Skip to content

Bayesian Coresets Construction with Accelerated Iterative Hard Thresholding (A-IHT).

License

Notifications You must be signed in to change notification settings

jackyzyb/bayesian-coresets-optimization

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

23 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Bayesian Coresets Construction with Accelerated Iterative Hard Thresholding

This repository provides python implementation for constructing Bayesian coreset using accelerated iterative hard thresholding (A-IHT).

Corresponding paper: Bayesian Coresets: Revisiting the Optimization Perspective. Jacky Y. Zhang, Rajiv Khanna, Anastasios Kyrillidis, and Oluwasanmi Koyejo. (AISTATS 2021)

Please refer to the paper for details regarding the two algorithms, i.e., A-IHT I (Algorithm 1) and A-IHT II (Algorithm 2), and the experiment settings.

Contents

  1. IHT Toolbox
  2. Experiments

IHT Toolbox

Three implementations of the A-IHT are provided in IHT_toolbox/accelerated_iht.py, and can be applied directly. Specifically,

  1. A-IHT I implemented with numpy
  2. A-IHT II implemented with numpy
  3. A-IHT II implemented with pytorch
    For large-scale problems, use the pytorch version on GPU for acceleration.

Experiments

The three experiments in our paper are in experiments/bayesiancoresets/examples/riemann_gaussian/ (Synthetic Gaussian posterior inference), experiments/bayesiancoresets/examples/riemann_linear_regression/ (Bayesian Radial Basis Function Regression), and experiments/bayesiancoresets/examples/riemann_logistic_poisson_regression/ (Bayesian logistic and Poisson regression), respectively. To run the experiments, simply run the run.sh under each directories.

Installation and Dependencies

To install the experiment with pip, download the repository and run pip3 install . --user in the experiments/ folder. The experiments depend on NumPy, SciPy, and SciKit Learn. The examples also depend on Bokeh and cairosvg for plotting.

Reference

The experiments are built on the framework of Trevor Campbell, etc.'s repository (commit a7d97b7 on Nov 27, 2019).

About

Bayesian Coresets Construction with Accelerated Iterative Hard Thresholding (A-IHT).

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published