Running the code in this repository requires the sgmcmc R package, which can be downloaded from the CRAN repository.
Each of the models from the paper can be found in one of the following folders:
- Diagnostic tests (Section 4) - Code to compute the kernel Stein discrepancy.
- Logistic regression (Section 6.1) - Code to compare the sgmcmc algorithms on a logistic regression model using simulated data.
- Bayesian neural networks (Section 6.2) - Code to run the sgmcmc algorithms on a Bayesian neural network model using the popular MNIST dataset.
- Bayesian probabilistic matrix factorisation (Section 6.3) - Code to run the sgmcmc algorithms on the Bayesian probabilistic matrix factorisation model using the MovieLens dataset.
Note that within each folder for the SGMCMC simulations there is a file run_algorithms.R which is the main file to run each of the SGMCMC algorithms. The other files, e.g. bpmf_setup.R and bpmf_model.R provide utility functions and the Tensorflow model for the posterior targets, respectively.