adriaj / pyfw-lasso

Implementation of the Polyatomic Frank-Wolfe algorithm to solve the LASSO problem.

pyfw-lasso

This python package provides two Frank-Wolfe algorithms to solve the LASSO problem

These algorithms are:

• the Vanilla Frank-Wolfe algorithm,
• the Polyatomic Frank-Wolfe algorithm.

Polyatomic Frank-Wolfe

The Polyatomic Frank-Wolfe algorithm has been proposed and described there: https://doi.org/10.1109/LSP.2022.3149377 (pre-print version available).

It is an optimization algorithm that builds upon the classical Frank-Wolfe algorithm by allowing to place multiple atoms at each iteration. This results is a significantly faster convergence. An additional approximate correction step is used in order to accelerate further more the solving time while preserving the accuracy of the solution.

Installation

Instructions are provided in the howtoinstall.txt file.

$ git clone [email protected]:AdriaJ/pyfw-lasso.git
$ cd pyfw-lasso
$ conda create --name pyfwl --strict-channel-priority --channel=conda-forge --file=conda/requirements.txt
$ conda activate pyfwl
$ pip install -e .
$ pytest

Citation

For citing this package, please refer to the Signal Processing Letters article: https://doi.org/10.1109/LSP.2022.3149377 .

@article{jarret2022,
    author={Jarret, Adrian and Fageot, Julien and Simeoni, Matthieu},
    journal={IEEE Signal Processing Letters},
    title={A Fast and Scalable Polyatomic Frank-Wolfe Algorithm for the LASSO},
    year={2022},
    volume={29},
    pages={637-641},
    doi={10.1109/LSP.2022.3149377}
}

pyfw-lasso's Issues

Update log_var

No more "x" needed as a log var since the implementation of the sparse iterates.

Fully corrective step for Vanilla FW

Would be interesting to feature the fully corrective steps to Vanilla FW. It is probably possible to that using the reweighting function from Polyatomic FW.