[FR] Use normal dict instead of `**model_kwargs` in `_from_pretrained`

The _from_pretrained function of the modeling module has an argument called **model_kwargs. See here:

https://github.com/huggingface/setfit/blob/fa1021d2355f0cb3a2c85732ee7ffe44b0cef0d1/src/setfit/modeling.py#L337

IMO this should be changed. The problem I see is that it “collects” all keyword arguments from the function that are not defined. Later it only handles those assigned to head_params. This then swallows all other keyword args that might have happened because of a mistype for example. This would then be a bug very hard to find.

My suggestion is to just expect a dict called logistic_regression_kwargs head_kwargs (so just a dict and no **). This can then just be passed like so:

                    clf = LogisticRegression(**head_kwargs)

and

                    model_head = SetFitHead(**head_kwargs)

What do you think?

@lewtun and @blakechi and all others…

[blakechi]

Hi @PhilipMay,

I remember @lewtun asked a related problem in here in #149. So I think maybe exposed arguments will be preferred in the future.

But if we put that aside a bit, would it be more general if we use head_params instead of logistic_regression_kwargs? Since it’s for two different versions (sklearn and pytorch) of heads.

[PhilipMay]

Ahh I see. Very nice to have a 2nd maintainer. 😃 Many thanks.