What it does This repository implements the Bayesian GAN, which replaces the standard GAN’s single point estimate of network weights with a full posterior distribution. It uses stochastic gradient Hamiltonian Monte Carlo to sample over both generator and discriminator parameters, producing an ensemble of models rather than one winner-takes-all network. The code supports unsupervised and semi-supervised training on MNIST, CIFAR-10, SVHN, CelebA, or custom datasets.

The interesting bit The core insight is that mode collapse—the bane of standard GANs—can be framed as an inference problem. By maintaining a multimodal posterior over weights, the Bayesian GAN literally keeps multiple “creative hypotheses” alive at once. The README’s synthetic demo is striking: a standard GAN collapses to a single blob, while the Bayesian version preserves the full multi-blob distribution.

Key highlights

• Semi-supervised learning out of the box: train with as few as 100 labeled MNIST examples or 4,000 CIFAR-10 labels
• Includes scripts for standard benchmarks plus a Digits class template for rolling your own dataset loader
• Supports Wasserstein-GAN variant via --wasserstein flag
• Can run baseline DCGAN or DCGAN ensemble for direct comparison
• Ships with a Jupyter notebook (synth.ipynb) for reproducing the paper’s Jensen-Shannon divergence analysis

Caveats

• Pinned to TensorFlow 1.0.0 and Python 2.7, which are both well past end-of-life
• The README warns that even setting --random_seed does not guarantee reproducibility on GPU
• Custom datasets require hand-editing bgan_util.py and bayesian_gan_hmc.py to wire in your class

Verdict Grab this if you’re doing research on probabilistic deep learning or need a semi-supervised baseline with theoretical grounding. Skip it if you want production-ready code; the dependency stack is a time capsule from 2017.