SGLD

An implementation of SGLD (and variants) with two key purposes in mind:

1. Easy to cherry-pick into your own research code for baseline evaluation.
2. Clean enough for dirty prototyping with fancy/ad-hoc extensions.

Setup

Use the environment.yml to add common dependencies to your working Python environment.

conda env create -n <env_name>

Add the slgd package to PYTHONPATH for imports.

export PYTHONPATH="$(pwd)/src:${PYTHONPATH}"

Usage

Toy SGLD Visualization

See viz_sgld.ipynb.

Full Example

Full examples which are easy to modify.

• SGLD: A full pipeline of using SGLD is provided in train_sgld.py

• cSGLD: A full pipeline of using SGLD + cyclical step size schedule (cSGLD) is provided in train_csgld.py

License

MIT