Skip to content

mgorinova/autoreparam

Repository files navigation

Automatic Reparameterisation of Probabilistic Programs

This repository contains code associated with the paper:

M. I. Gorinova, D. Moore, and M. D. Hoffman. Automatic Reparameterisation of Probabilistic Programs. 2019.

Usage

The script main.py is the main entry point. For example, to evaluate the German credit model with four leapfrog steps per sample, you might run:

# Run variational inference to get step sizes and initialization.
python main.py --model=german_credit_lognormalcentered --inference=VI --method=CP --num_optimization_steps=3000 --results_dir=./results/
# Run HMC to sample from the posterior
python main.py --model=german_credit_lognormalcentered --inference=HMC --method=CP --num_leapfrog_steps=4 --num_samples=50000 --num_burnin_steps=10000 --num_adaptation_steps=6000 --results_dir=./results/

Available options are:

  • method: CP, NCP, cVIP, dVIP, i. Note that i is only available when inference is set to HMC.
  • inference: VI or HMC. VI needs to be run first for every model in order for a log file to be created, which contains information such as initial step size to be adapted when running HMC.
  • model: radon_stddvs, radon, german_credit_lognormalcentered, german_credit_gammascale, 8schools, electric, election and time_series
  • dataset (used only for radon models): MA, IN, PA, MO, ND, MA, or AZ

To generate human-readable analysis, run

python analyze.py --elbos --all
python analyze.py --ess --all
python analyze.py --reparams --all
python analyze.py --elbos --model=8schools

The number of leapfrog steps will be automatically tuned if (1) no num_leapfrog_steps argument is supplied and (2) no entry num_leapfrog_steps exists in the respective .json file.

When the number of leapfrog steps is tuned, the best number of leapfrog steps is recorded in a .json file, so that it can be reused accordingly.

This code has been tested with TensorFlow 1.14 and TensorFlow Probability 0.7.0.

Releases

No releases published

Packages

No packages published

Languages