APFL

Implementation of the APFL [APFL20] algorithm.

References

[APFL20]

Yuyang Deng, Mohammad Mahdi Kamani, and Mehrdad Mahdavi. Adaptive Personalized Federated Learning. In arXiv (2020). URL: https://arxiv.org/abs/2003.13461

Classes included in fluke.algorithms.apfl

APFLClient

APFLServer

APFL

Classes

class fluke.algorithms.apfl.APFLClient(index: int, model: Module, train_set: FastDataLoader, test_set: FastDataLoader, optimizer_cfg: OptimizerConfigurator, loss_fn: Module, local_epochs: int = 3, fine_tuning_epochs: int = 0, lam: float = 0.25, **kwargs: dict[str, Any])[source]

Bases: PFLClient

class fluke.algorithms.apfl.APFLServer(model: Module, test_set: FastDataLoader, clients: Iterable[Client], weighted: bool = False, tau: int = 3, **kwargs: dict[str, Any])[source]

Bases: Server

aggregate(eligible: Iterable[Client], client_models: Iterable[Module]) → None[source]

Aggregate the models of the eligible clients every hyper_params.tau rounds.

Parameters:

  • eligible (Iterable[Client]) – The clients that are eligible to participate in the aggregation.

  • client_models (Iterable[Module]) – The models of the clients to aggregate.

class fluke.algorithms.apfl.APFL(n_clients: int, data_splitter: DataSplitter, hyper_params: DDict | dict[str, Any], **kwargs: dict[str, Any])[source]

Bases: PersonalizedFL