https://github.com/hiyouga/sagan-pytorch

A PyTorch implementation for Self-Attention Generative Adversarial Networks
https://github.com/hiyouga/sagan-pytorch

Last synced: 4 months ago
JSON representation

A PyTorch implementation for Self-Attention Generative Adversarial Networks

• Host: GitHub
• URL: https://github.com/hiyouga/sagan-pytorch
• Owner: hiyouga
• License: mit
• Created: 2021-07-14T09:30:16.000Z (about 4 years ago)
• Default Branch: main
• Last Pushed: 2021-08-26T03:38:41.000Z (about 4 years ago)
• Last Synced: 2024-12-17T23:12:29.056Z (10 months ago)
• Language: Python
• Size: 1.66 MB
• Stars: 7
• Watchers: 3
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# SAGAN-PyTorch

A [PyTorch](https://pytorch.org/) implementation for Goodfellow et al.'s ICML 2019 work "Self-Attention Generative Adversarial Networks". [[arXiv]](https://arxiv.org/abs/1805.08318) [[PMLR]](http://proceedings.mlr.press/v97/zhang19d/zhang19d.pdf)

![GitHub](https://img.shields.io/github/license/hiyouga/sagan-pytorch)

## Requirements

Please install requirements by `pip install -r requirements.txt`

- Python 3.7

- numpy 1.17.2

- torch 1.7.1

- torchvision 0.8.2

- Pillow 6.1.0

The following libraries are optional:

- tensorboard

- tqdm

## Usage

### Clone

```bash

git clone https://github.com/hiyouga/SAGAN-PyTorch.git

cd SAGAN-PyTorch

```

### Train

```bash

python main.py --batch_size 64 --im_size 32 --dataset cifar10 --adv_loss wgan-gp

```

## Results

Note that the wgan-gp loss is adopted by default, since we found that the wgan-gp loss performs much better than hinge loss in our experiments.

We use the CIFAR dataset as the unsupervised training set to generate images.

![Real images](assets/real.png)

### Generated images with wgan-gp loss

![Fake images with wgan-gp loss](assets/wgan.gif)

### Generated images with hinge loss

![Fake images with hinge loss](assets/hinge.gif)

Click to show the training details (visualized via TensorBoard)

### Attention weight and Inception score with wgan-gp loss

![Attention weight and Inception score with wgan-gp loss](assets/wgan_score.png)

### Attention weight and Inception score with hinge loss

![Attention weight and Inception score with hinge loss](assets/hinge_score.png)

### Loss curves with wgan-gp loss

![Loss curves with wgan-gp loss](assets/wgan_loss.png)

### Loss curves with hinge loss

![Loss curves with hinge loss](assets/hinge_loss.png)

## References

For SAGAN architecture:

1. Zhang et al. Self-Attention Generative Adversarial Networks. ICML. 2019.

1. https://github.com/heykeetae/Self-Attention-GAN

1. https://github.com/christiancosgrove/pytorch-spectral-normalization-gan

For inception score:

1. Salimans et al. Improved Techniques for Training GANs. NeurIPS. 2016.

1. Shane Barratt and Rishi Sharma. A Note on the Inception Score. ICML Workshop on Theoretical Foundations and Applications of Deep Generative Models. 2018.

1. https://github.com/sbarratt/inception-score-pytorch

1. https://github.com/w86763777/pytorch-gan-metrics

## License

MIT