A modern PyTorch implementation of SRGAN
It is deeply based on Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network paper published by the Twitter team (https://arxiv.org/abs/1609.04802) but modernizing it with some new promising discoveries such as DenseNets (https://arxiv.org/abs/1608.06993)
I also try to follow best practices described here: https://github.com/soumith/ganhacks
Still a work in progress for now, but hopefully it will serve as a guide for people implementing somewhat complex GANs with PyTorch.
Contributions are welcome!
Work in progress. Good results are starting to be visible with the CIFAR-100 dataset.
To start a training session:
./train --cuda
- PyTorch
- torchvision