SAGAN

Self-Attention Generative Adversarial Networks

Loss Function

• used hinge loss with D/G nets.

Architecture Networks

• Same with the SAGAN paper.

(even using image size 64, additional memory is needed :(. Over GTX 1080 is recommended!)

Tensorboard

Elapsed Time (64x64) : 3h 42m 18s with GTX 1060 6GB x 1

Elapsed Time (128x128) : 7h 55m 40s with GTX 1080 8GB x 1

Result

Name	Global Step 10k	Global Step 20k	Global Step 30k
SAGAN (64x64)
SAGAN (128x128)

To-Do