Closed PDomonkos closed 2 years ago
Thanks for the tip regarding pytorch version.
Re: division by 2, I don't think it was left out intentionally. I'll take a closer look later.
@KevinMusgrave Hi! Actually I'll take a look at this since I contributed to this loss function. It's a bit late in my time zone right now so I'll take a look tomorrow.
Thanks @cwkeam!
Fixed in v1.3.1
Hi,
The torch.cov in the covariance_loss function requires torch >= 1.10. Replacing:
Another remark is that in the reference implementation, the variance loss is divided by 2: https://github.com/facebookresearch/vicreg/blob/5e7b38f4586384bbb0d9a035352fab1d8f03b3b4/main_vicreg.py#L207 Was it intentionally left out, or am I missing something here?