导入torchvision时报ImportError：无法从torch._ops导入OpOverload

问题场景 #

运行以下代码时触发ImportError：

from torchvision import datasets,transforms

错误详情 #

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
Cell In[11], line 5
      3 import pickle
      4 import numpy as np
----> 5 from torchvision import datasets,transforms
      6 from PIL import Image
      7 from torch.utils.data import Dataset, DataLoader

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torchvision/__init__.py:10
      7 # Don't re-order these, we need to load the _C extension (done when importing
      8 # .extensions) before entering _meta_registrations.
      9 from .extension import _HAS_OPS  # usort:skip
---> 10 from torchvision import _meta_registrations, datasets, io, models, ops, transforms, utils  # usort:skip
     12 try:
     13     from .version import __version__  # noqa: F401

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torchvision/_meta_registrations.py:4
      1 import functools
      3 import torch
----> 4 import torch._custom_ops
      5 import torch.library
      7 # Ensure that torch.ops.torchvision is visible

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/_custom_ops.py:4
      1 # mypy: allow-untyped-defs
      2 import inspect
----> 4 from torch._custom_op.impl import (
      5     _custom_op_with_schema,
      6     _find_custom_op,
      7     infer_schema,
      8     parse_qualname,
      9     validate_namespace,
     10 )
     11 from torch.library import get_ctx
     14 __all__ = [
     15     "custom_op",
     16     "impl",
   (...)
     20     "impl_backward",
     21 ]

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/_custom_op/impl.py:14
     12 import torch
     13 import torch._C as _C
---> 14 import torch.library as library
     15 from torch.library import get_ctx
     17 from .autograd import autograd_kernel_indirection, construct_autograd_kernel

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/library.py:13
     10 from typing_extensions import deprecated
     12 import torch
---> 13 import torch._library as _library
     14 from torch._library.custom_ops import (
     15     _maybe_get_opdef,
     16     custom_op,
     17     CustomOpDef,
     18     device_types_t,
     19 )
     20 from torch._library.infer_schema import infer_schema  # noqa: F401

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/_library/__init__.py:1
----> 1 import torch._library.autograd
      2 import torch._library.fake_impl
      3 import torch._library.simple_registry

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/_library/autograd.py:9
      6 from torch import _C, _ops, autograd, Tensor
      7 from torch.utils import _pytree
----> 9 from . import utils
     12 class InfoProtocol(Protocol):
     13     _backward_fn: Optional[Callable]

File ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/_library/utils.py:9
      7 import torch
      8 from torch import _C, _utils_internal
----> 9 from torch._ops import OpOverload
     12 @dataclasses.dataclass
     13 class Kernel:
     14     """Models a (function, source location)"""

ImportError: cannot import name 'OpOverload' from 'torch._ops' (/root/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/_ops.py)

已尝试无效操作 #

• 执行 pip uninstall torch torchvision 后重新运行 pip3 install torch torchvision torchaudio，错误依旧。

解决方案 #

1. 彻底清理残留后重装 #

卸载后手动删除torch相关残留文件，避免旧版本文件干扰：

pip uninstall -y torch torchvision torchaudio
rm -rf ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch*
rm -rf ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torchvision*
rm -rf ~/miniconda3/envs/pytorch/lib/python3.10/site-packages/torchaudio*

之后根据你的CUDA版本安装对应兼容的torch版本（示例为CUDA 11.8）：

pip3 install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

2. 验证环境一致性 #

• 确保激活正确的conda环境，避免跨环境安装导致的路径混乱：

  conda activate pytorch
  

• 检查torch版本及模块可用性：

  import torch
  print(torch.__version__)
  from torch._ops import OpOverload  # 测试是否能正常导入
  

  如果导入失败，说明当前torch版本不支持OpOverload，需更换至torch 1.13及以上的兼容版本。

3. 改用conda安装 #

若pip安装仍有问题，尝试用conda安装以保证依赖兼容性：

conda uninstall -y torch torchvision torchaudio
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

内容的提问来源于stack exchange，提问作者Hongtauo