SSFL-Benchmarking-Semi-supervised-Federated-Learning

Benchmarking Semi-supervised Federated Learning

Introduction

This repository includes all necessary programs to implement Semi-supervised Federated Learning of the following paper. The code runs on Python 3.7.6 with PyTorch 1.0.0 and torchvision 0.2.2. We appreciate it if you would please cite the following paper if you found the repository useful for your work:

• Z. Zhang, Z. Yao, Y. Yang, Y. Yan, J. E. Gonzalez, and M. W. Mahoney, “Benchmarking semi-supervised federated learning”.

Usage

Please first clone the this library to your local system:

git clone https://github.com/jhcknzzm/SSFL-Benchmarking-Semi-supervised-Federated-Learning.git

After cloning, please use Anaconda to install all the dependencies:

conda env create -f environment.yml

To run the main scripte "train_parallel.py", one should first use ifconfig -a to find the ip address of the machine. Then, you can train a Semi-supervised Federated Learning experiment using the following command:

python train_parallel.py [--experiment_num]  [--GPU_list] [--ip_address] [--datasetid]
optional arguments:
--experiment_num      experiment number (default: 0), each experiment number corresponds to a specific experiment setting, you can modify these settings by modifying setting_cifar/svhn/emnist.py 
--GPU_list            GPUs used for training, e.g., --GPU_list 0123456789   
--ip_address          the ip address of the machine, e.g., --ip_address 128.32.162.169
--datasetid           the id of the datasets (default: 0), datasetid = 0/1/2 means the Cifar-10/SVHN/EMNIST dataset is used in the experiment. 

Following is the default experiment setting of Cifar-10 (datasetid is 0):

Experiment num	K	Ck	T	Ns	R	Batch size	Epochs	Average method H	Neural network model
0	10	10	16	1000	1.0	64	300	0	ResNet-18 with group normalization
1	10	10	16	1000	0.0	64	300	0	ResNet-18 with group normalization
2	10	10	16	1000	0.2	64	300	0	ResNet-18 with group normalization
...	...	...	...	...	...	...	...	...	...

• K: User number
• Ck: The number of communication users
• T: Communication period, T ∈ {2,4,8,16,32}
• Ns: The number of labeled samples in server
• R: The non-iidness, R ∈ {0,0.2,0.4,0.6,0.8,1.0}
• H: Average method, H=0/1 means the FedAvg/Grouping-based average method is used

Assume the ip address is 128.32.162.169, one can use the following command to run the experiment in the setting of K=10, Ck=10, R=1.0, T=16 and Ns=1000 on Cifar-10:

python train_parallel.py --experiment_num 0  --GPU_list 0123467895 --ip_address 128.32.162.169 --datasetid 0

You can also run the following scripts to reproduce the results reported in Table 2 of the paper.

nohup bash Exper_GroupingMethod.sh