search

jerryli27 / my_pix2pix

This is an adaption from https://github.com/affinelayer/pix2pix-tensorflow. It's for personal use only and this git repo serves as a backup.
3 stars 0 forks source link

WGAN experiences #1

Open kadarakos opened 7 years ago

kadarakos commented 7 years ago

I'm using the pix2pix for a project I'm working on https://github.com/tcsai/tcsai.github.io . I have implemented the WGAN independently of you, but then I've found your implementation, which seemed pretty helpful, so now the code looks exactly like yours.

However, the WGAN doesn't seem to learn to generate as good images as the default. What are your experiences? Did you get it to work?

jerryli27 commented 7 years ago

Hi. Sorry for replying so late. I don't usually check my repo because I wasn't expecting others to use it. Regarding your question, I've found a way for wgan to work. If you compare https://github.com/jerryli27/my_pix2pix/blob/master/pix2pix_w_hint_lab_wgan.py with https://github.com/jerryli27/my_pix2pix/blob/master/pix2pix_w_hint_512.py and only focus on the changes I've made in the loss function and the discriminator, you should see how it's implemented. I've tried it on my own dataset and it indeed increased the model's variety (meaning it outputs more types of correct output) Some minor things that is different from the paper is; I don't find training the discriminator extra rounds help at all. It only slows down the training process. (But that could be I was not implementing that part correctly). Hope this helps!

jerryli27 commented 7 years ago

And let me know if you have any questions. Shoot me a reminder at jiamingli2017 (at) u.northwestern.edu. if I'm not responding on github.