This repository contains code to replicate the experiments of the paper On Consistent Bayesian Inference from Synthetic Data.
conda env create
conda activate bayes-downstream-envMost of the code in lib/ is from our
NAPSU-MQ implementation
(Räisä et al. 2023).
A more user-friendly version of the core algorithm is available in the Twinify library.
The DP-GLM baseline uses code from Kulkarni et al. The code is not public, so we could not include it in the repo.
The Gaussian example is in the notebook bayesian-inference-gaussian-test.ipynb.
For the toy data and UCI Adult experiments, we use Snakemake. The experiments can be run with
snakemake -j 16which uses 16 cores in parallel. After the command finishes, the plotting
notebooks workflow/scripts/toy-data/report.py.ipynb can
and workflow/scripts/adult-reduced/report.py.ipynb be run to
plot the results.
Note: running this will likely take several days on a single computer.
You can edit the snakefiles in workflow/rules to run a subset of the
experiments.
Figures are saved in the figures/ directory, and separated into subdirectories
of each example.