Federated Learning Manager

The py-fl-manager project provides easy integration with NVFlare.

Warning

This project is currently in alpha.

Note

The documentation is in progress. Some sections may be incomplete, outdated, or missing details.

Running (example)

The base docker image used is fl-manager:dev, so make dev/build should be executed first.

See available examples at examples. Currently, only nvflare_mnist is available.

Access the example folder. There is a project.yml with the FL project configuration and a set of folders with jobs examples.

1. Start by provisioning the workspace with nvflare provision.

• For now, (04/02/2025):
  • Copy and paste the job's folders inside transfer from the admin folder.
  • Change the PYTHON_EXECUTABLE variable inside the .env.
  • Change fed_admin.json overseer endpoint from 'server1' to 'localhost'.

2. Run docker compose build and docker compose up to initialize the simulation workspace.
3. Login into the admin CLI (startup.sh) script.
4. Submit job!

For (1), the following command can be run to achieve it: EXAMPLE=nvflare_mnist make dev/provision.

Code Quality

This repository provides a configuration for SonarScanner. Some missing parameters that should be set before running are the following:

sonar.projectKey=${PROJECT_KEY}
sonar.projectName=${PROJECT_NAME}
sonar.token=${SONAR_TOKEN}

A self-hosted SonarQube can be deployed to perform a local analysis.

1. Deploy SonarQube with docker run --name sonarqube-custom -p 9000:9000 sonarqube:community.
2. Access SonarQube and configure a project (default credentials admin:admin).
3. Create a Project and copy the token.
4. Execute tests with coverage using make dev/test-cov.
5. Change in coverage.xml the <source>/opt/project/fl_manager</source> root to <source>/usr/src/fl_manager</source>, this enables the coverage support in this local scan.
6. Fill the properties file with the missing values.
7. Run docker run --rm --network host -e SONAR_HOST_URL="http://localhost:9000" -v $(pwd):/usr/src sonarsource/sonar-scanner-cli

Versioning

All packages follow semantic versioning. This means, all versions with the same major must be compatible between them. This way, packages can evolve independently between them. When a major release needs to be done, all packages must be bumped to the new major.

Contributing

Creating a new package

1. Execute make uv_runtime/run.
2. Create the necessary folders if they do not exist.
3. Run uv init --no-readme --no-pin-python --lib --name $NAME $PATH.
4. Mark src as source root. Create subfolders for namespace.
5. Add build target (in new package .toml) to the namespace package.

Configuration

We use pre-commit hooks.

Setup pre-commit hooks

Probably there will be other methods to install pre-commit we use the following steps:

1. Install pipx with (system's) pip (i.e. pip install pipx).
2. Install pre-commit with pipx (i.e. pipx install pre-commit).
3. Follow the instructions that are printed after the execution of the step 2 command, to ensure that pre-commit is in the path.
4. Run pre-commit install.

Check .pre-commit-config.yaml to check what hooks are installed!