Report is available here: ./report.pdf

This repository implements 3 Generative Adversarial Networks (GANs) on the MNIST dataset:

• Simple GAN (./First_GAN.ipynb)
• DC-GAN [1] (./DC_GAN.ipynb)
• WGAN-PG [2,3] (./W_GAN.ipynb)

Checkpoints for pretrained networks are available here, you should put them in the ./checkpoints/ directory.

Interpolation between two noise vectors is demonstrated in ./Interpolation.ipynb.

Example of fixing childish paint skills using StyleGAN2 [4] is demonstrated in ./StyleGAN2.ipynb. Additionally, StyleGAN2-ADA [4] wrapper scripts are provided in the ./tools/.

• Docker
• NVIDIA Container Runtime
• Conda (see ./conda-env.yml)
• NVIDIA GPU, NVIDIA drivers (see NVlabs/stylegan2-ada)

# --- StyleGAN2 ---
$ ./tools/stylegan2-generate.sh -h
$ ./tools/stylegan2-projector.sh -h

# --- MISC ---
# Clean large notebook
$ ./tools/jupyter-clean.sh my_notebook.ipynb

[1] Radford, Alec, Luke Metz, and Soumith Chintala. "Unsupervised representation learning with deep convolutional generative adversarial networks." arXiv preprint arXiv:1511.06434 (2015).

[2] Arjovsky, Martin, Soumith Chintala, and Léon Bottou. "Wasserstein gan." arXiv preprint arXiv:1701.07875 (2017).

[3] Gulrajani, Ishaan, et al. "Improved training of wasserstein gans." Advances in neural information processing systems. 2017.

[4] Karras, Tero, et al. "Training generative adversarial networks with limited data." arXiv preprint arXiv:2006.06676 (2020).

[5] Build Basic Generative Adversarial Networks (GANs) by deeplearning.ai