DCGAN-tensorflow icon indicating copy to clipboard operation
DCGAN-tensorflow copied to clipboard

Published 20 hours ago • verified publisher icon carpedm20

Training good till epoch 20 then goes horribly bad...

Open quintendewilde opened this issue 6 years ago • 3 comments

I had these settings CelebA 265 input 128 output 25 epochs --crop --train and everything was going fine till epoch 20 then things got very badly? Anybody an idea what's happening and if I can redo it from epoch 19?? Waste of amazon EC2 GPU time and money otherwise?

train_20_2198 train_20_2298 train_20_2398 train_21_1033

quintendewilde avatar Mar 04 '18 00:03 quintendewilde

What are the epochs for each of these pictures? The 3rd one looks like a very early epoch.

benckx avatar Mar 10 '18 20:03 benckx

It is mode collapse.

mode collapse

See this tutorial for instance.

woctezuma avatar Feb 15 '19 13:02 woctezuma

I believe that the mode collapse may be caused by updating the generator twice for every one update of the discriminator. Reverting to 1 to 1 training can solve this from my experience.

greerviau avatar Apr 18 '19 12:04 greerviau