HT-FL: Hybrid Training Federated Learning for Heterogeneous Edge-based IoT Networks

Dependencies and Setup

All code runs on Python 3.10.10 using PyTorch version 2.0.0 In addition, you will need to install

• torch
• numpy
• torchvision
• pickle
• argparse
• torchsummary
• random

Organization

• HT-FL.py: The main algorithm code.
• test_iid.py: Test the distribution of data.
• modelsize.py: Calculate the size of local model.
• 'run_exps2.sh': Code of our experiments.
• graphs/: Different type of hubs communication graph.
• Net/: Realization of different local models.
• results/: Results of our stimulation.

Description of main parameters

This is only a stimulation code for "HT-FL: Energy Efficient Federated Learning for IoT Applications with Non-IID Data" and have many aspects to improve. Experiments we run are written in run_exps2.sh. Key params are given as follows:
--data: dataset
--model: training model
--hubs: hub amount
--workers: worker amount
--tau: local training rounds
--q: global aggregation frequency
--graph: hub communication graph type
--epochs: training epoches
--batch: training batch size
--prob: probability for worker to attend training
--fed & --non_iid: data distribution paradigm
--percentage: controlling param
--num_class: amount of class on a worker
--uniform: uniform data percentage
--dir: dirichlet param