FEDn is a modular and model agnostic framework for hierarchical federated machine learning which scales from pseudo-distributed development to real-world production networks in distributed, heterogeneous environments. For more details see https://arxiv.org/abs/2103.00148.
- Scalable and resilient. FEDn is highly scalable and resilient via a tiered architecture where multiple aggregation servers (combiners) form a network to divide up the work to coordinate clients and aggregate models. Recent benchmarks show high performance both for thousands of clients in a cross-device setting and for large model updates (1GB) in a cross-silo setting. FEDn has the ability to recover from failure in all critical components.
- ML-framework agnostic. Model updates are treated as black-box computations. This means that it is possible to support any ML model type or framework. Support for Keras and PyTorch is available out-of-the-box.
- Security. A key feature is that clients do not have to expose any ingress ports.
- Track events and training progress. FEDn logs events in the federation and tracks both training and validation progress in real time. Data is logged as JSON to MongoDB and a user can easily make custom dashboards and visualizations.
- UI. A Flask UI lets users see client model validations in real time, as well as track client training time distributions and key performance metrics for clients and combiners.
Clone this repository, locate into it and start a pseudo-distributed FEDn network using docker-compose:
docker-compose up
Navigate to http://localhost:8090. You should see the FEDn UI, asking you to upload a compute package. The compute package is a tarball of a project. The project in turn implements the entrypoints used by clients to compute model updates and to validate a model.
Locate into 'examples/mnist-pytorch'.
Start by initializing a virtual enviroment with all of the required dependencies for this project.
bin/init_venv.sh
Now create the compute package and a seed model:
bin/build.sh
Uploade the generated files 'package.tar.gz' and 'seed.npz' in the FEDn UI.
The next step is to configure and attach clients. For this we download data and make data partitions:
Download the data:
bin/get_data
Split the data in 2 parts for the clients:
bin/split_data
Data partitions will be generated in the folder 'data/clients'.
Now navigate to http://localhost:8090/network and download the client config file. Place it in the example working directory.
To connect a client that uses the data partition 'data/clients/1/mnist.pt':
docker run \ -v $PWD/client.yaml:/app/client.yaml \ -v $PWD/data/clients/1:/var/data \ -e ENTRYPOINT_OPTS=--data_path=/var/data/mnist.pt \ --network=fedn_default \ ghcr.io/scaleoutsystems/fedn/fedn:develop-mnist-pytorch run client -in client.yaml --name client1
You are now ready to start training the model at http://localhost:8090/control.
To scale up the experiment, refer to the README at 'examples/mnist-pytorch' (or the corresponding Keras version), where we explain how to use docker-compose to automate deployment of several clients.
You will find more details about the architecture, compute package and how to deploy FEDn fully distributed in the documentation:
All pull requests will be considered and are much appreciated. Reach out to one of the maintainers if you are interested in making contributions, and we will help you find a good first issue to get you started. For more details please refer to our contribution guidelines.
Community support in available in our Discord server.
If you use FEDn in your research, please cite:
@article{ekmefjord2021scalable,
title={Scalable federated machine learning with FEDn},
author={Ekmefjord, Morgan and Ait-Mlouk, Addi and Alawadi, Sadi and {\AA}kesson, Mattias and Stoyanova, Desislava and Spjuth, Ola and Toor, Salman and Hellander, Andreas},
journal={arXiv preprint arXiv:2103.00148},
year={2021}
}
FEDn is licensed under Apache-2.0 (see LICENSE file for full information).