如何解决错误:RuntimeError:收到 0 项 ancdata

问题描述 投票:0回答:5

我有一个 torch.utils.data.DataLoader。我用以下代码创建了它们。

transform_train = transforms.Compose([
    transforms.RandomCrop(32, padding=4),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])

trainset = CIFAR100WithIdx(root='.',
                           train=True,
                           download=True,
                           transform=transform_train,
                           rand_fraction=args.rand_fraction)

train_loader = torch.utils.data.DataLoader(trainset,
                                           batch_size=args.batch_size,
                                           shuffle=True,
                                           num_workers=args.workers)

但是当我运行以下代码时出现错误。

train_loader_2 = []
for i, (inputs, target, index_dataset) in enumerate(train_loader):
    train_loader_2.append((inputs, target, index_dataset))

错误是

Traceback (most recent call last):
  File "main_superloss.py", line 460, in <module>
    main()
  File "main_superloss.py", line 456, in main
    main_worker(args)
  File "main_superloss.py", line 374, in main_worker
    train_loader, val_loader = get_train_and_val_loader(args)
  File "main_superloss.py", line 120, in get_train_and_val_loader
    for i, (inputs, target, index_dataset) in enumerate(train_loader):
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 804, in __next__
    idx, data = self._get_data()
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 771, in _get_data
    success, data = self._try_get_data()
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 724, in _try_get_data
    data = self.data_queue.get(timeout=timeout)
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/multiprocessing/queues.py", line 113, in get
    return _ForkingPickler.loads(res)
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/site-packages/torch/multiprocessing/reductions.py", line 284, in rebuild_storage_fd
    fd = df.detach()
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/multiprocessing/resource_sharer.py", line 58, in detach
    return reduction.recv_handle(conn)
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/multiprocessing/reduction.py", line 185, in recv_handle
    return recvfds(s, 1)[0]
  File "/home/C00423766/.conda/envs/dp/lib/python3.7/multiprocessing/reduction.py", line 161, in recvfds
    len(ancdata))
RuntimeError: received 0 items of ancdata

我想获取列表中的数据的原因是因为我想对样本重新排序。而且不是以随机的方式,而是以特定的方式。我怎样才能做到这一点?

python-3.x pytorch dataset dataloader pytorch-dataloader
5个回答
6
投票

我的代码也面临类似的问题,并且基于一些讨论(检查#1#2#3)。我使用

ulimit -n 2048
来增加进程可以拥有的文件描述符的最大数量。您可以在此处阅读有关 ulimit 的更多信息。

关于这个问题 - 讨论表明它必须与 pytorch 的分叉多处理代码有关。

关于问题的第二部分,如何重新排序数据加载器 - 您可以参考此答案


1
投票
import resource
rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
resource.setrlimit(resource.RLIMIT_NOFILE, (2048, rlimit[1]))

1
投票

我能够修复它

sudo vim /etc/security/limits.conf
# TODO add `* soft nofile 4096` to the end of the file without a `#`.
sudo vim /etc/pam.d/common-session
# TODO add `session required pam_limits.so` to the end of the file without a `#`
# Log out and log back in and you should be good. Check with ulimit -n
# NOTE: also need to restart ssh/any screen sessions as they remember the fd limit.

0
投票

在 Ubuntu 上,您需要执行以下操作来为所有用户解决此问题:

添加行

session required pam_limits.so 

common-session*
文件(有多个!)

$ sudo nano /etc/pam.d/common-session
$ sudo nano /etc/pam.d/common-session-noninteractive

然后添加行

* soft nofile 4096
* hard nofile 4096

limits.conf
文件

 $ sudo nano /etc/security/limits.conf

重新登录后您应该会看到

$ ulimit -a
...
open files                          (-n) 4096
...

这应该可以永远解决你的 Ubuntu 机器上的问题。

看到


0
投票

如果由于某种原因(例如不是管理员),您无法使用

ulimit
更改文件描述符的限制,您也可以使用

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

导入火炬后(在做其他事情之前)

(来源此 pytorch 问题评论

© www.soinside.com 2019 - 2024. All rights reserved.