vdutor / Toy_DGP_experiment

Original code to reproduce Toy DGP experiment

⚠️ This is old, untested and undocumented code. It's only use-case is to reproduce the experiment below using the original code. The original experiment and the code in experiment.py was written by Hugh Salimbeni for the paper Deep Gaussian Processes with Importance-Weighted Variational Inference. The code under src is an early version of GPflux.

Toy Deep Gaussian Process Experiment

Code to reproduce experiment Fig. 1 from Salimbeni et al. (2018) and Fig. 7 from Leibfried et al. (2021).

GP	GP + GP	Latent Layer (LL) + GP	LL + GP + GP

Installation

Disclaimer: legacy code written in Python 3.6 and TensorFlow 1.5.

The experiment uses a (very) early version of GPflux (contained in this repo under source) and GPflow 1.2.0 (commit hash).

Step 1.

Create a virtual environment with Python 3.6, using for example anaconda. After a successful installation of (mini) conda, a new virtual env can be created and activated using

conda create -n py36 python=3.6  && conda activate py36

Step 2.

Install the dependencies

pip install -r requirements.txt

Run the experiments

From the repo's root directory run

python experiment.py --configuration G1 --mode VI

which will create a shallow sparse GP model, train and evaluate it. The results will be stored in the directory results. For a latent-variable shallow sparse GP model, run

python experiment.py --configuration L1_G1 --mode VI

and for deep sparse GP versions of the former (with two layers each) run

python experiment.py --configuration G1_G1 --mode VI

and

python experiment.py --configuration L1_G1_G1 --mode VI

Note that each experiment will create separate folders in the results directory.