This repository contains an implementation for training a variational autoencoder (Kingma et al., 2014), that makes (almost exclusive) use of pytorch.

Training is available for data from MNIST, CIFAR10, and both datasets may be conditioned on an individual digit or class (using --training_digits). To initialize training, simply go ahead and python3 train.py.

For scoring anomalies on the respective test set, evoke python3 score_elbo.py and make sure to point toward a trained instance with --ckpt_path.

Other available commands are listed by calling python3 train.py -h.

---

Kingma, D. P. & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114.