To run the scenarios locally please install the requirements

pip install -r requirements.txt

If you want to run the experiments in RPIs please follow the rpi_setup_comands.md

If you want to run the experiments in docker please run docker compose up.

To run locally you only need the following commands

mpirun -np 4 python mpi_training.py -d dataset/one_hot_encoding/

or

mpirun -np 4 python mpi_custom_training.py -d dataset/one_hot_encoding/

If you want to run the single_host setting you can run

python model.py

To run on docker you will need to connect to the master container

docker exec -it --user mpiuser dl_costs-master-1 bash

Connect one time to every worker to confirm the fingerprint of the server.

ssh worker1 and then exit

After this you can run the command

mpirun -np 4 -hostfile hostfile python mpi_training.py -d dataset

or

mpirun -np 4 -hostfile hostfile python mpi_custom_training.py -d dataset

inside the code folder.

To change the hyperparameters please confirm the available options in the mpi_training.py and mpi_custom_training.py files

The results obtained with the different hyperparameters are presented in the paper_results folder the translation for the folder names is 10_g_50_l -> (decentralized optimization with 10 global epochs and 50 local epochs) and fed_sgd_64 -> (centralized optimization with batch size 64)

• Rafael Teixeira - rgtzths

This project is licensed under the MIT License - see the LICENSE file for details

If you use this code please site our work: Teixeira, Rafael & Antunes, Mário & Gomes, Diogo & Aguiar, Rui. (2023). The learning costs of Federated Learning in constrained scenarios. 10.1109/FiCloud58648.2023.00011.