Decentralized Langevin Dynamics for Bayesian Learning

This repository is the official implementation of Decentralized Langevin Dynamics for Bayesian Learning.

Requirements

python 3.x, numpy, matplotlib

pytorch (>=1.0), torchvision

scikit-learn

Scripts for experiments

Toy Data - GMM

Centralized ULA - toy_data_sgld.py

Decentralized ULA - toy_data_dist_sgld.py

Logistic Regression

Centralized ULA - a9a_sgld.py

Decentralized ULA - a9a_dist_sgld.py

Image Classification

SGD - mnist_sgd_svhn_pred_scores.py

Centralized ULA - mnist_sgld_svhn_pred_scores.py

Decentralized ULA - mnist_dist_sgld_svhn_pred_scores.py

All content in this repository is licensed under the MIT license.