Hi, I trained your model but result not very good

[lucasjinreal]

I run apple - > orange but when transfer apple to orange the result is very pool, using the default image size, I also trained the original pytorch version, it generates just nice picture, is I am not training enough epoch?

[vanhuyz]

Thank you for your information. I've just realized that my implementation of PatchGAN was incorrect and I'm working on it.

Regards,

[lucasjinreal]

Thanks, this work is very meaningfull, cause original can not even predict single image, If you worked out, update repo let me know, I'd like to follow this project.

[cpury]

Hey man, thanks for your awesome repository. Everything is set up awesomely, Tensorboard and all, and I'd love to get this to work.

Unfortunately, I can confirm @jinfagang observation that the resulting models are not good at all. I spent like 20 USD on an AWS EC2 GPU instance until I realized this. Maybe until you fix this issue, mention this in the readme so other people don't waste their time as well.

Thanks!

[vanhuyz]

I've removed the wrong part of PatchGAN and now the results should be better. But it takes time to confirm.

[lucasjinreal]

@ Would you add some test image result for current made? Have some sample result could be better I think.

[vanhuyz]

Results on training set after 37 epochs.

• apple2orange:

• orange2apple:

[vanhuyz]

I tried on horse2zebra dataset but the results are still blurry. Maybe it is more difficult to translate horse2zebra than orange2apple. Did you train horse2zebra with the original pytorch version?

[lucasjinreal]

I have trained original summer2winter and apple2orange, both get in some level fair result, at least the resolution are more clear.

[vanhuyz]

I fixed another issue and retrained horse2zebra. The results are quite promising. A screenshot after 87 epochs:

[lucasjinreal]

@vanhuyz Hi, I also managed predict using original torch version and reconstruct a production version, here is the repo https://github.com/jinfagang/pytorch_cycle_gan, I have some results, seems original still more clear, what did you see? Maybe we can discuss about the difference detail, do you have wechat ?

[lhao0301]

What does the another issue mean? Could you please point it out specifically?

[vanhuyz]

@TX2012LH just 1 line but critical https://github.com/vanhuyz/CycleGAN-TensorFlow/commit/960d1b085459856213d653d0f4e5a1c42dc1f0cf. G and F of cycle loss were reversed :flushed:

[vanhuyz]

@jinfagang Sorry I don't have wechat. How about via email?

[lucasjinreal]

This is my email [email protected], you can send me some information