Client

The client is a microservice that exposes the infrastructure required to communicate with the server. It is however independent of the machine learning tools. The training loop is defined externally by the user, and then is "proxied" by the client.

App

Entry point to start a client node. The setup is done automatically on startup and assumes the server is reachable.

Fed.Client.build_router — Function

build_router(node::Node)

Build the routes to the node endpoints.

source

Fed.Client.start — Function

start(node::Node)

Start the given node by setting up its router, and registering it to the server.

source

Node

Data structure containing the client's configuration and its API for communicating with the server.

Fed.Client.Node — Type

Node(host, port, fit, config)

Orchestrate the training on local data of the clients participating in the federated training. A Node is registerating itself to the server on startup, and then wait for server requests for training on its local data.

The fit function is called "proxied function". It should define the training loop and respect the following signature:

fit(weights::Vector{Float32})::Vector{Float32}

Where weights is a flatten 1D array contianing the model parameters.

source

Fed.Client.fit_service — Function

fit_service(node::Node, request::HTTP.Request)

Fit the received weights contained in the request with the node's local data, and return the updated weights to the caller. This function is calling the "proxied function" fit from the Node.

Info

API Endpoint: this is an HTTP client endpoint.

source

Fed.Client.register_to_server — Function

register_to_server(node::Node)

Info

API Client: this is an HTTP client request.

source