CycleGAN with perception loss

What is this repository for?

Implementation of CycleGan model in Keras (original implementation link).

Demonstration: De-raining images

The example below presents 18 rainy images of shape (128x128x3) where cycleGAN with perception loss has been used to de-rain.

How do I get set up ?

Install Anaconda 3 Import the conda environment named deepenv using :

conda env create -f deepenv.yml

Activate that environment using :

source activate deepenv

Now all the dependencies must be installed without problems (Keras 2, tensorflow 1 ...)

How do I train CycleGAN ?

you may have information on how to run train.py by:

python predict.py --help

you can train your own model by running (N.B.: example):

python train.py --path_trainA ./data/trainA --path_trainB ./data/trainB --pic_dir ./intermediate_res --lmbd 10

How do I train CycleGAN with perception loss ?

you can train CycleGan with Perception loss by running:

python train.py --path_trainA ./data/trainA --path_trainB ./data/trainB --pic_dir ./intermediate_res --lmbd 10 --lmbd_feat 1

How do I deploy CycleGAN on new images after training?

you can deploy the model on a given collection, in order to transform A to B or B to A (Possible only after training).

python test.py --path_images ./data/trainA --pic_dir ./results --model_path ./../a2b.h5

Contents

└── cyclegan
    ├── data                         # data folder contaning both A and B images
         ├── trainA                  # images belonging to class A
         └── trainB                  # images belonging to class B
    ├── pics                         # intermediate results folders (for training phase)
    ├── deepenv.yml                  # Environment (keras 2, tensorflow 1.1, etc ...)
    ├── discriminator.py             # discriminator
    ├── generator.py                 # generator (Resblock 6 & unet_128)
    ├── resnet_builder.py            # utils for perception loss (Resnet50)
    ├── resnet50.py                  # cnn for perception loss (Resnet50)
    ├── layers.py                    # ReflectPadding2D & InstanceNormalization2D
    ├── models.py                    # cycleGAN: fit & predict
    ├── README.md                    # Readme
    ├── test.py                      # deploy model
    ├── train.py                     # train model
    ├── utils.py                     # utils

Acknowledgement

• PiscesDream https://github.com/PiscesDream/CycleGAN-keras
• shadySource https://github.com/shadySource/cyclegan_keras
• tjwei https://github.com/tjwei/GANotebooks
• Elias Vansteenkiste https://github.com/EliasVansteenkiste/CycleGANwithPerceptionLoss