Keras implementation of CycleGAN
Implementation using a tensorflow backend. Testing and evaluation done on street view images.
Results - 256x256 pixel images
Day 2 night
| Input | Translation | Input | Translation |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
Night 2 day
| Input | Translation | Input | Translation |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
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
Code usage
- 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:
- data/Dataset-name/trainA
- data/Dataset-name/trainB
- data/Dataset-name/testA
- data/Dataset-name/testB
-
Set wanted training options, also found in the init function.
-
Train a model by:
python model.py
- Generate synthetic images by following specifications under:
- generate_images/ReadMe.rtf
The following gif shows an example of the training progression in a translation from day to night.
Left: Input image. Middle: Translated images. Right: Reconstructed images.
More results
Day 2 night - gif
Rainy 2 sunny - gif
Sunny 2 rainy - gif
