Github网址：https://github.com/diaoquesang/pytorchTutorials/tree/main
本教程创建于2023/7/31，几乎所有代码都有对应的注释，帮助初学者理解dataset、dataloader、transform的封装，初步体验调参的过程，初步掌握opencv、pandas、os等库的使用，😋纯手撸手写数字识别项目（为减少代码量简化了部分数据集相关操作），全流程跑通Pytorch！❤️❤️❤️
This tutorial was created on 2023/7/31. Almost all the code has corresponding comments, to help beginners understand dataset, dataloader, transform packaging, preliminary experience of the process of tuning the parameters, the initial grasp of the use of libraries such as opencv, pandas, os, etc., 😋 and get involved in this handwritten digit recognition project (we simplified some dataset-related operations in order to reduce the amount of code). Enjoy the whole process of running Pytorch!❤️❤️❤️如果喜欢本项目的话，留下你的⭐吧！
Give me a ⭐ if you like this project!
一、train.py
import torch
import torchvision
from torch import nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
import os
import cv2 as cv
import pandas as pd
class myDataset(Dataset): # 定义数据集类
def __init__(self, annotations_file, img_dir, transform=None,
target_transform=None): # 传入参数(标签路径,图像路径,图像预处理方式,标签预处理方式)
self.img_labels = pd.read_csv(annotations_file, sep=" ", header=None)
# 从标签路径中读取标签,sep为划分间隔符,header为列标题的行位置
self.img_dir = img_dir # 读取图像路径
self.transform = transform # 读取图像预处理方式
self.target_transform = target_transform # 读取标签预处理方式
def __len__(self):
return len(self.img_labels) # 读取标签数量作为数据集长度
def __getitem__(self, idx): # 从数据集中取出数据
img_path = os.path.join(self.img_dir, self.img_labels.iloc[
idx, 0])
# 从标签对象中取出第idx行第0列(第0列为图像位置所在列)的值(numberImages\5.bmp),并与图像路径(numberImages)进行拼接
image = cv.imread(img_path) # 用openCV的imread函数读取图像
label = self.img_labels.iloc[idx, 1] # 从标签对象中取出第idx行第1列(第1列为图像标签所在列)的值(5)
if self.transform:
image = self.transform(image) # 图像预处理
if self.target_transform:
label = self.target_transform(label) # 标签预处理
return image, label # 返回图像和标签
class myTransformMethod1(): # Python3默认继承object类
def __call__(self, img): # __call___，让类实例变成一个可以被调用的对象，像函数
img = cv.resize(img, (28, 28)) # 改变图像大小
img = cv.cvtColor(img, cv.COLOR_BGR2RGB) # 将BGR(openCV默认读取为BGR)改为RGB
return img # 返回预处理后的图像
# 测试函数
# print(pd.read_csv("annotations.txt", sep=" ", header=None))
# print(os.path.join("numberImages", pd.read_csv("annotations.txt", sep=" ", header=None).iloc[5, 0]))
# print(pd.read_csv("annotations.txt", sep=" ", header=None).iloc[5, 1])
# cv.imshow("1",cv.imread(os.path.join("numberImages", pd.read_csv("annotations.txt", sep=" ", header=None).iloc[5, 0])))
# cv.waitKey(0)
class myNetwork(nn.Module): # 定义神经网络
def __init__(self):
super().__init__() # 继承nn.Module的构造器
self.flatten = nn.Flatten(-3, -1)
# 继承nn.Module的Flatten函数并改为flatten,考虑到推理时没有batch(CHW),若使用默认值(1,-1)会导致C没有被flatten,故使用(-3,-1)
self.linear_relu_stack = nn.Sequential( # 定义前向传播序列
nn.Linear(3 * 28 * 28, 512),
nn.ReLU(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Linear(512, 10),
)
def forward(self, x): # 定义前向传播方法
x = self.flatten(x)
logits = self.linear_relu_stack(x)
return logits
# 设置运行环境,默认为cuda,若cuda不可用则改为mps,若mps也不可用则改为cpu
device = (
"cuda"
if torch.cuda.is_available()
else "mps"
if torch.backends.mps.is_available()
else "cpu"
)
print(f"Using {device} device") # 输出运行环境
model = myNetwork().to(device) # 创建神经网络模型实例
# 设置超参数
learning_rate = 1e-5 # 学习率
batch_size = 8 # 每批数据数量
epochs = 3000 # 总轮数
img_path = "./numberImages" # 设置图像路径
label_path = "./annotations.txt" # 设置标签路径
myTransform = transforms.Compose([myTransformMethod1(), transforms.ToTensor()])
# 定义图像预处理组合,ToTensor()中Pytorch将HWC(openCV默认读取为height,width,channel)改为CHW,并将值[0,255]除以255进行归一化[0,1]
myDataset = myDataset(label_path, img_path, myTransform) # 创建数据集实例
myDataLoader = DataLoader(myDataset, batch_size=batch_size,
shuffle=True)
# 创建数据读取器(可对训练集和测试集分别创建),batch_size为每批数据数量(一般为2的n次幂以提高运行速度),shuffle为随机打乱数据
def train():
# 根据epochs(总轮数)训练
for epoch in range(epochs):
totalLoss = 0
# 分批读取数据
for batch, (images, labels) in enumerate(myDataLoader):
# 数据转换到对应运行环境
images = images.to(device)
labels = labels.to(device)
pred = model(images) # 前向传播
myLoss = nn.CrossEntropyLoss() # 定义损失函数(交叉熵)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) # 定义优化器
loss = myLoss(pred, labels) # 计算损失函数
totalLoss += loss # 计入总损失函数
loss.backward() # 反向传播
optimizer.step() # 更新权重
optimizer.zero_grad() # 清空梯度
if batch % 1 == 0: # 每隔1个batch输出1次loss
loss, current = loss.item(), min((batch + 1) * batch_size,len(myDataset))
print(f"epoch: {epoch:>5d} loss: {loss:>7f} [{current:>5d}/{len(myDataset):>5d}]")
if epoch == 0:
minTotalLoss = totalLoss
if totalLoss < minTotalLoss:
print("······························模型已保存······························")
minTotalLoss = totalLoss
torch.save(model, "./myModel.pth") # 保存性能最好的模型
if __name__ == "__main__":
model.train() # 设置训练模式
train()二、eval.py
import torch
import torchvision
from torch import nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
import os
import cv2 as cv
import pandas as pd
class myTransformMethod1(): # Python3默认继承object类
def __call__(self, img): # __call___，让类实例变成一个可以被调用的对象，像函数
img = cv.resize(img, (28, 28)) # 改变图像大小
img = cv.cvtColor(img, cv.COLOR_BGR2RGB) # 将BGR(openCV默认读取为BGR)改为RGB
return img # 返回预处理后的图像
class myNetwork(nn.Module): # 定义神经网络
def __init__(self):
super().__init__() # 继承nn.Module的构造器
self.flatten = nn.Flatten(-3, -1)
# 继承nn.Module的Flatten函数并改为flatten,考虑到推理时没有batch(CHW),若使用默认值(1,-1)会导致C没有被flatten,故使用(-3,-1)
self.linear_relu_stack = nn.Sequential( # 定义前向传播序列
nn.Linear(3 * 28 * 28, 512),
nn.ReLU(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Linear(512, 10),
)
def forward(self, x): # 定义前向传播方法
x = self.flatten(x)
logits = self.linear_relu_stack(x)
return logits
if __name__ == "__main__":
model = torch.load("./myModel.pth").to("cuda") # 载入模型
model.eval() # 设置推理模式
myTransform = transforms.Compose([myTransformMethod1(), transforms.ToTensor()])
# 定义图像预处理组合,ToTensor()中Pytorch将HWC(openCV默认读取为height,width,channel)改为CHW,并将值[0,255]除以255进行归一化[0,1]
for i in range(10):
img = cv.imread("./numberImages/"+str(i)+".bmp") # 用openCV的imread函数读取图像
img = myTransform(img).to("cuda") # 图像预处理
print(torch.argmax(model(img)))
