scatter_and_gather_workflow.rst

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 :class:`nvflare.app_common.workflows.scatter_and_gather.ScatterAndGather` is a for loop:

Trainer

A :class:`Trainer<nvflare.app_common.executors.trainer.Trainer>` is a type of :class:`Executor<nvflare.apis.executor.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 :ref:`application`.

Learnable

:class:`Learnable<nvflare.app_common.abstract.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 :class:`LearnablePersistor<nvflare.app_common.abstract.learnable_persistor.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

:class:`Aggregators<nvflare.app_common.abstract.aggregator.Aggregator>` 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.

.. literalinclude:: ../../../nvflare/app_common/abstract/aggregator.py
    :language: python
    :lines: 22-

Example with Scatter and Gather Workflow

See the :github_nvflare_link:`Hello Numpy Scatter and Gather Example <examples/hello-world/hello-numpy-sag>` for an example application with the scatter and gather workflow.