湖南太平洋建设集团网站,专业网站开发制作,简单的模板网站,上传网站需要什么软件Github网址#xff1a;https://github.com/diaoquesang/pytorchTutorials/tree/main 本教程创建于2023/7/31#xff0c;几乎所有代码都有对应的注释#xff0c;帮助初学者理解dataset、dataloader、transform的封装#xff0c;初步体验调参的过程#xff0c;初步掌握openc… 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 torchvisionfrom torch import nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torchvision import transformsimport os
import cv2 as cv
import pandas as pdclass myDataset(Dataset): # 定义数据集类def __init__(self, annotations_file, img_dir, transformNone,target_transformNone): # 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 , headerNone)# 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RGBreturn img # 返回预处理后的图像# 测试函数
# print(pd.read_csv(annotations.txt, sep , headerNone))
# print(os.path.join(numberImages, pd.read_csv(annotations.txt, sep , headerNone).iloc[5, 0]))
# print(pd.read_csv(annotations.txt, sep , headerNone).iloc[5, 1])
# cv.imshow(1,cv.imread(os.path.join(numberImages, pd.read_csv(annotations.txt, sep , headerNone).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 (cudaif torch.cuda.is_available()else mpsif torch.backends.mps.is_available()else cpu
)
print(fUsing {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_sizebatch_size,shuffleTrue)
# 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(), lrlearning_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lossloss, current loss.item(), min((batch 1) * batch_size,len(myDataset))print(fepoch: {epoch:5d} loss: {loss:7f} [{current:5d}/{len(myDataset):5d}])if epoch 0:minTotalLoss totalLossif totalLoss minTotalLoss:print(······························模型已保存······························)minTotalLoss totalLosstorch.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 torchvisionfrom torch import nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torchvision import transformsimport os
import cv2 as cv
import pandas as pdclass 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RGBreturn 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 logitsif __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)))三、其余资料详见Github
