Pytorch 数据加载器：错误的文件描述符和 EOF > 0

问题描述

在使用由自定义数据集制作的 Pytorch 数据加载器进行神经网络训练期间，我遇到了奇怪的行为。数据加载器设置为workers=4、pin_memory=False。

大多数时候，训练都顺利完成。 有时，训练会随机停止，并出现以下错误：

1. OSError: [Errno 9] 错误的文件描述符
2. EOFError

看起来错误是在创建套接字以访问数据加载器元素期间发生的。 当我将工作人员数量设置为 0 时，错误消失，但我需要通过多处理来加速我的训练。 错误的根源可能是什么？谢谢 ！

python 和库的版本

Python 3.9.12、Pytorch 1.11.0+cu102
编辑：错误仅发生在集群上

错误文件的输出

Traceback (most recent call last):
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/multiprocessing/resource_sharer.py", line 145, in _serve
Epoch 17:  52%|█████▏    | 253/486 [01:00<00:55,  4.18it/s, loss=1.73]

Traceback (most recent call last):
  File "/my_directory/bench/run_experiments.py", line 251, in <module>
    send(conn, destination_pid)
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/multiprocessing/resource_sharer.py", line 50, in send
    reduction.send_handle(conn, new_fd, pid)
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/multiprocessing/reduction.py", line 183, in send_handle
    with socket.fromfd(conn.fileno(), socket.AF_UNIX, socket.SOCK_STREAM) as s:
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/socket.py", line 545, in fromfd
    return socket(family, type, proto, nfd)
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/socket.py", line 232, in __init__
    _socket.socket.__init__(self, family, type, proto, fileno)
OSError: [Errno 9] Bad file descriptor

    main(args)
  File "/my_directory/bench/run_experiments.py", line 183, in main
    run_experiments(args, save_path)
  File "/my_directory/bench/run_experiments.py", line 70, in run_experiments
    ) = run_algorithm(algorithm_params[j], mp[j], ss, dataset)
  File "/my_directorybench/algorithms.py", line 38, in run_algorithm
    data = es(mp,search_space,  dataset, **ps)
  File "/my_directorybench/algorithms.py", line 151, in es
   data = ss.generate_random_dataset(mp,
  File "/my_directorybench/architectures.py", line 241, in generate_random_dataset
    arch_dict = self.query_arch(
  File "/my_directory/bench/architectures.py", line 71, in query_arch
    train_losses, val_losses, model = meta_net.get_val_loss(
  File "/my_directory/bench/meta_neural_net.py", line 50, in get_val_loss
    return self.training(
  File "/my_directorybench/meta_neural_net.py", line 155, in training
    train_loss = self.train_step(model, device, train_loader, epoch)
  File "/my_directory/bench/meta_neural_net.py", line 179, in train_step
    for batch_idx, mini_batch in enumerate(pbar):
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/site-packages/tqdm/std.py", line 1195, in __iter__
    for obj in iterable:
  File "/my_directory/.local/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 530, in __next__
    data = self._next_data()
  File "/my_directory/.local/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1207, in _next_data
    idx, data = self._get_data()
  File "/my_directory/.local/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1173, in _get_data
    success, data = self._try_get_data()
  File "/my_directory/.local/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1011, in _try_get_data
    data = self._data_queue.get(timeout=timeout)
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/multiprocessing/queues.py", line 122, in get
    return _ForkingPickler.loads(res)
  File "/my_directory/.local/lib/python3.9/site-packages/torch/multiprocessing/reductions.py", line 295, in rebuild_storage_fd
    fd = df.detach()
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/multiprocessing/resource_sharer.py", line 58, in detach
    return reduction.recv_handle(conn)
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/multiprocessing/reduction.py", line 189, in recv_handle
    return recvfds(s, 1)[0]
  File "/my_directory/.conda/envs/geoseg/lib/python3.9/multiprocessing/reduction.py", line 159, in recvfds
    raise EOFError
EOFError

编辑：访问数据的方式

    from PIL import Image
    from torch.utils.data import DataLoader
    
    # extract of code of dataset
    
        class Dataset():
           def __init__(self,image_files,mask_files):
              self.image_files = image_files
              self.mask_files = mask_files
    
           def __getitem__(self, idx):
              img = Image.open(self.image_files[idx]).convert('RGB')
              mask=Image.open(self.mask_files[idx]).convert('L')
              return img, mask
    
    # extract of code of trainloader
      
        train_loader = DataLoader(
                        dataset=train_dataset,
                        batch_size=4,
                        num_workers=4,
                        pin_memory=False,
                        shuffle=True,
                        drop_last=True,
                        persistent_workers=False,
                    )

我终于找到了解决方案。将此配置添加到数据集脚本中可以正常工作：

import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')

默认情况下，共享策略设置为'file_descriptor'.

我尝试过一些解释的解决方案：

• 这个问题 https://github.com/pytorch/pytorch/issues/13246（增加共享内存，增加打开文件描述符的最大数量，每个时期结束时的 torch.cuda.empty_cache() ，...）
• and 这另一个问题 https://github.com/pytorch/pytorch/issues/11201，结果解决了问题

正如 @AlexMeredith 所建议的，该错误可能与某些集群使用的分布式文件系统（Lustre）有关。该错误也可能来自分布式共享内存。