A Simple Async FL Simulation Framework

A simple and easy-to-extend asynchronous federated learning (AFL) simulation framework.

The asynchronous simulation is based on a prior queue.

• Insert training-finished clients into the prior queue based on supposed finished time
• Pop the first item in the prior queue
• Update the simulated wall-clock time based on the item

We also provide simulation of synchronous federated learning like FedAvg.

Getting Started

• Download this project

git clone https://github.com/boyi-liu/Asynchronous-Federated-Learning-Simulation.git

• Install required packages
• Partition datasets

cd dataset
python generate_cifar10.py noniid balance dir

• Config hyperparameters

There are two places to config, one in /script/config.yaml, another in utils/options.py. The priority follows: args>yaml. If you config hyperparameters in utils/options.py, it will overwrite that in args.

• Run evaluation

cd script
bash run.sh {your_suffix}

How to extend

Create a new file

Create a new {your_algorithm}.py file inside trainer.alg

Extend the Client and Server

If you are working on a synchronous FL algorithm, just extend the Client and Server class in trainer.base

from trainer.base import BaseServer, BaseClient

class Client(BaseClient):
    def __init__(self, id, args, dataset):
        super().__init__(id, args, dataset)

class Server(BaseServer):
    def __init__(self, id, args, dataset, clients):
        super().__init__(id, args, dataset, clients)

Otherwise, you may extend the Client and Server in trainer.asyncbase

from trainer.asyncbase import AsyncBaseServer, AsyncBaseClient

class Client(AsyncBaseClient):
    def __init__(self, id, args, dataset):
        super().__init__(id, args, dataset)

class Server(AsyncBaseServer):
    def __init__(self, id, args, dataset, clients):
        super().__init__(id, args, dataset, clients)

Config your hyperparameters

For algorithm-specific hyperparameters, it is recommended to add a add_args() function inside your file

def add_args(parser):
    parser.add_argument('--{your_param}', type=int, default=1)
    return parser.parse_args()

And all general args could be found in utils/options.py

Implement your algorithms

We claim that each algorithm should overwrite the function run(), because it stands for the main workflow of your algorithm.

You can overwrite or add any function as you want then.

Acknowledgements

The data partitioning module is adopted from PFLlib.