Server

The server is the orchestrator of the training, aggregating the results from the clients to improve the global model. As the clients, it is independent of the machine learning tools. The server can evaluate the global model by proxying an evaluate() function.

App

Entry point to start the CentralNode. The server must be started before launching clients.

Fed.Server.build_router — Function

build_router(central_node::CentralNode)

Build the routes to the central node endpoints.

source

Fed.Server.start — Function

start(central_node::CentralNode)

Start a Fed.jl HTTP server.

source

CentralNode

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

Fed.Server.CentralNode — Type

CentralNode(host, port, weights, strategy, evaluate, config)

Orchestrate the entire federated learning process. It handles the clients registration and requests them to train on their local data.

source

Fed.Server.fit — Function

fit(central_node::CentralNode)

Fit the global model by orchestrating the federated training.

source

Fed.Server.fit_client — Function

fit_client(url::String, payload::Vector{UInt8})

Ask one client to train on its local data, by using the model weights contained in the payload. Return the serialized updated weights.

source

Fed.Server.fit_clients — Function

fit_clients(endpoint::String, clients::Vector{String}, payload::Vector{UInt8})

Ask asynchronously each client from clients to train on their local data by using the model weights contained in the payload. Return a Vector where each element is the serialized weights of one client.

source

Fed.Server.register_client! — Function

register_client!(cm::ClientManager, req::HTTP.Request)

Register a new client in the ClientManager. The body of the request should contain the client URL: http://hostname:port.

Info

API Endpoint: this is an HTTP server endpoint.

source

ClientManager

The ClientManager handles the nodes registrations and synchronize the server to make it start when the clients are ready.

Fed.Server.ClientManager — Type

ClientManager()

Structure which manages the clients registration and syncrhonizes the server to make it work with them.

source

Base.length — Method

length(cm::ClientManager)

Return the number of client managed by the ClientManager.

source

Fed.Server.num_available_clients — Function

num_available_clients(cm::ClientManager)

Return the number of available clients.

source

Fed.Server.register! — Function

register!(cm::ClientManager, client_url::String)

Example

register!(cm, "http://127.0.0.1:8081")

source

Fed.Server.sample_clients — Function

sample_clients(cm::ClientManager, fraction::Float32)

Sample randomly fraction% of the clients, without replacement. Return at least one client if the number of available clients > 0.

source

Fed.Server.wait_for — Function

wait_for(cm::ClientManager, num_clients::Int)

Block until at least num_clients are available.

source

Strategy

Fed.Server.federated_averaging — Function

federated_averaging(round_weights)

Implement the federated averaging algorithm descibed in:

Communication-Efficient Learning of Deep Networks from Decentralized Data

round_weights must be an array containing the weights from each clients participating in the round.

source