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

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.

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.

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.

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