FedLUAR

This repository contains the S/W framework used for all the experiments in the below paper. '[Jisoo Kim, Sungmin Kang, and Sunwoo Lee, Layer-wise Update Recycling for Communication-Efficient Federated Learning]'

Software Requirements

• tensorflow2 (< 2.16.0)
• tensorflow_datasets
• python3
• mpi4py
• tqdm

Instructions

Training

1. Set hyper-parameters properly in config.py.
2. Put the dataset files in the top directory of this program. The directory name should be the same as dataset in config.py.
3. Run training.

mpirun -n 8 python main.py

Output

This program evaluates the trained model after every epoch and then outputs the results as follows.

1. loss.txt: An output file that contains the training loss for every epoch.
2. acc.txt: An output file that contains the validation accuracy for every epoch.
3. ./checkpoint: The checkpoint files generated after every epoch. This directory is created only when checkpoint is set to 1 in config.py.

Results

We will provide a few key experimental results here once the papers are published.

Supported Federated Learning Features

• FedAvg
• FedLUAR

Supported Datasets

• CIFAR-10
• CIFAR-100
• FEMNIST
• AG News

Questions / Comments

• Jisoo Kim (starprin3@gmail.com)
• Sunwoo Lee (sunwool@inha.ac.kr)