Play with pretrained GAN

Part 2 & Alumni (2018)
1

Here is my jupyter notebook.
Let you to play with pretrained GAN (SAGAN + SNGAN + D2GAN)

GITHUB:

Requirement:

  • old fastai 0.7 (put in the same file)
  • python 3.6
  • pytorch 0.4
  • test on windows 10, cuda 9.0

change to your dir of dataset: put your image in file train

It is ok without any dataset and just play the pretrained GAN
image

image
image.png721×644 64 KB

Train on Guassian but generate on uniform will make the image looks better,
but the uniform images look more similar.

image
image.jpg732×843 256 KB

Continues change on row and colume

image
image.jpg827×1035 361 KB

multi_noise display

show different results in mult std
small std make image looks more similar,
large std make image looks more difference.

image
image.jpg706×1717 606 KB

4 Likes
2

Hello. I have spent several hours today trying to retrofit this code into the newest fastai package and finally losing my mind and giving up. Has anyone here tried to generate faces with a GAN and a pretrained network using fastai?
Thank you…