Optimal Client Sampling for Federated Learning

This is a Python 3 implementation of Federated Averaging (FedAvg) algorithm with Optimal Client Sampling. The code is based on TensorFlow Federated (TFF) and is an extension of simple FedAvg example provided in TFF examples. For detailed description of the method, please read our manuscript.

Install and Test Dependencies

Set up a new environment and install dependencies:

conda create -n fl python=3.7
conda activate fl
pip install tensorflow_federated==0.16.1
pip install nest_asyncio==1.4.0

For the EMNIST experiments, unbalanced datasets modified from the EMNIST dataset can be downloaded here. They are expected to be located in dataset directory.

Run the following command to test dependencies:

python emnist_fedavg_main_cookup.py --total_rounds 2

Details on the parameters can be found in the scripts.

Reference

In case you find the method or code useful for your research, please consider citing

@article{chen2022optimal,
 title={Optimal Client Sampling for Federated Learning},
 author={Chen, Wenlin and Horvath, Samuel and Richtarik, Peter},
 journal={Transactions on Machine Learning Research},
 year={2022}
}

License