joseph-rance/oort-flower

Flower implementation of https://www.usenix.org/system/files/osdi21-lai.pdf

Flower implementation of Oort client selection

Joseph Rance | Federated Learning final project

This codebase contains experiments to test a flower implementation of the Oort client selection algorithm. This is part of my Federated Learning module project. The original paper on Oort can be found at:

https://www.usenix.org/system/files/osdi21-lai.pdf

---

Installation

To run this project, first modify lines 204 and 205 of project/task/cifar10/dataset_preparation.py to point to a suitable location to save the CIFAR10 dataset. Then run:

pip install .
bash run_scripts/download_cifar10.sh
bash run_scripts/launch.sh

The default configuration uses random sampling with $\sigma=2$. Oort can be activated by modifying line 198 of project/main.py.

Structure

This codebase closely follows the structure of the CaMLSys template.

THE MAIN CONTRIBUTION OF THIS PROJECT IS IN project/fed/server/oort_client_manager.py.

The main additions are discussed in my technical report, and can be found in project/task/cifar10 (CIFAR10 model implementation), project/task/speech (Google speech model implementation), and project/fed/server/stategy/strategy.py (custom strategy for aggregating client utility values).