Fine-tune CLIP towards macro-f1 loss

[fkemeth]

See here:

https://stackoverflow.com/a/76842670/12438757

[HCA97]

I think we can close this ticket, @fkemeth what do you think?

[fkemeth]

I though about it recently, but I see an issue with the batch size. The macro-f1 only makes sense really if we have large enough batch sizes - for the head+backbone finetuning I cannot use large batch sizes because of GPU memory. But maybe if we fine tune the head later on this might help. But I agree, I am not sure if this will help much, so closing this ticket would be fine for me.

[fkemeth]

FIne tuned head on ce+f1 loss. Submitted model under

http://gitlab.aicrowd.com/hca97/mosquitoalert-2023-phase2-starter-kit/issues/140