Have you tested the FID for CelebaHQ?

tqch / ddpm-torch

Unofficial PyTorch Implementation of Denoising Diffusion Probabilistic Models (DDPM)

MIT License

200 stars 35 forks source link

Have you tested the FID for CelebaHQ? #8

Closed zhuyu-cs closed 1 year ago

zhuyu-cs commented 1 year ago

Hi, thanks for your great work for implementing the DDPM baseline. I'm curious about the FID in CelebaHQ, since I have got 21.64 which may be too higher for DDPM. So if possiable, please let me know the correctness of the above results. Thanks!

tqch commented 1 year ago

Thank you for taking the time to verify the CelebA-HQ model checkpoint and report the performance mismatch here! After inspecting the code as well as the released checkpoint, I have identified some potential causes of the performance degradation. I am currently working on addressing these issues and plan to release an update to the repository in the near future.

tqch commented 1 year ago

It turned out that the previous CelebA-HQ checkpoint released by this repo was under-trained. I have updated the new checkpoint (1200 epoch model [link]), whose FID is ~8.9.