Make progress bar in `SentenceTransformer.encode()` configurable

[lbhm]

I regularly use the show_progress_bar parameter in SentenceTransformer.encode() to track the progress of larger jobs. While the current functionality is already helpful, I would love to be able to configure the progress bar behavior. For example, I often use leave=False or dynamic_ncols=True (see the tqdm docs) in my self-created progress bars.

I understand that these parameters are not helpful in all situations, so I was wondering whether it would be acceptable to change the type of show_progress_bar to bool | dict[str, Any] | None = None. In addition to the current behavior, if the user passes a dict, it is interpreted as tqdm parameters and passed to the trange() call. Otherwise, the current default parameters are used.

I'm not sure if the suggested change has implications for other parts of the code base, so I wanted to raise this as a discussion point first. Since the behavior for bool or None values is unchanged, there should be no breaking behavior changes.

[tomaarsen]

Hello!

I totally understand the desire to update your tqdm, but I'm hesitant to place this functionality directly in ST. Perhaps a cleaner solution is to use either:

1. Use environment variables to override defaults (granted: you can't override 'call' parameters, i.e. ones that I explicitly define in ST), see env var override support tqdm/tqdm#1491
2. Override tqdm with a decorator or a partial, e.g. in support global config to override defaults tqdm/tqdm#614 (comment)

What do you think about something like this?

• Tom Aarsen

[lbhm]

Thank you for the quick response @tomaarsen.

The second option sounds nice and would reduce config repetition across my own usage of progress bars. However, I am not completely sure if this would work to modify ST's progress bars.

If I created a custom_tqdm.py in my library and overwrote tqdm.__init__ in there, would this also affect other modules (like ST) that import the "original" tqdm functions? Since I cannot modify the ST imports directly, I would somehow need to modify the global defaults for tqdm. Ideally, I would want to avoid env vars since I prefer having a self-contained config in the code.