Add classifier_dropout param for CrossEncoder.

[PhilipMay]

This PR adds the option to set / change the classifier_dropout value. This way it is possible to set the dropout value of the classification head. This can be a very important hyperparameter for classification tasks (and also regression).

This is a non breaking change: When the classifier_dropout is None (the default) then nothing happens.

For more background please also see:

• add classifier_dropout to classification heads huggingface/transformers#12794
• Set dropout for ClassificationHead huggingface/transformers#12781
• Declaring classifier_dropout in model config but not using it huggingface/transformers#12792

[PhilipMay]

One unrelated test fails: FAILED tests/test_sentence_transformer.py::test_save_to_hub - TypeError: __init__() missing 1 required positional argument: 'pull_requests'

@tomaarsen do you know why?

[tomaarsen]

@PhilipMay I suspect that there's been a new release of huggingface_hub that changed the API for GitRefs that I use in one of the tests. I will look into this next week (as I'm off this week).

[PhilipMay]

@PhilipMay I suspect that there's been a new release of huggingface_hub that changed the API for GitRefs that I use in one of the tests. I will look into this next week (as I'm off this week).

Many thanks @tomaarsen ! 🚀

[PhilipMay]

@PhilipMay I suspect that there's been a new release of huggingface_hub that changed the API for GitRefs that I use in one of the tests. I will look into this next week (as I'm off this week).

Last release is from 20th Dec. 2023:

[PhilipMay]

I found the bug and reported it: huggingface/huggingface_hub#1956

For this PR I applied a workaround.

[PhilipMay]

CI is green - can be merged from my point of view. :-)

[tomaarsen]

This looks great, thanks! And feel free to share any trained models with me, I'm curious to see what you'll cook up :)