What we expect Improved GAN and Mode GAN (with dynamic activation maximization) to learn.

What Mode GAN learns.

Run this example on cpu with:

python main.py --d_loss nll --g_loss activation_maximization --n_real_classes 10 --variance 2.0 --ngpu 0

Omit --ngpu to run on default GPU.

See:

python main.py

Don't forget to run visdom in a different screen.

Works on torch 1.8.0 with python 3.6.12.

Works on ROCM 3.10 also.

Originally written on torch 1.0.1 python 3.6.8 cuda 10.0, cudnn 7.5.0