Error when running run_clm.py on Python3.9/MacOS

[liyucheng09]

Environment info

• transformers version: 4.1.1
• Platform: macOS-11.0-arm64-arm-64bit
• Python version: 3.9.1
• PyTorch version (GPU?): 1.8.0a0+c20b916 (False)
• Tensorflow version (GPU?): not installed
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

Information

Model I am using (Bert, XLNet ...): GPT2

The problem arises when using:

• [yes ] the official example scripts: (give details below)

The tasks I am working on is:

• [no] an official GLUE/SQUaD task: language-modeling task; dataset: wikitext

To reproduce

Steps to reproduce the behavior:

1. install transformers from the master branch of version 4.1.1
2. run examples/language-modeling/run_clm.py
3. arguments are as following: --model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train --do_eval --output_dir test-clm/

/Users/liyucheng/miniforge3/bin/python /Users/liyucheng/projects/comments_generation/run_clm.py --model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train --do_eval --output_dir test-clm/
Traceback (most recent call last):
  File "/Users/liyucheng/projects/comments_generation/run_clm.py", line 388, in <module>
    main()
  File "/Users/liyucheng/projects/comments_generation/run_clm.py", line 145, in main
    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
  File "/Users/liyucheng/miniforge3/lib/python3.9/site-packages/transformers/hf_argparser.py", line 52, in __init__
    self._add_dataclass_arguments(dtype)
  File "/Users/liyucheng/miniforge3/lib/python3.9/site-packages/transformers/hf_argparser.py", line 85, in _add_dataclass_arguments
    elif hasattr(field.type, "__origin__") and issubclass(field.type.__origin__, List):
  File "/Users/liyucheng/miniforge3/lib/python3.9/typing.py", line 829, in __subclasscheck__
    return issubclass(cls, self.__origin__)
TypeError: issubclass() arg 1 must be a class

Process finished with exit code 1

This error is bizarre cause it only occurs on my OSX and I cannot reproduce it on my PC.

I think the main reason is about the decorator dataset, but I am not sure about that.

Thanks for any helps.

[LysandreJik]

It seems quite cryptic, but maybe @sgugger has already been confronted to that issue, so pinging him here.

[sgugger]

Never seen this before. There is some code in the HFArgumentParser to make it work with Python 3.9 that was added by @julien-c so maybe he has more insight?

[liyucheng09]

I want to provide more valuable information about this issue.

The field of the corresponding argument --model_name_of_path on my Mac/Python3.9 is like the following:

Field(name='model_name_or_path',type=typing.Optional[str],default=None,default_factory=<dataclasses._MISSING_TYPE object at 0x1065a6220>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({'help': "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."}),_field_type=_FIELD)

However, it is different on my PC/Python3.7.9.

Field(name='model_name_or_path',type=typing.Union[str, NoneType],default=None,default_factory=<dataclasses._MISSING_TYPE object at 0x00000227D9888A48>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({'help': "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."}),_field_type=_FIELD)

The critical change of it is the type attribute. The function in transformers/data_classes.py do not give type=typing.Optional[str] a appropriate solution.

But, I have no idea why the type attribute has that different value when I run it on Mac/Python3.9.1.

[sgugger]

#9479 will fix this I believe.

[LysandreJik]

Closed by #9479!