Auxiliary Classifier Generative Adversarial Networks

Loss Function

• used sce_loss with D/G nets
• used softce_loss with C nets

Architecture Networks

• Same with the AC-GAN paper.
• But, i just used hyper-parameters like weight initializer, etc...

DIFFS	AC-GAN Paper	ME
weight initializer	Isotropic Gaussian	HE Initializer
Activation noise std	0 ~ 0.2	None

Isotropic Gaussian parameters : (µ = 0, σ = 0.02)
HE Initializer parameters : (factor = 1, FAN_AVG, uniform)

Tensorboard

Elapsed Time : 1h 17m 56s with GTX 1060 6GB x 1

Result

Name	Global Step 10k	Global Step 25k	Global Step 50k
ACGAN

To-Do

• prevent model collapse. Sometimes it happens... :(