search

TencentARC / Open-MAGVIT2

Open-MAGVIT2: Democratizing Autoregressive Visual Generation
Apache License 2.0
261 stars 5 forks source link

Noisy points in reconstruction results #12

Closed shinshiner closed 1 hour ago

shinshiner commented 3 days ago

Hi, thanks for your great reproduction, I've trained it in ImageNet. An issue is there are many noisy points in reconstruction images, especially for black background, as shown in the picture. Do you have any idea about this? I've trained the model for 2 epochs, and the loss seems to not decrease, does it mean the model has converged?

Thanks!

bad1 bad2

yyyouy commented 3 days ago

Maybe the epoch should be 150 (in this GitHub repo), and the paper shows the result which is trained with 270 epochs.

shinshiner commented 3 days ago

Maybe the epoch should be 150 (in this GitHub repo), and the paper shows the result which is trained with 270 epochs.

I'm not sure whether it is a common case in the early of training, or it has coverged to a bad local minimal. Since I observed the loss does not decrease after 2 epochs, and worry that there is no use even I train 1000 epochs. @RobertLuo1 could u provide some suggestions, is it common?

RobertLuo1 commented 2 days ago

Hi, Thanks for your interest in our work. I assume that perhaps you did not clamp the output into (-1, 1) and then map back to RGB space. At the early training stage, since the model is still learning, some values of pixels will be out of the range of (-1,1). I hope that will be helpful because we haven't encountered that problem during our training.