Learning image-to-image translation using paired and unpaired trainingsamples

Project | Arxiv | ACCV-2018

---

This is the part of implementation for the "Learning image-to-image translation using paired and unpaired training samples" (https://arxiv.org/pdf/1805.03189.pdf). This paper is accepted in ACCV 2018.

Prerequisites

1. Python 3.5.4
2. Pytorch 0.3.1
3. Visdom and dominate

Training

1. Downlaod cityscapes datasets as in pix2pix and cyclegan as sugggested in here
2. Create a folder name datasets with the subfolder structures as given in this repo.
3. Keep the paired data in train-subfolder and unpaired data in trainA and trainB subfolders.
4. Then run: python train.py --dataroot ./datasets --model cycle_gan --dataset_mode unaligned --which_model_netG resnet_9blocks --which_direction AtoB --super_epoch 50 --super_epoch_start 0 --super_mode aligned --super_start 1 --name mygan_70 --no_dropout

Testing

1. Downlaod cityscapes test data as in cyclegan as sugggested in here
2. Keep the test data in testA and testB subfolders within datasets folder.
3. Then run: python test.py --dataroot ./datasets --model cycle_gan --dataset_mode unaligned --which_model_netG resnet_9blocks --which_direction AtoB --name mygan_70 --how_many 100

Training Tips:

1. With less paired data, increase the --super_epoch value for better results.
2. With No paired data, set --super_start 0.
3. For no unpaired data, set --super_epoch and --niter to same value. We have not included the VGG loss in the training script (Commented part). We will update this soon. For any help, please contact us at: soumya.tripathy@tuni.fi

If you are using this implementation for your research work then please cite us as:

#Citation 

@article{tripathy+kannala+rahtu,
  title={Learning image-to-image translation using paired and unpaired training samples},
  author={Tripathy, Soumya and Kannala, Juho and Rahtu, Esa},
  journal={arXiv preprint arXiv:1805.03189},
  year={2018}
}

Related Work

1. Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio "Generative Adversarial Networks", in NIPS 2014. 
2. Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. "Image-to-Image Translation with Conditional Adversarial Networks", in CVPR 2017.
3. J. Y. Zhu, T. Park, P. Isola, and A. A. Efros. "Unpaired image-to-image translation using cycle-consistent adversarial networks",

NOTE: Code borrows heavily from pix2pix