thu-ml/fpovi

Code for "Function Space Particle Optimization for Bayesian Neural Networks"

f-POVI

Code for Function Space Particle Optimization for Bayesian Neural Networks, ICLR 2019.

Please reach out at wzy196@gmail.com for questions.

For an imperative implementation, check out the tf2 branch, or a JAX version by @yardenas.

Dependencies

• requirements.txt
• thu-ml/zhusuan (@1011dd9)
• meta-inf/experiments (@60c0a77)

Usage

See scripts in exps/.

You are recommended to look at ffn/ if you want a minimal implementation to build upon.

Acknowledgements

This repository contains code adapted from other sources:

• The dcb directory is a fork of the deep contextual bandit library by Riquelme et al.
• The convnet implementation is adapted from Tensorpack;
• The adversarial attack code (for feed-forward network) is taken from Yingzhen Li's code for alpha-dropout;
• The UCI data processing script is taken from the doubly stochasitc DGP code, from Salimbeni et al.

Citation

@inproceedings{
    wang2018function,
    title={Function Space Particle Optimization for {B}ayesian Neural Networks},
    author={Ziyu Wang and Tongzheng Ren and Jun Zhu and Bo Zhang},
    booktitle={International Conference on Learning Representations},
    year={2019},
    url={https://openreview.net/forum?id=BkgtDsCcKQ},
}