Ask about the return loss "simclr loss + classification loss"

[KimSoybean]

Hi, I see the code in https://github.com/vturrisi/solo-learn/blob/8ab3c7ef7fdca8e0471c64700de3d9df60d7e47c/solo/methods/simclr.py#L188

I am confused that your code optimize the "unsupervision loss + supervised loss"? Or I misunderstand it?

[ankitpatnala]

Yes, even i got confused before. Yes, both loss are calculated simulatenously but the author uses detach https://github.com/vturrisi/solo-learn/blob/8ab3c7ef7fdca8e0471c64700de3d9df60d7e47c/solo/methods/base.py#L416

Detach function doesn't allow the class_loss to optimize (manipulate) the weights of encoder network.

[KimSoybean]

Yes, even i got confused before. Yes, both loss are calculated simulatenously but the author uses detach

https://github.com/vturrisi/solo-learn/blob/8ab3c7ef7fdca8e0471c64700de3d9df60d7e47c/solo/methods/base.py#L416

Detach function doesn't allow the class_loss to optimize (manipulate) the weights of encoder network.

Thanks a lot!

[pragyasrivastava0805]

How do they deal with the semi-supervised case where the class loss needs to be optimized?

[pragyasrivastava0805]

@ankitpatnala

[vturrisi]

Hi @pragyasrivastava0805, just override this function https://github.com/vturrisi/solo-learn/blob/f25dc1fccb3c3ed2dbd848138adc2d5594f671a4/solo/methods/base.py#L481-L496 in your method class removing the .detach(). In this case, the method will be optimized with the self-sup loss and the classification loss. If you need some extra variation of this, you will possibly need to override the _base_shared_step method.

[DonkeyShot21]

Afaik the semi-supervised experiments presented in self-supervised learning papers do not optimize the supervised loss at pre-training time. They just use a fraction of the labels for linear evaluation.