Personalized Federated Learning with Exact Distributed Stochastic Gradient Descent Updates

Source code for paper "Personalized Federated Learning with Exact Distributed Stochastic Gradient Descent Updates" at ALA 2023 - Workshop at AAMAS 2023.

Poster

Requirements

Ubuntu 18.04 (OS)
Anaconda3

To install requirements, first download Anaconda3:

wget https://repo.anaconda.com/archive/Anaconda3-2020.11-Linux-x86_64.sh
chmod +x Anaconda3-2020.11-Linux-x86_64.sh
su root
cd ~/anaconda3/bin
./conda init bash

then run the following command to create the anaconda environment ( for cpu only, we will create a section for gpu when we upload the code to github):

conda env create -f mytf23.yml

Create basic folders and place main.py into the exp folder:

mkdir exp out out/results

Hardware Requirements

We recommend a GPU with memory >= 11GB.

Datasets

All datasets (MNIST,Fashion-MNIST,EMNIST,CIFAR-10,Omniglot) are downloaded automatically from our given code in main.py through the tensorflow_datasets library.

Training

First we will explain the basic setup of the input arguments:

inner_stepss (int): Number of local gradient updates at the client side
rounds (int): Number of communication rounds
dataset_name (str): Dataset to use e.x. cifar10, mnist, fashion_mnist, emnist, omniglot
algo (str): Algorithm to use i.e fedavg, pflego, feper
num_clients (int): Number of clients
client_lr (float): Learning rate of client
server_lr (float): Learning rate of server
users_fraction (float): The fraction of users we sample at each round
user_gpu (str): Add this argument only if you want to use a gpu, i.e. --use_gpu True
scenario (str): RoundRobin partition i.e. highpers for high personalization, mediumpers for medium personalization, and nopers for no personalization
optimizer (str): The optimizer of the server, i.e sgd or adam

Use these commands to run the code:

First, activate the environment

source activate mytf23

For GPU,

PFLEGO with Omniglot

python -u main.py  --dataset_name omniglot --inner_steps 50 --num_clients 50 --algo pflego --rounds 5000 --client_lr 0.007 --server_lr 0.003 --scenario highpers --use_gpu True

FedPer with CIFAR-10

python -u main.py  --dataset_name mnist --inner_steps 50 --num_clients 100 --algo fedper --rounds 200 --client_lr 0.007 --server_lr 1.0 --scenario highpers --use_gpu True

For CPU,

FedAvg with Mnist

python -u main.py  --dataset_name mnist --inner_steps 50 --num_clients 100 --algo fedavg --rounds 200 --client_lr 0.007 --server_lr 1.0 --scenario highpers