This repository contains the code associated with the manuscript SAE: Sequential Anchored Ensembles [link]. This idea has been developed in the context of the Approximate Inference in Bayesian Deep Learning NeurIPS2021 competition and achieved the
The code has been tested with the version 3.7.11 of python. The requirements can be found in the file requirements.txt. To install all the dependencies, run the following command.
pip install -r requirements.txt
Instructions for downloading the data can be found here.
The pipeline (training and evaluation) can be executed by running the command
python main.py --config_file "config file path" --index "index"
The structure of a config file and examples of config files can be found here. The index argument is optional and is used to run several times the same experiments.
This repository contains the code to train
- Simple neural networks
- Deep ensembles and anchored ensembles (AE)
- Sequential anchored ensembles (SAE)
- Graphical anchored ensembles (an attempt to improve SAE not included in the paper)
A description of all those models can be found here.
@article{delaunoy2021sae,
title={SAE: Sequential Anchored Ensembles},
author={Delaunoy, Arnaud and Louppe, Gilles},
journal={arXiv preprint arXiv:2201.00649},
year={2021}
}