About training stage III

[salvadog]

Thanks for sharing your great job! I've trained CodeFormer on my task for stage I and II. When testing on real images, the overall performance is good except for some corner cases (e.g. large poses and heavy shadow). I understand this is due to my limited dataset (~10k images), and the skipped features in stage III is indeed needed to complement the limited coverage of codebook. However, since I've trained stage I and II with w=0, the training of stage III quickly break down when I change w to 1, the discriminator is immediately too strong. Also, I fixed the quantizer and decoder and change the lr to 2e-5 in stage III. Can you give me some suggestions?

[salvadog]

And did you supervise with images generated with w=0 in stage III? I think fixing quantizer and decoder, and finetuning encoder only with w=1 will affect the originally good results generated with w=0.

[donghaoye]

@salvadog Hello, could you share the training codes?

[salvadog]

@sczhou Any suggestions on stages III training?

[salvadog]

@salvadog Hello, could you share the training codes?

Sorry, I can't share my training codes since I implemented it with some non-open-source framework.