Self-Supervised GAN

Implementation of the paper:

Bingchen Liu, Yizhe Zhu, Kunpeng Song and Ahmed Elgammal. Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis.

See here for the official Pytorch implementation.

Examples

Dependencies

• Python 3.8
• Tensorfow 2.7
• Tensorflow Addons 0.15

Usage

Train

Use --train_dir=<file_pattern> to provide the dataset file_pattern. The file pattern "./dataset_path/*.jpg" builds a jpg image dataset.

python train.py --dataset_path=./dataset_path/*.jpg

Hparams setting

Set batch_size and learning_rate on the hparams.py file.

Tensorboard

Run tensorboard --logdir ./.

References

Code:

• This model depends on other files that may be licensed under different open source licenses.
• Self-Supervised GAN uses Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu and Song Han. Differentiable Augmentation. Under BSD 2-Clause "Simplified" License.

Implementation notes:

• Self-supervised discriminator with a single reconstruction decoder and perceptual loss.
• Hinge loss GAN and WGAN gradient penalty.
• Orthogonal initialization.
• Skip-layer excitation generator.
• Adam with β1 = 0.5 and β2 = 0.99.
• Batch size = 8.

Licence

MIT