Skip to content

Latest commit

 

History

History
20 lines (13 loc) · 1.66 KB

README.md

File metadata and controls

20 lines (13 loc) · 1.66 KB
Raw

cntk-cyclegan

CNTK based implementation of Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks [paper]

How-to run the code

  1. Yosemity dataset Yosemity dataset is expected to be unzipped into ./data folder
  2. dataUtils.py generates map files for input
  3. If you ran on GPU and see out-of-memory exception => lower batch size

Results

I have ran trainCycleGan.py on Yosemity dataset and batch size 4. This dataset is not super clean, the set of summer imagages has several winter images and vice versa. I did quick clean up of those before training. Also I noticed that in current implementation I have G(X) that transfers summer Yosemity to winter works better than F(X) (winter to summer). Also Generator tends to change daytime to evening\night time.