Fully Bayesian Autoencoders with Latent Sparse Gaussian Processes

Code for the paper "Fully Bayesian Autoencoders with Latent Sparse Gaussian Processes".

Dependencies

This repository was tested with Python 3.7, CUDA 10.0 and cuDNN 7.6.0 on Ubuntu 20.4.

Run the following command to install necessary python packages for the repository:

pip install -r requirements.txt

Examples

Here are examples of SGP-BAE and DSGP-BAE models using shallow and deep GP priors, respectively:

SGP-BAE

python3 sgpbae_experiment.py \
    --out_dir="exp/mnist/sgpbae" \
    --lr=0.01 \
    --mdecay=0.01 \
    --n_samples=200

DSGP-BAE

python3 dsgpbae_experiment.py \
    --out_dir="exp/dsgpbae/jura" \
    --lr=0.002 \
    --mdecay=0.05 \
    --n_samples=50

Acknowledgement

Our code and experiments are based on the following repositories:

• https://github.com/srossi93/Sparse-Gaussian-Processes-Revisited
• https://github.com/ratschlab/SVGP-VAE
• https://github.com/MattAshman/sgpvae

Contact

Feel free to contact me via email ([email protected]) if you have any issues or questions.

Citation

When using this repository in your work, please consider citing our paper

@inproceedings{Tran2023,
  author    = {Tran, Ba-Hien and Shahbaba, Babak and Mandt, Stephan and Filippone, Maurizio},
  title     = {{Fully Bayesian Autoencoders with Latent Sparse Gaussian Processes}},
  booktitle = {Proceedings of the 40th International Conference on Machine Learning, ICML 2023},
  address   = {Honolulu, Hawaii, USA},
  publisher = {PMLR},
  series    = {Proceedings of Machine Learning Research},
  year      = {2023}
}