CoGAN-tensorfow is generating nothing but noise on other dataset

andrewliao11 / CoGAN-tensorflow

Implement Coupled Generative Adversarial Networks in Tensorflow

101 stars 28 forks source link

CoGAN-tensorfow is generating nothing but noise on other dataset #5

Open ArghyaPal opened 7 years ago

ArghyaPal commented 7 years ago

Hi, Thank you for this neat and clean code of CoGAN. I have one comment:

The architecture you have given is giving very good results on MNIST dataset. But, generating nothing but noise on other datasets like: CIFAR 10, Fashion MNIST, LSUN and SVHN.

I will be very much obliged if you share with me necessary hacks to get good results on those datasets too.

With Best Regards, Arghya Pal.

andrewliao11 commented 7 years ago

Hi @ArghyaPal , directly changes to other dataset might lead to unexpected results. There are some factors that you might have to take care

different image size -> different configuration of the generator (e.g. Checkerboard Artifacts in Deconvolution layer)
different complexity of the dataset (e.g. MNIST -> Fashion MNIST) -> might need different number of gradient updates for G and D

Welcomed for any pull request 😄