A Fully Bayesian Gradient-Free Supervised Dimension Reduction Method using Gaussian Processes

Github repository accompanying the paper entitled "A Fully Bayesian Gradient-Free Supervised Dimension Reduction Method using Gaussian Processes".

Setup

The setup instructions assume you are using anaconda/miniconda.

• clone this repository and cd to the project root
• create a conda environment using the provided environment.yml file (conda env create --file=environment.yml)
• activate this environment (its default name is bayesian_dimension_reduction)
• install this project as a library (pip install -e .)
• finally, clone and install this fork of pymanopt

Run the code

The script run_one_case.py is the main point of entry. It features a set of parameters that can be modified to run a particular case. Running this script produces an hdf5 file in a subfolder of the results directory that contains all relevant training and validation artifacts. hdf5 files may be programatically opened, e.g. using h5py or browsed using a utility such as HDFView.