markdjwilliams

Substituding `GMSDLoss` for `SRSIMLoss` also raises the exact same exception, with autograd also flagging the line `return torch.sqrt(torch.sum(grads ** 2, dim=-3, keepdim=True))` as being the source.

Thank you, please let me know if this cannot be reproduced on your side.

Thank you for looking at this so quickly! I might be testing the pull request incorrectly, but I think the problem remains. I see the change added a "r" literal...

Changing the python flags to `-B -Wall` seems to make this repro more reliably for me.

Thanks again, this is now good with your latest changes.

It turned out to be easier than I imaged to create a repro case. The scene is a single sphere, with the camera close enough to fill the entire image....

Thank you, that appears to help in the repro case (and reveals some issues with how I was handling devices) but in the context of my original example there's now...

Thank you so much, I'll update to 0.4.14 and try again!

Thank you this is working well in 0.4.14 - I just had to remove the bit in ```pyredner/image.py``` which attempts to download freeimage (I don't have unrestricted web access within...

Is there a good workaround for this?