Table 3.

Hyperparameters used in the experiments.

Hyperparameter	Value
Batch size ($n$)	64
CelebA resolutions ($h\times w$)	64 $\times$ 64, 128 $\times$ 128
CIFAR-10 resolution ($h\times w$)	32 $\times$ 32
Channel unit ($c$)	64
Regularization coefficient for training ($\alpha$)	12,288
Regularization coefficient for detecting ($\alpha$)	1
Adam (${\beta }_1,{\beta }_2$)	0.9, 0.999
Proportional gain (${\lambda }_k$)	0.001
Diversity ratio ($\gamma$)	0.5
Learning rate	0.001
Total global steps	3000,000