Capturing label characteristics in VAEs

Offical repository for Capturing Label Chacteristics in VAEs ICLR 2021. This codebase uses pyro, which some users may not find useful, conseqently we also released a purely pytorch version here. We kindly ask that you cite our work if you plan to use this codebase:

@inproceedings{
    joy2021capturing,
    title={Capturing Label Characteristics in {\{}VAE{\}}s},
    author={Tom Joy and Sebastian Schmon and Philip Torr and Siddharth N and Tom Rainforth},
    booktitle={International Conference on Learning Representations},
    year={2021},
    url={https://openreview.net/forum?id=wQRlSUZ5V7B}
}

Dependencies

• Python 3.6
• Pytorch 1.8
• Pyro-ppl 1.5

Usage

Ensure that CelebA is in the directory data/datasets/celeba, such that the path data/datasets/celeba/celeba/img_align_celeba/* is accessable.

To train, run:

python ss_vay.py -sup <sup-frac> --cuda>

where <sup-frac> is the fraction of supervised data (e.g. 0.004, 0.06, 0.2, 1.0).

Results

Classification accuracies and latent traversals will be stored in data/vae_results/f_<sup-frac>.