readme.md

Boundary Seeking Generative Adversarial Networks

Loss Function

• used GAN loss at D net
• used L2 loss at G net

Architecture Networks

• just 2 fc layers network for D/G nets
• In original BGAN paper, using simple 4 conv layers * 3 fc layers

Tensorboard

Elapsed time : 920s with GTX 1060 6GB x 1

Result

Name	Global Step 50k	Global Step 100k	Global Step 200k
BGAN

To-Do

• Add f-divergences
  • KL
  • Reverse-KL
  • JS
  • Squared-Hellinger
  • Pearson x^2
• Add Reinforce-based BGAN