BMSG-GAN icon indicating copy to clipboard operation
BMSG-GAN copied to clipboard

Published 20 hours ago • verified publisher icon akanimax

Model collapse

Open dance-ing opened this issue 5 years ago • 5 comments
trafficstars

Why are all the images generated by me using the trained network exactly the same? I observed the training process, and the images generated using different noises are still somewhat different, but gradually they become the same image. What is the reason? Is my training data too small?

dance-ing avatar Jun 05 '20 14:06 dance-ing

I faced the same problem even trying to change the GAN structure, add batch normalization,.. Did you find the answer yet?

nguyenpbui avatar Jun 16 '20 15:06 nguyenpbui

sorry, not yet...................

dance-ing avatar Jun 20 '20 14:06 dance-ing

I'm facing the same issue while trying to train my own dataset. What is the size of your dataset?

nihadguluzade avatar Jun 22 '20 12:06 nihadguluzade

400 images from LFW dataset. But I'm using celaba dataset now.

nguyenpbui avatar Jun 23 '20 03:06 nguyenpbui

Hi guys! imho, 400 images is an extremely small dataset and mode collapse is expected. I'd suggest trying the data - augmentation suggested in https://arxiv.org/abs/2006.06676 rather than changing the network architecture. There also a very informative video about the paper here. Hope this helps!

Cheers :beers:!

akanimax avatar Sep 06 '20 10:09 akanimax