This repository hosts code for Variational Auto-Regressive Gaussian Processes for Continual Learning by Sanyam Kapoor, Theofanis Karaletsos, Thang D. Bui.

From the root of the directory

conda env create

This creates the environment from environment.yml.

In addition, ensure availability of the this directory in PYTHONPATH for module imports. As an example in Bash,

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

All experiment scripts utilize Fire and the arguments can directly be converted to CLI arguments. Appropriate functions are mentioned to look for CLI arguments to change. The default arguments are enough to reproduce results in the paper.

CLI Arguments: See toy method.

python experiments/vargp.py toy

CLI Arguments: See split_mnist method.

python experiments/vargp.py s-mnist

CLI Arguments: See permuted_mnist method.

python experiments/vargp.py p-mnist

The checkpoints files can be downloaded here. Extract contents of the zip file into notebooks/results.

All graphs in the paper can now be generated via code in the notebooks.

SK was supported by the Uber AI Residency program.

Apache 2.0