share log.txt

[romanvey]

@zhanglonghao1992 could you provide your log.txt from your training? I think it would be beneficial for people to understand are they going in the right direction or not.

[Vijayue]

ye! I hope so, the training process is really hard. 😢😢

[zhanglonghao1992]

@romanvey @Vijayue I'll release the checkpoint and the training log file (if I can still find it - -!).

[romanvey]

Thank you!

[vinayak015]

log.txt

I was able to train for an epoch taking @zhanglonghao1992 checkpoints and the losses and the visualisation looks like this. The expressions of driving is not matching with the source. I think, either expression network or generator needs more tuning.

I was able to train with the 2D flow and 2D features, but getting similar results

[zhanglonghao1992]

log.txt @vinayak015 @romanvey @Vijayue

[vinayak015]

Thank you @zhanglonghao1992, In FOMM issues they mention the perceptual should be around 80 (although you reached 92). I think we should just train for more epochs.

[zhanglonghao1992]

@vinayak015 For now, my perceptual loss can only converge to about 90.

[romanvey]

Thank you @zhanglonghao1992 ! Yeah, I had the same perceptual loss and longer training helped me to increase quality.

[vinayak015]

It's mentioned here: https://github.com/AliaksandrSiarohin/first-order-model/issues/460#issuecomment-897689413