Pybiscus is a simple tool to perform Federated Learning on various models and datasets. It aims at automated as much as possible the FL pipeline, and allows to add virtually any kind of dataset and model.
Pybiscus is built on top of Flower, a mature Federated Learning framework; Typer (script and CLI parts) and Lightning/Fabric for all the Machine Learning machinery.
It is managed using the uv package manager
The architecture is built around a plugin system and extensible hooks, and we are continuously evolving the framework to maximize flexibility and adaptability across all aspects of the federated learning process.
In addition to the pybiscus command-line application, we have introduced two key tools:
-
a lightweight client agent that dynamically generates the configuration interface;
-
a session manager that synchronizes configuration of the server and clients, and handles the orchestration of the training session.
From experiments on one machine to sessions across several organisations, the usage scenarios describe what Pybiscus brings in each case, how to set it up and its limits.
uv is a python project and package manager ( https://github.com/astral-sh/uv )
extend your user's path to include the bin directory
source ./extend_path.shdownload the required packages
uv syncYou can find example launch scripts in ./launch/uv subdirectories
Documentation is available at docs.
If you are interested in contributing to the Pybiscus project, start by reading the Contributing guide.
Pybiscus is on active development at Thales, both for internal use and on some collaborative projects. One major use is in the European Project PAROMA-MED, dedicated to Federated Learning in the context of medical data distributed among several Hospitals. Pybiscus is also used in the European Defence Fund project NEWSROOM, dedicated to cyber situational awareness, where collaborative and federated learning serve the detection of attacks.
The License is Apache 2.0. You can find all the relevant information here LICENSE