FedGroup

The source code of the Arxiv preprint article:

FedGroup: Efficient Clustered Federated Learning via Decomposed Data-Driven Measure

NOTE: The code base of FedGroup is FedProx.

🎁 Why not try the wholly new FlexCFL, which added many exciting improvements and technical fixes.

Overview

FedGroup can simulate following (Clustered) Federated Learning framework:

• FedAvg & FedSGD -> Communication-Efficient Learning of Deep Networks from Decentralized Data
• FedProx -> Federated optimization in heterogeneous networks
• FedGrop & FedGrouProx -> FedGroup: Efficient Clustered Federated Learning via Decomposed Data-Driven Measure
• IFCA -> An Efficient Framework for Clustered Federated Learning
• FeSEM -> Multi-center federated learning

Requirement

Python packages:

• Tensorflow (>2.0)
• Jupyter Notebook
• scikit-learn
• matplotlib
• tqdm

You need to download the dataset (e.g. FEMNIST, MNIST, Sent140, Synthetic) and specify a GPU id follow the guidelines of FedProx.

The directory structure of the datasets should look like this:

GrouProx-->data-->mnist-->data-->train--> ***train.json
               |              |->test--> ***test.json
               |
               |->nist-->data-->train--> ***train.json
               |                     |-> ***test.json
               |
               |->sent140--> ...
               |
               ...

Quick Start

Just run GrouProx_notebook.ipynb.

You can modify the configurations by directly modifying the code of GrouProx_notebook.ipynb. The common hyperparameters of FedGroup is:

# Name of dataset, should be list in GrouProx-->data-->...
params['dataset'] = 'sent140' 

# Name of model, should be list in GrouProx-->flearn-->models-->params['dataset']-->...
params['model'] = 'stacked_lstm'

# Name of optimizer, should be one of ['fedavg', 'fedprox', 'grouprox']
params['optimizer'] = 'grouprox'

# The dropout rate as demonstrated in the FedProx paper
params['drop_percent'] = 0

# Total communication rounds
params['num_rounds'] = 200

# Local epoch E, same as FedProx
params['num_epochs'] = 20

# Local mini-batch size, same as FedProx
params['batch_size'] = 10

# Evaluate the group model every $params['eval_every'] rounds
params['eval_every'] = 1

# The number of clients K selected per round
params['clients_per_round'] = 20

# Random seed
params['seed'] = 233

# Inter-group learning rate
params['agg_lr'] = 0.01

# Number of 'Groups'
params['num_group'] = 5

# Some specific hyperparameters of FedGroup
if params['optimizer']  == 'grouprox':
  # Whether to use Proximal method, True for FedGrouProx
  params['proximal'] = False

  # Radomly Assign Clients and Random Cluster Centers strategy, please see the paper for details 
  params['RAC'] = False
  params['RCC'] = False

  # The Group may be empty if True
  params['allow_empty'] = True
  
  # We implement IFCA and FeSEM base on grouprox
  # Set 'ifca' or 'fesem' to True to enable it. 
  params['ifca'] = False
  params['fesem'] = False

Experimental Results

All evaluation results will save in the GrouProx-->results-->... directory as csv format files.

Reference

Please cite the preprint version of FedGroup if the code helped your research 😊

• FedGroup: Efficient Clustered Federated Learning via Decomposed Data-Driven Measure

BibTeX

@article{duan2020fedgroup,
  title={FedGroup: Efficient Clustered Federated Learning via Decomposed Data-Driven Measure},
  author={Duan, Moming and Liu, Duo and Ji, Xinyuan and Liu, Renping and Liang, Liang and Chen, Xianzhang and Tan, Yujuan},
  journal={arXiv preprint arXiv:2010.06870},
  year={2020}
}