Automatic Reparameterisation in Probabilistic Programming

This directory contains code for the paper "Automatic Reparameterisation in Probabilistic Programming", submitted to AABI 2018. This includes interceptors to transform Edward2 programs into non-centered or partially non-centered parameterizations, code for interleaved HMC (iHMC) and Variationally Inferred Parameterisation (VIP), and infrastructure for running experiments, including the models and data used in the paper.

Usage

The script run_experiments.py is the main entry point. For example, to evaluate the German credit model with four leapfrog steps per sample, you might run (from the toplevel google_research directory):

python -m edward2_autoreparam.run_experiments --method=baseline --model=german_credit_lognormalcentered --num_leapfrog_steps=4 --num_mc_samples=16 --num_optimization_steps=2000 --num_samples=50000 --burnin=8000 --num_adaptation_steps=6000 --results_dir=/tmp/results
python -m edward2_autoreparam.run_experiments --method=vip --model=german_credit_lognormalcentered --num_leapfrog_steps=4 --num_mc_samples=16 --num_optimization_steps=2000 --num_samples=50000 --burnin=8000 --num_adaptation_steps=6000 --results_dir=/tmp/results

Available options are:

• method:
  • vip
  • vip_iaf (runs VIP with an inverse autoregressive flow rather than mean-field normal posterior)
  • baseline (runs CP-HMC, NCP-HMC, and iHMC)
• model: radon_stddvs, radon, german_credit_lognormalcentered, german_credit_gammascale, and 8schools
• dataset (used only for radon models): MA, IN, PA, MO, ND, MA, or AZ

Each run will save results as Python pickle files to the specified directory /tmp/results. To generate human-readable analysis, run

python -m edward2_autoreparam.analyze_results --results_dir=/tmp/results --model_and_dataset=german_credit_lognormalcentered_na
cat /tmp/results/analysis/german_credit_lognormalcentered_na_analysis.txt

Authors

• Maria Gorinova ([email protected]): primary contact.
• Dave Moore ([email protected])