Hanjun-Dai / cvb

Code for Coupled Variational Bayes via Optimization Embedding

Code for Coupled Variational Bayes via Optimization Embedding (http://papers.nips.cc/paper/8177-coupled-variational-bayes-via-optimization-embedding.pdf)

1. Setup

Get the source code, and do pip install:

git clone --recursive https://github.com/Hanjun-Dai/cvb
cd cvb
pip install -e .

Download the data via the dropbox link:

https://www.dropbox.com/sh/ho3tntcr0a7jsp4/AACpmE_3ErMw2JV3y9Kr8bTXa?dl=0

And create a symbolic link to this dropbox folder. Finally the folder structure should look like this:

cvb (project root)
|__  README.md
|__  cvb
|__  dropbox  -> symbolic link
|__  |__ data
|    |......
|......

2. Run synthetic data

Directly run the script:

cd cvb/toy_img
./run_fenchel.sh

You will get a sequence of posterior samples.

3. Run MNIST and CelebA

There are several variants of the model.

For MNIST, we implemented cvb with gaussian, parametric optimization and also gaussian + fenchel duality;

For CelebA, we implemented cvb with parametric optimization and fenchel duality

To run the training:

cd cvb/celeb
./run_celeb_parametric_opt.sh

To visualize the image samples:

cd cvb/celeb
./vis_celeb_parametric_opt.sh

Reference

@inproceedings{dai2018coupled,
  title={Coupled Variational Bayes via Optimization Embedding},
  author={Dai, Bo and Dai, Hanjun and He, Niao and Liu, Weiyang and Liu, Zhen and Chen, Jianshu and Xiao, Lin and Song, Le},
  booktitle={Advances in Neural Information Processing Systems},
  pages={9712--9722},
  year={2018}
}