Scatter and Gather Workflow

The Federated scatter and gather workflow is an included reference implementation of the default workflow of previous versions of NVIDIA FLARE with a Server aggregating results from Clients that have produced Shareable results from their Trainer.

At the core, the control_flow of nvflare.app_common.workflows.scatter_and_gather.ScatterAndGather is a for loop:

Trainer

A Trainer is a type of Executor in NVIDIA FLARE.

The execute() method needs to get the required information from the Shareable, use that in its training process, then returning the local training result as a Shareable.

You will need to configure your own Trainer in config_fed_client.json. Example FL configurations can be found in NVIDIA FLARE Application.

Learnable

Learnable is the result of an FL application. For example, in the deep learning scenario, it can be the model weights. In the AutoML case, it can be the network architecture.

A LearnablePersistor defines how to load and save a Learnable. Learnable is a subset of the model file (which can contain other data like LR schedule) which is to be learned, like the model weights.

Aggregator

Aggregators define the aggregation algorithm to aggregate the Shareable. For example, a simple aggregator would be just average all the Shareable of the same round.

Below is the signature for an aggregator.

class Aggregator(FLComponent, ABC):
    @abstractmethod
    def accept(self, shareable: Shareable, fl_ctx: FLContext) -> bool:
        """Accept the shareable submitted by the client.

        Args:
            shareable: submitted Shareable object
            fl_ctx: FLContext

        Returns:
            first boolean to indicate if the contribution has been accepted.

        """
        pass

    @abstractmethod
    def aggregate(self, fl_ctx: FLContext) -> Shareable:
        """Perform the aggregation for all the received Shareable from the clients.

        Args:
            fl_ctx: FLContext

        Returns:
            shareable
        """
        pass