kirthevasank / dragonfly-experiments

Experiments presented in the paper "Tuning Hyperparameters without Grad Students: Scalable and Robust Bayesian Optimisation with Dragonfly"

Dragonfly Experiments

To get started, first follow the instructions in the Dragonfly repository to insall Dragonfly.

Then, clone this repository.

$ git clone https://github.com/dragonfly/dragonfly-experiments.git

 

This repository provides three Python scripts to run experiments.

1. euclidean/run_euclidean_experiments.py: Executes the experiments on Euclidean domains. This include the synthetic experiments on the maximum likelihood problem on luminous red galaxies.

2. non_euclidean/run_non_euclidean_synthetic_experiments.py: Executes the synthetic experiments on non-Euclidean domains.

3. non_euclidean/run_non_euclidean_realtime_experiments.py: Executes the model selection and astrophysical maximum likelihood experiments on non-Euclidean domains.

The experiments use the examples in the dragonfly/examples directory. You need to specify the path to this directory via the DRAGONFLY_EXPERIMENTS_DIR variable at the beginning of each script.

To run these experiments, simply cd into the relevant directory and execute the script. For example, the second script above can be run via the following commands. You may select the specific experiment in the script.

(env)$ cd dragonfly-experiments/euclidean
(env)$ python run_non_euclidean_synthetic_experiments.py

Once the experiment is done, you may use plotting.py to plot the results.

(env)$ python plotting.py --file non_eucildean/syn_results/<file_name>.mat