【Bug】Pytorch RuntimeError: DataLoader worker (pid(s) 15904) exited unexpectedly

【Bug1】RuntimeError: DataLoader worker (pid(s) 15904) exited unexpectedly

知乎：https://zhuanlan.zhihu.com/p/712407893

环境

Windows 11
Python 3.10
torch 2.0.1
numpy 1.25.0

问题详情

在使用 PyTorch 的 DataLoader 时出现的错误。详情

RuntimeError:
        An attempt has been made to start a new process before the
        current process has finished its bootstrapping phase.

        This probably means that you are not using fork to start your
        child processes and you have forgotten to use the proper idiom
        in the main module:

            if __name__ == '__main__':
                freeze_support()
                ...
.....
RuntimeError: DataLoader worker (pid(s) ) exited unexpectedly

意思是，这是运行时错误，是由于在主进程完成初始化之前试图启动了新的进程导致的错误。

错误的代码示例

import torch
from torch.utils.data import Dataset, DataLoader

import numpy as np

# 假设我们有一些简单的数据
data = np.array([1,2,3,4,5,6,7])  # np, tensor 格式都可以
targets = torch.tensor([1,1,1,1,0,0,0])  # 标签

# 定义自定义数据集
class SimpleDataset(Dataset):
    def __init__(self, data, targets):
        self.data = data
        self.targets = targets

    def __getitem__(self, index):
        x = self.data[index]
        y = self.targets[index]
        return x, y

    def __len__(self):
        return len(self.data)
        
# 实例化数据集
dataset = SimpleDataset(data, targets)

# 创建 DataLoader, 如果启动多线程num_workers>=1，需要将启动代码放置在 if __name__ == "__main__": 下, 否则会报错
dataloader = DataLoader(dataset, batch_size=4, shuffle=True, num_workers=2)

# 使用 DataLoader 迭代数据
for i, (batch_data, batch_label) in enumerate(dataloader):
    print(f"Batch {i}: batch_data: {batch_data}, batch_label: {batch_label}")

解决方法

【方法1】（不推荐）

这是由于多线程加载数据使用不当使用的错误，因此可以设置为单线程即可，num_workers 不设置或设置为0。

dataloader = DataLoader(dataset, batch_size=4, shuffle=True, num_workers=0)

【方法2】

将涉及dataloader 的代码放置在if __name__ == "__main__": 下运行，修改代码如下

import torch
from torch.utils.data import Dataset, DataLoader
import numpy as np

# 定义自定义数据集
class SimpleDataset(Dataset):
    def __init__(self, data, targets):
        self.data = data
        self.targets = targets

    def __getitem__(self, index):
        x = self.data[index]
        y = self.targets[index]
        return x, y

    def __len__(self):
        return len(self.data)
        
def train():
    # 假设我们有一些简单的数据
    data = np.array([1,2,3,4,5,6,7])  # np, tensor 格式都可以
    targets = torch.tensor([1,1,1,1,0,0,0])  # 标签
    # 实例化数据集
    dataset = SimpleDataset(data, targets)

    # 创建 DataLoader, 如果启动多线程，需要将启动代码放置在 if __name__ == "__main__": 下, 否则会报错
    dataloader = DataLoader(dataset, batch_size=4, shuffle=True, num_workers=2)

    # 使用 DataLoader 迭代数据
    for i, (batch_data, batch_label) in enumerate(dataloader):
        print(f"Batch {i}: batch_data: {batch_data}, batch_label: {batch_label}")

if __name__ == "__main__":
     train()

参考

解决pytorch报错:RuntimeError: DataLoader worker (pid(s) ***, ***, ***, ***) exited unexpectedly - 知乎 (zhihu.com)