A CycleGAN model that converts the style of English letters to look like Mandarin or Katakana characters
A lot of code was taken from aitorzip/PyTorch-CycleGAN. Although I made it a lot more simpler (with a potential compromise in quality of results).
More results can be seen here.
For both these datasets, I got the best results when discrinimator weights were set to 0.1
(currently set to 0.3
in the code) and batch size was increased to 16. I couldn't go higher with the batch size due to memory issues but it's worth trying.
- Comment out this line to train on Mandarin characters.
- I also tried training on two image sets: real images and stickers. The files named
*im2im.py
contain the code for it. - You can run
python main.py
for training on character sets orpython main_im2im.py --/path/to/images1 --/path/to/images2
for training on image sets.