FedNH

Implementation of the [FedNH23] algorithm.

References

[FedNH23]

Yutong Dai, Zeyuan Chen, Junnan Li, Shelby Heinecke, Lichao Sun, Ran Xu. Tackling Data Heterogeneity in Federated Learning with Class Prototypes. In AAAI (2023). URL: https://arxiv.org/abs/2212.02758

Classes included in fluke.algorithms.fednh

ProtoNet

Neural network with prototypes.

FedNHClient

FedNHServer

FedNH

Classes

class fluke.algorithms.fednh.ProtoNet(encoder: Module, n_protos: int, proto_size: int)[source]

Bases: Module

Neural network with prototypes. The network is composed of an encoder and a set of prototypes. The prototypes are initialized as orthogonal vectors and are not trainable. In the forward pass, the network computes the cosine similarity between the normalized embeddings (output of the encoder) and the prototypes.

Parameters:

  • encoder (Module) – Encoder network.

  • n_protos (int) – Number of prototypes.

  • proto_size (int) – Size of the prototypes.

class fluke.algorithms.fednh.FedNHClient(index: int, model: Module, train_set: FastDataLoader, test_set: FastDataLoader, optimizer_cfg: OptimizerConfigurator, loss_fn: Module, local_epochs: int, n_protos: int, fine_tuning_epochs: int = 0, clipping: float = 5, **kwargs: dict[str, Any])[source]

Bases: Client

class fluke.algorithms.fednh.FedNHServer(model: Module, test_set: FastDataLoader, clients: Iterable[Client], weighted: bool = True, n_protos: int = 10, rho: float = 0.1)[source]

Bases: Server

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

Bases: CentralizedFL