BicycleGAN Implemented In Keras
Paper: https://papers.nips.cc/paper/6650-toward-multimodal-image-to-image-translation.pdf
Github: https://github.com/junyanz/BicycleGAN
Slight Modifications:
- Only 1 patch discriminator, with 16 8x8 patches
- Uses CNNs for both E and D
- Uses Pooling and Upsampling instead of strides
- Modified learning rates
(1 x Labels, 1 x Ground Truth, 6 x Generated Image)
Dependencies:
- Numpy
- Pillow
- Matplotlib (Somewhat redundant)
- Tensorflow
- Keras
Be sure to load data into /data/ folder.
data/DomainA and data/DomainB are for image pairs. data/DomainAs and data/DomainBs are for non-paired images.
Run pretrained model: Run main.py, 100 sample sheets will be generated.
Train a new model: Clear parameters in declaration of BicycleGAN near the end of main.py. Comment or delete the line where an old model is loaded near the end of main.py - "model.load(6)". Set train_model to true near the end of main.py. Run main.py.