General

This is the source code for "Boosting Black Box Variational Inference," by Francesco Locatello, Gideon Dresdner, Rajiv Khanna, Isabel Valera, Gunnar Rätsch. https://arxiv.org/abs/1806.02185.

@article{locatello2018boosting,
  title={Boosting Black Box Variational Inference},
  author={Locatello, Francesco and Dresdner, Gideon and Khanna, Rajiv and Valera, Isabel and R{\"a}tsch, Gunnar},
  journal={arXiv preprint arXiv:1806.02185},
  year={2018}
}

Setup

Setup the dependencies using conda:

conda env create -n bbbvi --file conda-env.txt

Activate the environment

source activate bbbvi

Install the package for development

python setup.py develop

Run

Bayesian Logistic Regression

To recreate the Bayesian Linear Regression results in Table 1 of the paper, run

python blr.py \
    --base_dist lpl \
    --exp $EXPERIMENT \
    --outdir $OUTDIR \
    --seed $seed"

Where $EXPERIMENT is either chem or wine.

Bayesian Matrix Factorization

To recreate the Bayesian matrix factorization results in Table 1 of the paper, run

python matrix_factorization.py --D $D --outdir $OUTDIR --seed $seed

Toy data (mixture model)

To recreate Figure 1 of the experiment run

python mixture_model_relbo.py \
    --relbo_reg 1.0 \
    --relbo_anneal linear \
    --exp mixture \
    --fw_variant $variant \
    --outdir $OUTDIR"

Where $variant is fixed, line_search, or fc.

License

This project is licensed under the MIT License - see the LICENSE file for details.

ratschlab / boosting-bbvi