No longer needlessly deepcopy the original model state

[tomaarsen]

Resolves #199

Hello!

Pull request overview

• Remove deepcopy of model.state_dict.
• Modify scripts/setfit/run_fewshot_multilabel.py to reflect that model.model_original_state has been removed.

Details

As mentioned by @Yongtae723 in #199, the model state is copied and then never used again, with exception of scripts/setfit/run_fewshot_multilabel.py. This is an unnecessary waste of memory, and hence we believe that these lines should be removed.

Furthermore, I modified scripts/setfit/run_fewshot_multilabel.py to perform the deepcopying in that script itself. Note that this is somewhat different than the original situation, as now these two lines are executed before the deepcopying is performed, rather than the deepcopy occurring first:

setfit/src/setfit/modeling.py

Lines 47 to 48 in b66760c

	if add_normalization_layer:
	self.model._modules["2"] = models.Normalize()

I suspect that this won't matter too much, but I'll leave that up to your judgement.
The script still works like before, e.g. using:

python scripts/setfit/run_fewshot_multilabel.py --sample_sizes 64 --datasets=go_emotions

• Tom Aarsen

[Yongtae723]

@tomaarsen
Hi Tom!

thank you for making excellent modifications and PR!

[tomaarsen]

And thanks to you for spotting this issue!

[blakechi]

Looks good to me. Nice work!

[lewtun]

Thanks for the fix @tomaarsen - this looks great 🔥 !

I agree this is a "bug" we should have removed before the release, but great to finally solve it now :)