omkaark / simple-federated-learning

Federated Learning for local networks (photo credit: Parth Sareen's expert prompt engineering talent)

Federated Learning Setup Guide

This guide provides step-by-step instructions on how to set up and run a federated learning system with a central leader and multiple learners.

Prerequisites

Before you begin, ensure you have the following requirements installed:

• Python 3.6+
• pip
• Access to a terminal or command line interface

Installation

Install Required Packages

Navigate to the project directory and install the required Python packages (preferrable in an environment:

pip install -r requirements.txt

Prepare Model Artifacts

Add your model.py file to the model_artifacts directory. This file should define the necessary variables including device, model, criterion, and optimizer_function. Check out my resnet-18 example model at model_artifacts/model.py If you want to change the model from Resnet to anything else, make sure you change it here: model = <MODEL-DEF-GOES-HERE>

Configure Data Loader

Modify the data loader in utils.py to suit your setup or continue using the default setup configured for CIFAR-10.

Network Configuration

Ensure all computers (leader and learners) are connected to the same network to facilitate communication.

Usage

Start the Leader

Run the leader script specifying the number of learners to wait for:

python leader.py --learner-count X

Replace X with the number of learners you want (# of computers you want to run training on).

Start the Learners

Once the leader is running, it will display its address in the format ADDRESS:PORT. Use this address to start each learner:

python learner.py --leader-address ADDRESS:PORT

Training Process

• After the specified number of learners have joined, training will start automatically.
• The system will train the model distributed across all learners and periodically synchronize their updates with the leader.
• Upon completion of the training, the validation accuracy will be printed, and a model.pth file will be generated. This file contains the trained model binary.

Final Notes

• The training session's progress and results will be logged in the terminal.
• For troubleshooting and detailed logs, refer to the log files generated in the logs directory.