Parse arguments with HfArgumentParser

[chocoded]

I was using HfArgumentParser to parse SentenceTransformerTrainingArguments and met the following error:

parser = HfArgumentParser(SentenceTransformerTrainingArguments)`
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/admin/.local/lib/python3.10/site-packages/transformers/hf_argparser.py", line 137, in __init__
    self._add_dataclass_arguments(dtype)
  File "/home/admin/.local/lib/python3.10/site-packages/transformers/hf_argparser.py", line 264, in _add_dataclass_arguments
    self._parse_dataclass_field(parser, field)
  File "/home/admin/.local/lib/python3.10/site-packages/transformers/hf_argparser.py", line 172, in _parse_dataclass_field
    field.type.__args__[0] if isinstance(None, field.type.__args__[1]) else field.type.__args__[1]
  File "/opt/conda/envs/python3.10/lib/python3.10/typing.py", line 994, in __instancecheck__
    return self.__subclasscheck__(type(obj))
  File "/opt/conda/envs/python3.10/lib/python3.10/typing.py", line 997, in __subclasscheck__
    raise TypeError("Subscripted generics cannot be used with"
TypeError: Subscripted generics cannot be used with class and instance checks

This error is due to that SentenceTransformerTrainingArguments.prompts has a type of dict[str, dict[str, str]] | dict[str, str] | str | None, while in _parse_dataclass_field, field.type.__args__[1] got a generic type of dict[str, str], and isinstance(None, dict[str, str]) caused the error. Could you fix this? thanks.

[tomaarsen]

Hello!

Thanks for reporting.
Based on https://github.com/huggingface/transformers/blob/4c1388f48e72dde96b666f8fe24f74bf4056e410/src/transformers/hf_argparser.py#L167-L171
it seems like the HfArgumentParser only accepts one type per argument (and optionally None). The new prompts parameter has 3 types plus None, so it might be incompatible. I think this is a limitation by argparse itself. I'm also not sure if you can load dictionaries with argparse.

I'm not sure how to tackle this other than e.g. updating the type for prompts to str | None or something. Looking into this some more now.

• Tom Aarsen

[Hypothesis-Z]

Same issue.

Overwriting the type seems to work.

@dataclass
class MySentenceTransformerTrainingArguments(SentenceTransformerTrainingArguments):
    prompts: Optional[str] = field(
        default_factory=dict
    )

[tomaarsen]

Thanks for providing a quick fix @Hypothesis-Z. Perhaps the best approach is to just use Optional[str] in the actual implementation. The other typings are also described in the docs, and I'd like to avoid errors like these.

I'll open a PR.

• Tom Aarsen