Basic PPFL Training#

APPFL provides users with the capabilities of simulating and training PPFL on either a single machine, a cluster, or multiple heterogeneous machines. We refer

simulation as running PPFL experiments on a single machine or a cluster without actual data decentralization
training as running PPFL experiments on multiple (heterogeneous) machines with actual decentralization of client datasets

Hence, we describe two types of PPFL run:

Simulating PPFL is useful for those who develop, test, and validate new models and algorithms for PPFL, whereas Training PPFL for those who consider actual PPFL settings in practice.

Sample template#

Note

Before reading this section, it is highly recommended to check out Tutorials for more detailed examples in notebooks.

For either simulation or training, a skeleton of the script for running PPFL can be written as follows:

from appfl import *
from appfl.config import Config
from omegaconf import OmegaConf

def main():

    # load default configuration
    cfg = OmegaConf.structured(Config)

    # change configuration if needed
    ...

    # define model, loss, and data
    model = ... # user-defined model
    loss_fn = ... # user-defined loss function
    data = ... # user-defined datasets

    # The choice of PPFL runs

if __name__ == "__main__":
    main()

Some remarks are made as follows:

Line 8 loads the default configuration for PPFL run. See How to set configuration for more details.
User-defined model, loss, and data can be read as in lines 14 to 16; see How to define model and How to define datasets for more details.
Depending on our choice of PPFL run, we then call API functions in line 12.