transformers.__spec__ returning None. Causing downstream import errors

[prikmm]

Environment info

• transformers version: Tried on 4.6.1(current default kaggle version)/4.8.1/4.8.2 and 4.9.1
• Platform: Colab/Kaggle/ My Local Runtime
• Python version: 3.7.11

Who can help

Information

This is causing downstream import errors, like right now I am not able to import lightning-flash properly as it uses __spec__ in order to find the availability of transformers.

ValueError                                Traceback (most recent call last)
<ipython-input-3-76e523923a79> in <module>
      5 print(transformers.__version__)
      6 print(transformers.__spec__)
----> 7 from flash import Trainer
      8 #from flash.core.data.utils import download_data
      9 #from flash.text import SummarizationData, SummarizationTask

/opt/conda/lib/python3.7/site-packages/flash/__init__.py in <module>
     16 
     17 from flash.__about__ import *  # noqa: F401 F403
---> 18 from flash.core.utilities.imports import _TORCH_AVAILABLE
     19 
     20 if _TORCH_AVAILABLE:

/opt/conda/lib/python3.7/site-packages/flash/core/utilities/imports.py in <module>
     75 _PYTORCHVIDEO_AVAILABLE = _module_available("pytorchvideo")
     76 _MATPLOTLIB_AVAILABLE = _module_available("matplotlib")
---> 77 _TRANSFORMERS_AVAILABLE = _module_available("transformers")
     78 _PYSTICHE_AVAILABLE = _module_available("pystiche")
     79 _FIFTYONE_AVAILABLE = _module_available("fiftyone")

/opt/conda/lib/python3.7/site-packages/flash/core/utilities/imports.py in _module_available(module_path)
     36     """
     37     try:
---> 38         return find_spec(module_path) is not None
     39     except AttributeError:
     40         # Python 3.6

/opt/conda/lib/python3.7/importlib/util.py in find_spec(name, package)
    112         else:
    113             if spec is None:
--> 114                 raise ValueError('{}.__spec__ is None'.format(name))
    115             return spec
    116 

ValueError: transformers.__spec__ is None

To reproduce

import transformers
print(transformers.__version__)
print(transformers.__spec__)

4.9.1
None

Expected behavior

Properly defined __spec__

[sgugger]

__spec__ is used by the Python import system internally, I am not reading anywhere that it should be defined manually by the package creators. If you have more resources about this I'm happy to look into what we could add, but a quick Google search yields nothing.

[prikmm]

__spec__ is used by the Python import system internally, I am not reading anywhere that it should be defined manually by the package creators. If you have more resources about this I'm happy to look into what we could add, but a quick Google search yields nothing.

My bad, at the time of error I found this issue on tensorflow/tensorflow#30028, and thought it was the same. After reading this, I somewhat understood the the functionality of __spec__.:thumbsup:

[laurahanu]

@sgugger I'm also getting the same error with the latest transformers version (4.9.2) when I'm trying to use torch.hub to load a model that has transformers as a dependency. It seems that torch.hub tries to check if dependencies exist by verifying that transformers.__spec__ is not None (source code here) resulting in an error otherwise.

Before I was using an older version of transformers (3.9.2) that returned a ModuleSpec object for transformers.__spec__ so loading the same model with torch.hub worked, just wondering why this has changed and whether it should be defined?

[laurahanu]

After investigating this further, it does seem particular to the transformers library that __spec__ returns None after importing it (other libraries still return something without having it explicitly defined).

Although it does seem that normally python's import system handles __spec__ internally and it does not need to be defined manually, it should return something automatically and not doing so could cause downstream problems e.g. when checking that dependencies exist:

Looks like the difference lies in whether transformers is manually imported or not:

In [1]: import importlib

In [2]: importlib.util.find_spec("transformers") is not None
Out[2]: True

In [3]: import transformers

In [4]: importlib.util.find_spec("transformers") is not None
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-4-6fdb35471f82> in <module>
----> 1 importlib.util.find_spec("transformers") is not None

~/opt/miniconda3/envs/pt/lib/python3.8/importlib/util.py in find_spec(name, package)
    112         else:
    113             if spec is None:
--> 114                 raise ValueError('{}.__spec__ is None'.format(name))
    115             return spec
    116

ValueError: transformers.__spec__ is None

This looks like something specific to the transformers package though, it doesn't happen e.g. with numpy:

In [5]: importlib.util.find_spec("numpy") is not None
Out[5]: True

In [6]: import numpy

In [7]: importlib.util.find_spec("numpy") is not None
Out[7]: True

[pratikchhapolika]

How to solve this issue?

[LysandreJik]

This issue should be solved in transformers versions v4.10.x

[talhaanwarch]

This issue should be solved in transformers versions v4.10.x

i tried transformers-4.15.0 and error is still there