News

• We re-implement CycleGAN by Tensorflow 2! The old versions are here: v1, v0.

Tensorflow 2 implementation of CycleGAN.

Paper: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

Author: Jun-Yan Zhu et al.

row 1: summer -> winter -> reconstructed summer, row 2: winter -> summer -> reconstructed winter

row 1: horse -> zebra -> reconstructed horse, row 2: zebra -> horse -> reconstructed zebra

row 1: apple -> orange -> reconstructed apple, row 2: orange -> apple -> reconstructed orange

• Environment

  • Python 3.6

  • TensorFlow 2.2, TensorFlow Addons 0.10.0

  • OpenCV, scikit-image, tqdm, oyaml

  • we recommend Anaconda or Miniconda, then you can create the TensorFlow 2.2 environment with commands below

    conda create -n tensorflow-2.2 python=3.6
    
    source activate tensorflow-2.2
    
    conda install scikit-image tqdm tensorflow-gpu=2.2
    
    conda install -c conda-forge oyaml
    
    pip install tensorflow-addons==0.10.0

  • NOTICE: if you create a new conda environment, remember to activate it before any other command

    source activate tensorflow-2.2

• Dataset

  • download the summer2winter dataset

    sh ./download_dataset.sh summer2winter_yosemite

  • download the horse2zebra dataset

    sh ./download_dataset.sh horse2zebra

  • see download_dataset.sh for more datasets

• Example of training

  CUDA_VISIBLE_DEVICES=0 python train.py --dataset summer2winter_yosemite

  • tensorboard for loss visualization

    tensorboard --logdir ./output/summer2winter_yosemite/summaries --port 6006

• Example of testing

  CUDA_VISIBLE_DEVICES=0 python test.py --experiment_dir ./output/summer2winter_yosemite