PyTorch-progressive_growing_of_gans icon indicating copy to clipboard operation
PyTorch-progressive_growing_of_gans copied to clipboard

Published 20 hours ago • verified publisher icon github-pengge

It is unstable using your new release code in 64*64 resolution

Open akafen opened this issue 7 years ago • 6 comments

It is unstable using your newest release code in 64*64 resolution,why?

akafen avatar Dec 05 '17 06:12 akafen

Can't figure out why. In my experiment, when fading in 256x256, the model crashed.

github-pengge avatar Dec 06 '17 06:12 github-pengge

Try train_no_tanh.py with latest data.py script.

github-pengge avatar Dec 13 '17 04:12 github-pengge

yeah, same thing happened to me as well. In my case, the model collapsed very early on at 16*16 fade in. ( I used default value for all the hyper-parameters listed in the README)

--g_lr_max 1e-3 --d_lr_max 1e-3 --train_kimg 600 --transition_kimg 600 --beta1 0 --beta2 0.99 --gan lsgan --first_resol 4 --target_resol 256 --no_tanh

RooseveltAdvisors avatar Dec 21 '17 17:12 RooseveltAdvisors

Me too, the model coIlapsed very early. BTW, the model is trained on my own data set. Is there something important??

KylinYee avatar Dec 22 '17 08:12 KylinYee

Use train_no_tanh.py instead.

github-pengge avatar Dec 25 '17 09:12 github-pengge

Thx!! It works. I can successfully obtain meaningful images.

KylinYee avatar Dec 26 '17 09:12 KylinYee