Keras implementation of CycleGAN
Implementation using a tensorflow backend. Testing and evaluation done on street view images.
Results - 256x256 pixel images
Day 2 night
Night 2 day
Model additions as training options
- Identity learning (on different modulus of training iterations)
- PatchGAN in discriminators
- Multi-scale discriminators
- Resize convolution in generators
- Supervised learning with training weight
- Data generator (if using a large dataset)
- Weight on discriminator training labels on real images
- Prepare your dataset under the directory 'data' and set dataset name to parameter 'image_folder' in CycleGAN init function.
- Directory structure on new dataset needed for training and testing:
Set wanted training options, also found in the init function.
Train a model by:
- Generate synthetic images by following specifications under:
The following gif shows an example of the training progression in a translation from day to night.