# optimizer = optim.SGD(vgg16.parameters(), lr=learning_rate, momentum=0.9)
optimizer = optim.Adam(vgg16.parameters(), lr=learning_rate)
losses = [] # 用于存储损失值，便于后面画损失变化曲线图开始训练
for i in range(epoch):
print('-----epoch{}-----'.format(i))
file.write('-----epoch{}-----'.format(i)+'\n')
# 训练步骤开始
vgg16.train()
for data in train_dataloader:
images, targets = data
output = vgg16(images)
loss = loss_fn(output, targets)
# 优化器优化模型
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_step += 1
if train_step % 50 == 0:
print('train_step:{},loss:{}'.format(train_step, loss.item()))
file.write('train_step:{},loss:{}'.format(train_step, loss.item())+'\n')
losses.append(loss.item())
# 测试步骤开始
vgg16.eval()
accuracy = 0
with torch.no_grad():
for data in test_dataloader:
images, targets = data
output = vgg16(images)
current_acc = (output.argmax(1) == targets).sum()
accuracy += current_acc
acc = accuracy / test_data_size
print('-------eval accuracy:{}'.format(acc))
file.write('-------eval accuracy:{}'.format(acc)+'\n')
torch.save(tudui, 'checkpoints/vgg16_{}.pth'.format(i))
print('The model has been saved')
file.write('The model has been saved')
file.close()保存的记录文档和打印输出如下所示
画出损失曲线变化图
plt.subplot(1, 1, 1)
plt.plot(losses, label='Training Loss')
plt.title('Training Loss')
plt.legend()
plt.show()编写测试文件，将训练好的模型用于预测任意图片的类别。
import torch
from PIL import Image
from torchvision import transforms
image_path = 'imgs/dog.jpg' # 随意选择一张图片进行测试
image = Image.open(image_path)
image = image.convert('RGB')
transform = transforms.Compose([transforms.Resize(32, 32), transforms.ToTensor()]) # 将图片大小改为32×32，并转换为Tensor类型
image = transform(image)
classes = ['plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
model = torch.load('checkpoints/vgg16_30.pth') # 加载训练好的模型
image = torch.reshape(image, (1, 3, 32, 32))
model.eval()
with torch.no_grad():
output = model(image)
index = output.argmax(1)
pred = classes[index]
print(pred)