nvflare.edge.assessors.buff_model_manager module

class BuffModelManager(num_updates_for_model: int, max_model_history: int | None = None, global_lr: float = 1.0, staleness_weight: bool = False)[source]

Bases: ModelManager

Initialize the ModelManager. The aggregation scheme and weights are calculated following FedBuff paper “Federated Learning with Buffered Asynchronous Aggregation”. The staleness_weight can be enabled to apply staleness weighting to model updates.

Special cases for max_model_history: - If None: Keep every model versions, only remove a version when all devices processing it reports back (version no longer related with any device_id in the current_selection from device_manager).

Parameters:

  • num_updates_for_model (int) – Number of updates required before generating a new model version.

  • max_model_history (int) – Maximum number of historical model versions to keep in memory. - None (default): keep every version until all devices processing a particular version report back. - positive integer: keep only the latest n versions

  • global_lr (float) – Global learning rate for model aggregation, default is 1.0.

  • staleness_weight (bool) – Whether to apply staleness weighting to model updates, default is False.

generate_new_model(fl_ctx: FLContext) → None[source]

Generate a new model version based on accumulated updates.

Parameters:

  • fl_ctx – FLContext object

  • Returns – none

initialize_model(model: DXO, fl_ctx: FLContext)[source]

Initialize the model manager with an initial model.

Parameters:

  • model – The initial model

  • fl_ctx – FLContext object

  • Returns – none

process_updates(model_updates: Dict[int, ModelUpdate], fl_ctx: FLContext) → bool[source]

Process incoming model updates from clients.

Parameters:

  • model_updates – updates collected from clients

  • fl_ctx – FLContext object

Returns:

Whether the updates were successfully processed

Return type:

bool

prune_model_versions(versions_to_keep: Set[int], fl_ctx: FLContext) → None[source]

Prune the model versions that are no longer active.

Parameters:

  • versions_to_keep – Set of model versions to keep

  • fl_ctx – FLContext object

  • Returns – none