Train different types of Generative Adversarial Networks using TensorFlow-GAN. Provided types:
- Vanilla GAN without label switching
- Wasserstein GAN with gradient penalty (WGAN-GP)
- Spectral normalization GAN (SNGAN)
- WGAN-GP with spectral normalization
This repository cover the basic files of the Advanced GAN Tutorial given at the IML Workshop 2019 at CERN.
The full tutorial (including the gan_tutorial) can be found at IML 2019 tutorial.
Which you can directly load on Binder:
The repository provides GAN models for the following datasets:
- CIFAR10 / CIFAR100
- MNIST
Furthermore, you can use 2 physics datasets:
- CMS prototype Calorimeter dataset
- Simulated using Geant4. For further details see Comput Softw Big Sci (2018) 2: 4.. In contrast to the publication, this datsets contains only data of 3 layers and a single energy bin (for simplicity reasons).
- Footprints of cosmic ray induced air
- The dataset contains footprints of cosmic ray induced air showers with energies between 1 - 100 EeV. The details of the simulation can be found in Astropart. Phys. 97 (2017) 46.
For starting the training just run one of the GAN scripts in the main folder. For performance reason the usage of a GPU is highly recommended, otherwise the training lasts >10h, before producing reasonable results. Runs produce TensorBoard outputs -> for monitoring training and plots after each "epoch".
Training duration~14 h NVIDIA 1080 GTX
Training duration~6 h NVIDIA 1080 GTX