CS6216 Project: Pyro Implementation of Nonparametric Hamiltonian Monte Carlo Method

This repo serves as a record of a course project. It contains a forked Pyro repo as the submodule, and some scripts to facilitate development.

Setup

Remember to update the submodule:

git submodule update --init --recursive

It is recommended to do the setup in a virtual environment (assuming you are using bash):

python -m venv ./my-venv && source ./my-venv/activate

Then install the forked pyro and the utility package in develop mode:

pip install -e pyro[extras]

With -e flag, you don't have to re-install pyro after making changes in the forked repo.

Install dependencies:

pip install -r requirements.txt

Usage

The python module walk.py contains scripts for different purposes. To get help strings for different commands, run

./walk.py --help  # if walk.py is executable
python walk.py --help

All commands have an option of --output-path, whose default is ./output/. It will have the following structure

<output-path>
|- random_walk
| |- samples
| | |- nuts/*.pickles
| | |- hmc/*.pickles
| | |- npdhmc/*.pickles
| | |- groudtruth/*.pickles
| |- importance_draws/*.pickles
| |- result.png
|- geometric/...
|- <other_probabilistic_model>/...

The program will take care of generating and storing data according to this structure. You don't have to worry about it unless you are downloading pre-generated data.

Most commands have an option of --model-name, whose default is random_walk. It will determine which probabilistic program is used.

Groundtruth from importance sampling and systematic resampling

Run enough repetitions of ./walk.py imp to collect draws from importance sampling, then use the draws to do systematic resampling ./walk.py gt.

MCMC methods

Do ./walk.py run nuts -r 1. See ./walk.py run --help for help.

Generate the plot