Incorrect type annotation & slightly unintuitive functionality of `available_model_names()` - Githubissues

TRI-ML / prismatic-vlms

A flexible and efficient codebase for training visually-conditioned language models (VLMs)

MIT License

425 stars 194 forks source link

Incorrect type annotation & slightly unintuitive functionality of `available_model_names()` #7

Closed RylanSchaeffer closed 4 months ago

RylanSchaeffer commented 6 months ago

The available_model_names() function https://github.com/TRI-ML/prismatic-vlms/blob/main/prismatic/models/load.py#L32C5-L32C26 promises to return a list of strings but it actually returns a list of tuples (the keys and values of the global registry). Consequently,

the type annotation of the returned value List[str] is incorrect
Usages like model_name not in model_available_names() will evaluate to true because the model_name is compared against a tuple.

I'm happy to make whatever changes but I don't feel sufficiently well informed to know what the preferred fix is.

RylanSchaeffer commented 6 months ago

I'm wondering if available_models() is what available_model_names() meant to be?