README

This is the code for the IEEE MLSP 2024 workshop paper "Joint Optimization of Piecewise Linear Ensembles" [arxiv] [IEEE (coming soon)]. For now, please cite as

@misc{raymond2024,
    title={Joint Optimization of Piecewise Linear Ensembles},
    author={Matt Raymond and Angela Violi and Clayton Scott},
    year={2024},
    eprint={2405.00303},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2405.00303},
}

The associated GitLab issue tracker is currently limited to internal use. Please email the current maintainer listed on PyPI with any questions or concerns, and they will open an issue on your behalf.

Installation

Installation via PyPI

pip install joplen

Installation from source

NOTE: pip install -e . will only work if you have setuptools v64 or higher and pip version 24 or higher.

Clone the repository to your local machine, then run the following commands (which assume that you already have Conda installed):

conda create --prefix ./my_env python=3.10
conda activate ./my_env
conda config --set env_prompt '({name}) '

pip install -r requirements.txt
pip install -e .

JAX must be installed manually according to this link because the installation is hardware-dependent. Please follow these instructions to install JAX.

Usage

Each module has example usage. You can run them by executing the module as a script. Note that single-task JOPLEn is much more modular than the multitask implementation. This is for practical reasons, but there's no reason it couldn't be made more modular.

python -m JOPLEn.singletask # single-task joplen
python -m JOPLEn.multitask # multi-task joplen
python -m JOPLEn.competing # Friedman ensemble refitting

Original implementation

To see the original implementation for the workshop submission, please see https://gitlab.eecs.umich.edu/mattrmd-public/joplen-repositories/joplen-mlsp2024.