Adaptive Cholesky Gaussian Process (AC⚡ GP)

Link to accepted AISTATS'23 paper.

Setup

In the project folder:

conda create --prefix ./env
conda activate ./env
conda install -y -c conda-forge -c gpytorch --file requirements.txt

Installation as python package (optional)

Run the following from the main directory (where this README file is also located) to install the package in development mode (that is, modifications to the source code is directly visible to file importing it without the need for reinstallation).

pip install -e .

Using OpenBLAS wrappers (optional)

Copy /usr/lib/libopenblas.so into acgp/blas_wrapper/openblas/lib.

Test by running python run_hyper_parameter_tuning.py. If it throws an error like the following:

AttributeError: [...]/blas_wrappers/openblas/lib/libopenblas.so: undefined symbol: dpotrf_

copy instead /usr/lib/liblapack.so to blas_wrapper/openblas/lib/libopenblas.so (yes, to libopenblas.so!).

Example Code

The file example_script.py guides through the main concepts of ACGP.

Running large scale experiments on a slurm cluster

Switch to the experiments folder.

cd experiments

Bound Experiments

python run_ground_truth_experiments.py -m generate_batch_jobs
python run_cglb_experiments.py -m generate_batch_jobs

To recreate the plots run

python make_bound_plotting.py
python make_llh_plotting.py

Hyper-parameter tuning experiments

python run_hyper_parameter_tuning.py -m generate_batch_jobs

After all runs have finished execute

python local_auxilary_computations.py

To recreate the result tables run

python make_results_table.py

For the plots run

python make_optimization_plotting.py