RevGAN

RevGAN copied to clipboard

RevGAN: Memory-efficient Image-to-Image Translation in 2D and 3D

RevGAN implementation in PyTorch. We extend the Pix2pix and CycleGAN framework by exploring approximately invertible architectures in 2D and 3D. These architectures are approximately invertible by design and thus partially satisfy cycle-consistency before training even begins. Furthermore, since invertible architectures have constant memory complexity in depth, these models can be built arbitrarily deep without requiring additional memory. In the paper we demonstrate superior quantitative output on the Cityscapes and Maps datasets at near constant memory budget.

Project | Paper

@inproceedings{vanderOuderaa2019revgan,
title={Reversible GANs for Memory-efficient Image-to-Image Translation},
author={van der Ouderaa, Worrall, Daniel E},
booktitle={CVPR 2019},
}

Getting Started

Installation

• Install [PyTorch](http://pytorch.org and) 0.4+ and other dependencies (e.g., torchvision, visdom and dominate).
  • For pip users, please type the command pip install -r requirements.txt.
• Clone this repo:

git clone https://github.com/tychovdo/RevGAN
cd RevGAN

Training

• Download a CycleGAN dataset (e.g. cityscapes):

bash ./datasets/download_cyclegan_dataset.sh cityscapes

• Train a model:

python train.py --dataroot ./datasets/cityscapes --name example_run --model unpaired_revgan

During training, results will be saved to ./checkpoints/example_run/web/index.html

Testing

• Test a model:

python test.py --dataroot ./datasets/cityscapes --name example_run --model unpaired_revgan

The test results will be saved to ./results/example_run/test_latest/index.html

Prerequisites

• Linux or macOS
• Python 2 or 3
• NVIDIA GPU + CUDA cuDNN

Acknowledgements

We grateful to the Diagnostic Image Analysis Group (DIAG) of the Radboud University Medical Center, and in particular Prof. Dr. Bram van Ginneken for his collaboration on this project. We also thank the Netherlands Organisation for Scientific Research (NWO) for supporting this research and providing computational resources.

This code relies heavily on the image-to-image translation framework from pytorch-CycleGAN-and-pix2pix and extends the framework with paired and unpaired reversible models in 2D and 3D. The reversible blocks are implemented using a modified version of MemCNN.