Forward $\chi^2$ Divergence Based Variational Importance Sampling [ICLR 2024 Spotlight]

Chengrui Li, Yule Wang, Weihan Li, and Anqi Wu

[paper] [arXiv] [slides] [video] [poster] [文章]

1 Tutorial

demo.ipynb is a step-by-step tutorial that run VI or VIS on a toy mixture model.

2 Paper's Results Reproduction

For example, consider the toy mixture model in our paper.

Go to the folder mixture. No installation is needed.

Create three folders in mixture: model, np, and csv.

Run main.py with different idx ranging from 0 to 39.

python main.py [idx]

This idx specifies the method and the random seed via

method_list = ['VI', 'CHIVI', 'VBIS', 'VIS']
seed_list = np.arange(10)

arg_index = np.unravel_index(args.idx, (len(method_list), len(seed_list)))
method, seed = method_list[arg_index[0]], seed_list[arg_index[1]]

The learned model $p(x,z;\theta)$ and $q(z|x;\phi)$ are saved in model. The learning curves are saved in np. The quantitative results are saved in csv.

Open the visualization.ipynb. This jupyter notebook plots Fig. 2 in our paper.