nvflare.app_common.workflows.scaffold module¶
- class Scaffold(*args, min_clients: int = 1000, num_rounds: int = 5, start_round: int = 0, persist_every_n_rounds: int = 1, **kwargs)[source]¶
Bases:
BaseFedAvg
Controller for Scaffold Workflow. Note: This class is based on WFController. Implements [SCAFFOLD](https://proceedings.mlr.press/v119/karimireddy20a.html).
- Provides the implementations for the run routine, controlling the main workflow:
def run(self)
The parent classes provide the default implementations for other routines.
- Parameters:
min_clients (int, optional) – The minimum number of clients responses before Workflow starts to wait for wait_time_after_min_received. Note that the workflow will move forward when all available clients have responded regardless of this value. Defaults to 1000.
num_rounds (int, optional) – The total number of training rounds. Defaults to 5.
persistor_id (str, optional) – ID of the persistor component. Defaults to “persistor”.
ignore_result_error (bool, optional) – whether this controller can proceed if client result has errors. Defaults to False.
allow_empty_global_weights (bool, optional) – whether to allow empty global weights. Some pipelines can have empty global weights at first round, such that clients start training from scratch without any global info. Defaults to False.
task_check_period (float, optional) – interval for checking status of tasks. Defaults to 0.5.
persist_every_n_rounds (int, optional) – persist the global model every n rounds. Defaults to 1. If n is 0 then no persist.
The base controller for FedAvg Workflow. Note: This class is based on the WFController.
Implements [FederatedAveraging](https://arxiv.org/abs/1602.05629).
A model persistor can be configured via the persistor_id argument of the WFController. The model persistor is used to load the initial global model which is sent to a list of clients. Each client sends it’s updated weights after local training which is aggregated. Next, the global model is updated. The model_persistor will also save the model after training.
- Provides the default implementations for the follow routines:
def sample_clients(self, min_clients)
def aggregate(self, results: List[FLModel], aggregate_fn=None) -> FLModel
def update_model(self, aggr_result)
The run routine needs to be implemented by the derived class:
def run(self)
- Parameters:
min_clients (int, optional) – The minimum number of clients responses before Workflow starts to wait for wait_time_after_min_received. Note that the workflow will move forward when all available clients have responded regardless of this value. Defaults to 1000.
num_rounds (int, optional) – The total number of training rounds. Defaults to 5.
start_round (int, optional) – The starting round number.
persist_every_n_rounds (int, optional) – persist the global model every n rounds. Defaults to 1. If n is 0 then no persist.