Transfer Learning for GANs?

alvisanovari · November 25, 2018, 11:01pm

I am interested in re-creating the results the NVIDIA team did with Progressive GANs (pg-GANs). Has anyone in the fastai community worked on this?
I am also curious to see if there is a way to use transfer learning for GANs. From what I know it took the NVIDIA team about 20 days to get those photo-realistic images. Can we have a model that starts with that and generates new images based on our custom dataset?

gan · March 3, 2019, 4:19pm

@alvisanovari Hi, not sure if your still working on this but I trained a model on stylegan which was just released last month and used it for portrait art faces. Yes you can do transfer learning with it and it will cut down your training time alot.

alvisanovari · March 3, 2019, 10:02pm

Nice - have a link/github?

gan · March 3, 2019, 10:11pm

yes, github link stylegan-art

alvisanovari · March 4, 2019, 1:57am

Awesome - will check it out.

jankelowitz · October 4, 2019, 1:49pm

Hi,

I am curious if you wound up getting this to work. I have been slowly trying to get transfer learning work for image generation on wgan in a simple example.