CycleGAN reimplimented in Keras
What has been made?
- Instance Norm layer for keras
- Cycle loss function for keras
- Test cyclegan training script
What needs to be done?
- Bigger models for non-mnist images need to be created
- All of the generator training models should probably be convered into one model for a real problem.
- Fix the upscaling on generator models
For the test training script, I decided to have cyclegan learn a mapping from cat pictures to mnist pictures (and back).
This took like 10 hours on a Nvidia GTX 1060 using the training parameters in mnist_test.py.
Here are some results (from various stages of training):