mhamilton723 / FeatUp

Official code for "FeatUp: A Model-Agnostic Frameworkfor Features at Any Resolution" ICLR 2024
MIT License
1.38k stars 78 forks source link

implicit_upsampler train problem #25

Open xutanxing opened 7 months ago

xutanxing commented 7 months ago

Hi @mhamilton723, Thank you for sharing this great work!

When training my implicit upsampler, I encountered a concern regarding the negativity of the rec_loss, leading to an entirely negative value for the overall loss. I suspect this may be due to the logarithm operation.

Could you confirm whether the loss is intended to allow negative values, or if there might be an issue with my code?

Thank you.