This is a clone of Hugh Salimbeni's Bayesian benchmarks repo, found here
In order to run experiments on the UCI-large datasets from the paper, please run mkdir data
from experiments/uci_exps
folder.
Then, download uci.tar.gz
from Google Drive and put
it in the data
folder.
Run cd data; tar -xzvf uci.tar.gz
.
The scripts for reproducing results on UCI problems from the paper can be found at
experiments/uci_exps/bayesian_benchmarks/tasks/run_ucilarge.sh
(UCI-large) and
experiments/uci_exps/bayesian_benchmarks/tasks/run_ucismall.sh
(UCI-small).
You can run these scripts as follows
cd experiments/uci_exps/bayesian_benchmarks/tasks/
chmod u+x run_ucismall.sh
./run_ucismall.sh
In the paper we report results averaged over splits 1,...,20, and use splits 21,...,23 for validation.
The results of scripts run_ucismall.sh
and run_ucilarge.sh
will be saved in databases
results_small.db
and results_large.db
respectively.
You can then view the results using the script view_results.py
at experiments/uci_exps/bayesian_benchmarks/tasks/
.
For example, to view results on UCI-small run
python3 view_results.py --database results_small.db