fed.md

heflwr.fed

The heflwr.fed module provides direct support for aggregation and distribution operations in system heterogeneous federated learning. For federated learning under system heterogeneity, direct aggregation and distribution of parameters are not feasible. heflwr.fed provides support for PyTorch and Flower aggregation and distribution functions.

The heflwr.fed module provides APIs supporting PyTorch and Flower to facilitate research on federated learning under system heterogeneity. This module provides three functions: aggregate_layer, extract, and merge.

• With the aggregate_layer function and the reset_parameters_from_father method of members in heflwr.nn, you can quickly build a simulated federated learning system implemented using PyTorch.
• With the extract and merge functions, you can quickly build a federated learning system implemented using Flower, which can be deployed in real-world application environments without the need to focus on low-level parameter transmission serialization protocols and communication details.

aggregate_layer

aggregate_layer(global_layer: heflwr.nn.SUPPORT_LAYER,
                subset_layers: List[heflwr.nn.SUPPORT_LAYER],
                weights: List[int],
) -> None

Aggregate parameters from multiple client layers (subset_layers) into the global layer (global_layer), where the influence of each client layer on the global layer is proportional to its associated weight. For the definition of SUPPORT_LAYER, please refer to heflwr.nn documentation.

Parameters

• global_layer (heflwr.nn.SUPPORT_LAYER) - The global parameter layer to aggregate as the target.
• subset_layers (List[heflwr.nn.SUPPORT_LAYER]) - List of multiple local parameter layers.
• weights (List[int]) - Weights for each client layers, typically set to a list consisting of the number of training samples for each federated client.

Returns and Side Effects

• Returns - None.
• Side Effects - Parameters of the global_layer object are reset to the weighted average of subset_layers.

extract

extract(parameters: flwr.common.typing.Parameters,
        client_net: torch.nn.Module,
        server_net: torch.nn.Module,
) -> flwr.common.typing.Parameters

Extract complete parameters parameters matching server_net according to the pruning structure of client_net, used for the scenario where the server distributes model parameters to clients in Flower federated learning. Returns the extracted local parameters, resets the parameters of server_net to parameters, and resets the parameters of client_net to the returned values.

Parameters

• parameters (flwr.common.typing.Parameters) - Parameters object of the current global model.
• client_net (torch.nn.Module) - Client model object constructed with SUPPORT_LAYER or a model object consistent with the client model structure.
• server_net (torch.nn.Module) - Global model object constructed with SUPPORT_LAYER or a model object consistent with the global model structure.

Returns and Side Effects

• Returns - Parameters extracted from parameters that match the structure of client_net.
• Side Effects
  • Resets the parameters of server_net to parameters.
  • Resets the parameters of client_net to the returned values.

merge

merge(results: List[Tuple[flwr.server.client_proxy.ClientProxy,
                          flwr.common.typing.FitRes]],
      client_nets: List[torch.nn.Module],
      server_net: torch.nn.Module,
) -> flwr.common.typing.Parameters

The parameters from models of multiple clients (contained in results) are aggregated using heterogeneous weighted aggregation. used for the scenario where the server aggregates local training results from various clients in Flower federated learning. Returns the parameters after heterogeneous weighted aggregation, resets the parameters of each object in client_nets to the corresponding parameters in results, and resets the parameters of server_net to the returned values.

Parameters

• results (List[Tuple[flwr.server.client_proxy.ClientProxy, flwr.common.typing.FitRes]]) - Local training results of various clients.
• client_nets (List[torch.nn.Module]) - Multiple client model objects constructed with SUPPORT_LAYER or model objects consistent with the client model structure.
• server_net (torch.nn.Module) - Global model object constructed with SUPPORT_LAYER or a model object consistent with the global model structure.

Returns and Side Effects

• Returns - Parameters after heterogeneous weighted aggregation.
• Side Effects
  • Resets the parameters of server_net to the returned values.
  • Resets the parameters of each object in client_nets to the corresponding parameters in results.