PyTorch Implementation of CycleGAN

PyTorch implementation of Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks on the Facades dataset.

Prerequisites

• PyTorch
• torchvision

DATASET

In the CycleGAN folder, run:

wget https://people.eecs.berkeley.edu/~tinghuiz/projects/pix2pix/datasets/facades.tar.gz
tar -zxvf facades.tar.gz
rm facades.tar.gz

Go to the scripts folder and run:

python PrepareDataset.py --dataPath ../facades

This script will split paired training image into unpaired training images. At the end of this script, it will ask you whether to delete original paired data in order to save disk space, please be aware that deleted data is unrecoverable.

Training

python CycleGAN.py --cuda

Generate

python generate.py --G_AB checkpoints/G_AB_40000.pth --G_BA checkpoints/G_BA_40000.pth -cuda --dataPath facades/val/

To train or generate on other dataset, change dataPath accordingly.

• Generations:

A -> B -> A

B -> A -> B

A -> B -> A

B -> A -> B

Notes

• DiscoGAN can't generate high quality reconstruction images on the Facades dataset, one thing that CycleGAN resolves this is by using a generator containing 6 residual blocks.
• It's important not to chain the parameters of two discriminators together, otherwise severe mode collapse when batchSize=1 will be observed.

Reference

1. https://github.com/junyanz/CycleGAN
2. Zhu J Y, Park T, Isola P, et al. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks[J]. arXiv preprint arXiv:1703.10593, 2017.