PeterouZh / CIPS-3D

3D-aware GANs based on NeRF (arXiv).
MIT License
610 stars 60 forks source link

Problem about reproducing the results #13

Closed 0three closed 2 years ago

0three commented 2 years ago

Hi, PeterouZh,

I'm reproducing your results at the same pace with you. Honestly speaking, this model takes about 40 hours to reach 64x64 at FID 15.97 with 8 A100 gpus. While I change the resolution to 128x128, the FID reach to 23.58. I'm still traning it and it reach FID 20.03 yet.

How can this model reach FID 6.XX as you described in paper? Do we miss some key things? It looks that this model can only reach 10+ FID in 256 resolutions because the performance increases very lowly when the FID reach 16 at 64x64 resolution.

By the way, I try to reproduce your results few weeks ago but I met problems about moxing. Does moxing provide very important tricks for this work?

0three commented 2 years ago

Thanks for your kind share and your hard work on reproducing your work in pure pytorch!

PeterouZh commented 2 years ago

The FID is not a standard FID during training, and it is calculated using only 2048 real images and fake images, taking into account the training efficiency.

The FID of 6.XX described in the paper is a standard FID, which is calculated using all training images and 50,000 generated images.

0three commented 2 years ago

Thank you for your professional answer! Let me regain my confidence. LOL

0three commented 2 years ago

I've continue to train r128 for a day. The FID cannot drop and are around 19.5 right now.

Surprisingly, I test the FID on the whole FFHQ dataset and 50,000 generated images and get FID 6.93 in 128x128 resolution.

It looks like that the training FID will take a very long time to reach 16.

PeterouZh commented 2 years ago

It seems that the 128 experiment did not meet expectations. I am currently debugging.

0three commented 2 years ago

Hi, I tried current sh script for training but failed.

0three commented 2 years ago

May you provide a detail readme for current sh scripts?