Self-Supervised GAN

Implementation of the FastGAN model in the paper:

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 2.8
• Tensorflow Addons 0.16

Usage

Train

Use --file_pattern=<file_pattern> to specify the dataset path and file pattern.

python train.py --file_pattern=./dataset_path/*.png

Generate

Use --main_dir=<main_dir> to specify the model directory name.

python generate.py --main_dir=<main_dir>

Hparams setting

Adjust hyperparameters in 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 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.
• Skip-layer excitation generator.
• Orthogonal initialization.
• Adam with β1 = 0.5 and β2 = 0.99.
• Batch size = 8.

Licence

MIT