FedGS: Data Heterogeneity-Robust Federated Learning via Group Client Selection in Industrial IoT

Preparation

• For instructions on generating data, please go to the folder of the corresponding dataset. For FEMNIST, please refer to femnist.

• NVIDIA-Docker is required.

• NVIDIA CUDA version 10.1 and higher is required.

How to run FedGS

Build a docker image

Enter the scripts folder and build a docker image named fedgs.

sudo docker build -f build-env.dockerfile -t fedgs .

Modify /home/lizh/fedgs to your actual project path in scripts/run.sh. Then run scripts/run.sh, which will create a container named fedgs.0 if CONTAINER_RANK is set to 0 and starts the task.

Use ./preprocess.sh -s niid --sf 0.1 -k 100 -t sample --smplseed 0 --spltseed 0 in data/femnist to generate dataset and then run

chmod a+x run.sh && ./run.sh

The output logs and models will be stored in a logs folder created automatically. For example, outputs of the FEMNIST task with container rank 0 will be stored in logs/femnist/0/.

Hyperparameters

We categorize hyperparameters into default settings and custom settings, and we will introduce them separately.

Default Hyperparameters

These hyperparameters are included in utils/args.py. We list them in the table below (except for custom hyperparameters), but in general, we do not need to pay attention to them.

Variable Name	Default Value	Optional Values	Description
--seed	0	integer	Seed for client selection and batch splitting.
--metrics-name	"metrics"	string	Name for metrics file.
--metrics-dir	"metrics"	string	Folder name for metrics files.
--log-dir	"logs"	string	Folder name for log files.
--use-val-set	None	None	Set this option to use the validation set, otherwise the test set is used. (NOT TESTED)

Custom Hyperparameters

These hyperparameters are included in scripts/run.sh. We list them below.

Environment Variable	Default Value	Description
CONTAINER_RANK	0	This identify the container (e.g., fedgs.0) and log files (e.g., logs/femnist/0/output.0).
BATCH_SIZE	32	Number of training samples in each batch.
LEARNING_RATE	0.01	Learning rate for local optimizers.
NUM_GROUPS	10	Number of groups.
CLIENTS_PER_GROUP	10	Number of clients selected in each group.
SAMPLER	gbp-cs	Sampler to be used, can be random, brute, bayesian, probability, ga and gbp-cs.
NUM_SYNCS	50	Number of internal synchronizations in each round.
NUM_ROUNDS	500	Total rounds of external synchronizations.
DATASET	femnist	Dataset to be used, only FEMNIST is supported currently.
MODEL	cnn	Neural network model to be used.
EVAL_EVERY	1	Interval rounds for model evaluation.
NUM_GPU_AVAILABLE	2	Number of GPUs available.
NUM_GPU_BEGIN	0	Index of the first available GPU.
IMAGE_NAME	fedgs	Experimental image to be used.

NOTE: If you wish to specify a GPU device (e.g., GPU0), please set NUM_GPU_AVAILABLE=1 and NUM_GPU_BEGIN=0.

NOTE: This script will mount project files /home/lizh/fedgs from the host into the container /root, so please check carefully whether your file path is correct.

Visualization

The visualizer metrics/visualize.py reads metrics logs (e.g., metrics/metrics_stat_0.csv and metrics/metrics_sys_0.csv) and draws curves of accuracy, loss and so on.

Note

This project is licensed under the terms of the apache-2.0 license.

Reference

• This demo is implemented on LEAF-MX, which is a MXNET implementation of the well-known federated learning framework LEAF.

• Li, Zonghang, Yihong He, Hongfang Yu, et al.: "Data heterogeneity-robust federated learning via group client selection in industrial IoT." IEEE Internet of Things Journal, vol. 9, no. 18, pp. 17844-17857. IEEE, 2021, doi: 10.1109/JIOT.2022.3161943.

• If you get trouble using this repository, please kindly contact us. Our email: lizhuestc@gmail.com