About train progress about cyclegan-keras

[simontomaskarlsson]

Hi @lindalibjchn,

How many epochs it takes for the generators to learn the translation good enough to see some kind of result depends a lot on the specific image domains, i.e. A and B. From my experience you could usually see some results after about 50 epochs, and after 100 only minor details where improved by the generators.

If I understand it correctly your examples are after three epochs (?) and my guess is that more training is needed.

[simontomaskarlsson]

Hi @VirtualAI,

Using the settings as @lindalibjchn is a good choice I think.

From my research I found that identity learning helped with preserving for example colors of cars but could also leed to other problems if applied too much. As for resize convolution it did not effect the results in any major way.

PatchGAN in discriminators will almost always improve the results I believe.

[nalibjchn]

I am using two options according to the original paper.

• identity learning
• PatchGAN in discriminators
• reflection padding

[VirtualAI]

Hello,
what options to lead to best results?
Do you recommend something out of

-dentity learning (on different modulus of training iterations)
-PatchGAN in discriminators
-Multi-scale discriminators
-Resize convolution in generators