gletarte/pbrff

Pseudo-Bayesian Learning with Kernel Fourier Transform as Prior

Pseudo-Bayesian Learning with Kernel Fourier Transform as Prior

This python code has been used to conduct the experiments presented in Section 6 of the following paper:

Gaël Letarte, Emilie Morvant, Pascal Germain. Pseudo-Bayesian Learning with Kernel Fourier Transform as Prior http://proceedings.mlr.press/v89/letarte19a.html

Content

• experiment.py contains the code used to launch experiments and save the results in the results folder.
• pbrff.ipynb is a jupyter notebook to process the results and produce relevant figures.
• pbrff/landmarks_based.py and pbrff/landmarks_selector.py implement algorithms used for Landmarks-Based Learning experiments (section 6.1).
• pbrff/greedy_kernel.py implements algorithms used for Greedy Kernel Learning experiments (section 6.2).
• pbrff/data_loader.py contains the code to load the datasets (located in data folder) used in the experiments.

Launching an experiment

In order to launch an experiment, launch experiment.py

python experiment.py

with the following arguments:

• -d, --datasets with the dataset name to process from {"breast", "ads", "adult", "farm", "mnist17", "mnist49", "mnist56"}.
• -e, --experiments with either "landmarks_based", "greedy_kernel" or both.
• -l, --landmarks-method with either "random" or "clustering" to select the landmarks selection method for the landmarks_based experiment.
• -n, --n-cpu with the desired number of cpus to be used or "-1" to use all available.

Here is an example:

python experiment.py -d breast -e landmarks_based greedy_kernel -l random -n -1

Of note, to change the various parameters explored in the experiments, modify the values in experiment.py hps dictionnary.

BiBTeX

@inproceedings{letarte2019pseudo,
  title={Pseudo-Bayesian Learning with Kernel Fourier Transform as Prior},
  author={Letarte, Ga{\"e}l and Morvant, Emilie and Germain, Pascal},
  booktitle={The 22nd International Conference on Artificial Intelligence and Statistics},
  pages={768--776},
  year={2019}
}