This is an implementation of HeteroFL: Computation and communication efficient federated learning for heterogeneous clients using MPI and Pytorch.

The official implementation of this paper is here.

• Global hyperparameters are defined in config.yml
• server_main.py is the main file to run server code
• client_main.py is the main file to run client code

If you want to run HeteroFL with 10 clients selected in each epoch, you can input this command in the console:

mpiexec --oversubscribe -n 1 python server_main.py : -n 10 python client_main.py

Each client and the server runs as a process, which communicates with each other by MPI.

Make sure that the value of hyperparameter "active_client_num" in config.yml equals the number of clients in the command.