MB-Conjugate-Gradient-DP

Implementation of Mini-Batch Conjugate Gradient with Data Persistency (MBCG-DP) proposed in

Lapucci M. and Pucci D. - Effectively Leveraging Momentum Terms in Stochastic Line Search Frameworks for Fast Optimization of Finite-Sum Problems - arXiv pre-print (2024)

Installation

In order to execute the code, you will need a working Anaconda installation. We suggest the creation of a new conda environment with Python 3.11.3 or above. Requirements can be installed through:

pip install -r requirements.txt

Usage

In order to run the experiments, execute the following:

python trainval.py [options]

The following arguments shall be specified:

Short Option	Long Option	Type	Description	Default
-e	--exp_group_list	str	List of experiments to be executed	None (required)
-d	--datadir	str	Path to save the datasets (downloaded automatically)	None (required)
-sb	--savedir_base	str	Path to save the output logs	None (required)

The file exp_configs.py specifies the experiments avaialable and their configurations.

An example usage is:

python trainval -e mushrooms -d Datasets -sb Results

The PlotResults functionalities enable the generation of plots based on logs produced during the experiments.

Acknowledgements

Our experimental framework is a simplified version of that of SLS and PoNoS.

In case you employed our code for research purposes, please cite:

@misc{lapucci2024effectivelyleveragingmomentumterms,
      title={Effectively Leveraging Momentum Terms in Stochastic Line Search Frameworks for Fast Optimization of Finite-Sum Problems}, 
      author={Matteo Lapucci and Davide Pucci},
      year={2024},
      eprint={2411.07102},
      archivePrefix={arXiv},
      primaryClass={math.OC},
      url={https://arxiv.org/abs/2411.07102}, 
}