8. pytorch 快速搭建、训练一个小型网络

目录

• 下载读取数据
• 搭建网络
• 准备日志、损失函数和优化器
• 网络的训练和测试
• 模型文件的保存
• 关闭日志并查看训练效果
• 完整的代码
• 注意事项

 

正文

  本篇博客中，我将快速搭建一个小型的网络，并对其进行训练、优化器调参，最后查看模型训练效果。
我将本次搭建网络分为一下几个部分

1. 下载、读取数据
2. 搭建网络
3. 准备日志、损失函数和优化器
4. 进行网络的训练与测试，
5. 模型文件的保存
6. 关闭日志并查看训练效果

回到顶部

下载读取数据

本篇博客所写代码使用 python，并且大量使用了 pytorch 第三方库，其中的 torvision.datasets.CIFAR10() 用于读取数据，torch.utils.data.DataLoader()用于封装遍历数据，倘若对这两种函数不熟悉的同学，可以看我前几篇博客。

# load the data and set the constant values
dataset_path = "../data_cifar10"
dataset_train = torchvision.datasets.CIFAR10(root=dataset_path, train=True, transform=torchvision.transforms.ToTensor(),
                                             download=True)
dataset_test = torchvision.datasets.CIFAR10(root=dataset_path, train=False, transform=torchvision.transforms.ToTensor(),
                                            download=True)
 
dataloader_train = DataLoader(dataset=dataset_train, batch_size=64)
dataloader_test = DataLoader(dataset=dataset_test, batch_size=64)
 
dataset_train_len = len(dataset_train)
dataset_test_len = len(dataset_test)

回到顶部

搭建网络

搭建网络的内容在前几篇博客已经写过了，不熟悉的同学请自行查看。

import torch
import torch.nn as nn
 
 
class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.model = nn.Sequential(
            nn.Conv2d(in_channels=3, out_channels=32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(2),
            nn.Conv2d(in_channels=32, out_channels=32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(2),
            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Flatten(start_dim=1, end_dim=-1),
            nn.Linear(in_features=1024, out_features=64),
            nn.Linear(in_features=64, out_features=10)
        )
 
    def forward(self, x):
        return self.model(x)
 
 
if __name__ == "__main__":
    test_data = torch.ones((64, 3, 32, 32))
    my_model = MyModel()
    data_proceed = my_model(test_data)
    print(data_proceed.shape)

回到顶部

准备日志、损失函数和优化器

网络搭建完成之后，我们实例化网络，并打开日志 SummaryWriter，准备好损失函数（交叉熵损失函数）和优化器（SGD），该部分也包括了设置一些常量

# define the net and constant
my_model = MyModel()
learning_rate = 1e-2
my_optimize = torch.optim.SGD(my_model.parameters(), lr=learning_rate)
my_loss_fn = torch.nn.CrossEntropyLoss()
train_step = 0
test_step = 0
max_epoch = 100
writer = SummaryWriter("./logs")

回到顶部

网络的训练和测试

网络的训练过程时，训练前加入 my_model.train()，测试前加入 my_model.test()
因为前几篇博客并没有讲到 train()和eval()函数，这里我贴一下官方帮助文档的翻译：

还望注意一下 train 和 test 的区别，是否会 optim.zero_grad()，损失函数数值backward()，和 optim.step()

# begin training and testing
for epoch in range(max_epoch):
    print("-------The {} Epoch is Running!-------".format(epoch))
 
    # train the data
    my_model.train()
    train_sum_loss = 0
    for images, targets in dataloader_train:
        outputs = my_model(images)
        train_loss = my_loss_fn(outputs, targets)
        train_sum_loss += train_loss.item()
        if train_step % 100 == 0:
            print(f"train {epoch}, step:{train_step}, train_loss{train_loss}")
        my_optimize.zero_grad()
        train_loss.backward()
        my_optimize.step()
 
        train_step += 1
 
    print(f"train {epoch}, train_epoch_loss{train_sum_loss}")
    writer.add_scalar("train epoch loss", train_sum_loss, epoch)
 
    # test the data
    my_model.eval()
    with torch.no_grad():
        test_sum_loss = 0
        predict_right_cnt = 0
        for images, targets in dataloader_test:
            output = my_model(images)
            test_loss = my_loss_fn(output, targets)
            test_sum_loss += test_loss.item()
            predict_right_cnt += (torch.argmax(output, dim=1) == targets).sum()
 
    writer.add_scalar(f"predict right rate", predict_right_cnt / dataset_test_len, epoch)
    print(f"test {epoch}, test_epoch_loss{test_sum_loss}")
    writer.add_scalar("test epoch loss", test_sum_loss, epoch)

回到顶部

模型文件的保存

之前的博客讲过模型保存的两种方式，这里采取官方推荐的第二种

    # save the model!!!!!!!
    torch.save(my_model.state_dict(), f"./project_models/train_model_1_{epoch}.pth")

回到顶部

关闭日志并查看训练效果

writer.close() # 关闭日志

因为 没有进行 dropout，感觉训练的有一点过拟合了我去。。。

回到顶部

完整的代码

项目结构

main.py，主文件

import torch
import torchvision
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
 
# load the data and set the constant values
dataset_path = "../data_cifar10"
dataset_train = torchvision.datasets.CIFAR10(root=dataset_path, train=True, transform=torchvision.transforms.ToTensor(),
                                             download=True)
dataset_test = torchvision.datasets.CIFAR10(root=dataset_path, train=False, transform=torchvision.transforms.ToTensor(),
                                            download=True)
 
dataloader_train = DataLoader(dataset=dataset_train, batch_size=64)
dataloader_test = DataLoader(dataset=dataset_test, batch_size=64)
 
dataset_train_len = len(dataset_train)
dataset_test_len = len(dataset_test)
 
 
class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.model = nn.Sequential(
            nn.Conv2d(in_channels=3, out_channels=32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(2),
            nn.Conv2d(in_channels=32, out_channels=32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(2),
            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Flatten(start_dim=1, end_dim=-1),
            nn.Linear(in_features=1024, out_features=64),
            nn.Linear(in_features=64, out_features=10)
        )
 
    def forward(self, x):
        return self.model(x)
 
 
# define the net and constant
my_model = MyModel()
learning_rate = 1e-2
my_optimize = torch.optim.SGD(my_model.parameters(), lr=learning_rate)
my_loss_fn = torch.nn.CrossEntropyLoss()
train_step = 0
test_step = 0
max_epoch = 100
writer = SummaryWriter("./logs")
 
# begin training and testing
for epoch in range(max_epoch):
    print("-------The {} Epoch is Running!-------".format(epoch))
 
    # train the data
    my_model.train()
    train_sum_loss = 0
    for images, targets in dataloader_train:
        outputs = my_model(images)
        train_loss = my_loss_fn(outputs, targets)
        train_sum_loss += train_loss.item()
        if train_step % 100 == 0:
            print(f"train {epoch}, step:{train_step}, train_loss{train_loss}")
        my_optimize.zero_grad()
        train_loss.backward()
        my_optimize.step()
 
        train_step += 1
 
    print(f"train {epoch}, train_epoch_loss{train_sum_loss}")
    writer.add_scalar("train epoch loss", train_sum_loss, epoch)
 
    # test the data
    my_model.eval()
    with torch.no_grad():
        test_sum_loss = 0
        predict_right_cnt = 0
        for images, targets in dataloader_test:
            output = my_model(images)
            test_loss = my_loss_fn(output, targets)
            test_sum_loss += test_loss.item()
            predict_right_cnt += (torch.argmax(output, dim=1) == targets).sum()
 
    writer.add_scalar(f"predict right rate", predict_right_cnt / dataset_test_len, epoch)
    print(f"test {epoch}, test_epoch_loss{test_sum_loss}")
    writer.add_scalar("test epoch loss", test_sum_loss, epoch)
 
    # save the model!!!!!!!
    torch.save(my_model.state_dict(), f"./project_models/train_model_2_{epoch}.pth")
 
writer.close()

回到顶部

注意事项

1. train 和 test 过程中，有着较大的区别，首先是 optim 是否 zero_grad()，loss是否进行backward()，optim 是否step()。
2. train 和 test 过程中，还应该注意模式的设置， my_model.train() 和 my_model.eval() 是选择不同的模式
3. test 过程中，为了减少硬件负担，提高运算速度，我们往往还会 加上 with torch.no_grad()，左右如下所示：

Author:luckylight(xyg)
Date:2021/11/13