LightingFlower

Pre-packaged federated learning framework using Flower and PyTorch-Lightning.

Installation

To install this library, simply run the following command:

pip install lightningflower

Installing the lightningflower framework should automatically install suitable versions of [Flower] and [PyTorch-Lightning].

##Features

Integrated Argument Parser

LightningFlower provides integrated argument parsing for data, server and client handling:

# initialize the argument parser
parser = ArgumentParser()

# Data-specific arguments like batch size
parser = LightningFlowerData.add_data_specific_args(parser)

# Trainer-specific arguments like number of epochs
parser = pl.Trainer.add_argparse_args(parser)

# Client-specific arguments like host address
parser = LightningFlowerClient.add_client_specific_args(parser)

# Parse arguments
args = parser.parse_args()

LightningFlowerBaseStrategy

A basic strategy to enable saving and loading of model weights as well as resuming the training procedure.

Full Pytorch-Lightning Trainer integration

LightningFlower supports full use of the Pytorch-Lighting training routines on federated Flower clients:

# Configure the client trainer
parser = pl.Trainer.add_argparse_args(parser)
args = parser.parse_args()

# Define a Pytorch-Lighting compatible model to train
model = pl.RandomModel()

# Create a client, pass Trainer configuration to and model to client
LightingFlowerClient(model=model, trainer_args=args, ...)

Federated transfer learning / backbone support

LightningFlower enables transfer learning by only transmitting trainable model parameters from server to clients, saving bandwidth and computation time.

ToDo

This is a work in progress.