SimCLR loss calculation seems deviating from original paper

[likenneth]

🐛 Bug

as title

To Reproduce

I the denominator shall substract exp(1/temp) rather than exp(1)

Code sample

Suspect at here:
https://github.com/PyTorchLightning/lightning-bolts/blob/9044ed143c7f1e20164de12208f451c7a8d9718f/pl_bolts/models/self_supervised/simclr/simclr_module.py#L284

Expected behavior

I haven't tested whether my proposed implementation would improve performance, I am suggesting any deviation from original paper be noted.

[github-actions]

Hi! thanks for your contribution!, great first issue!

[Moadab-AI]

Actually exp(1/Temperature) should be subtracted not exp(-Temperature)

[likenneth]

Actually exp(1/Temperature) should be subtracted not exp(-Temperature)

Exactly! I revised my comments.
BTW, I tried the new formula on my project, which is based on Pytorch Lightning SimCLR implementation, and found no drastic difference in performance.