flip.flower.strategy

FLIP’s Flower base strategy — FedAvg plus Central Hub telemetry.

FlipFedAvg is the durable home for everything a FLIP Flower server app reports to the hub: the RUNNING status, per-client metric and exception forwarding (flip.flower.metrics), and the typed round events (flip.flower.progress). App templates subclass it and keep only their app-specific behaviour; the telemetry evolves here, in flip-utils, without touching uploaded apps.

Round events attach to the phase the strategy actually runs: training rounds emit around configure_train/aggregate_train; an evaluation-only strategy (fraction_train=0.0) emits around the evaluate phase instead, so every job type yields a coherent round timeline without double-emission.

Unlike the rest of flip.flower, importing this module requires the flwr package (present in the Flower fl-server images) — it subclasses the real FedAvg. Keep logic in the helpers; this class is wiring.

Classes

FlipFedAvg

FedAvg with FLIP hub telemetry: status, metrics/exceptions, round events.

Module Contents

class flip.flower.strategy.FlipFedAvg(flip: flip.FLIP, model_id: str, *args, **kwargs)

Bases: flwr.serverapp.strategy.FedAvg

FedAvg with FLIP hub telemetry: status, metrics/exceptions, round events.

Parameters:

flip – FLIP instance used by the fl-server to reach the Central Hub.
model_id – FLIP model ID (UUID) for the current run.
*args – Passed through to FedAvg.
**kwargs – Passed through to FedAvg.

flip

model_id

num_rounds: int | None = None

_telemetry

start(grid: flwr.serverapp.Grid, initial_arrays: flwr.app.ArrayRecord, num_rounds: int = 3, **kwargs): Capture the round total and mark the run as executing on the hub.

configure_train(server_round: int, arrays: flwr.app.ArrayRecord, config: flwr.common.ConfigRecord, grid: flwr.serverapp.Grid) → collections.abc.Iterable[flwr.app.Message]: Dispatch the round’s train tasks, reporting the round start.

aggregate_train(server_round: int, replies: collections.abc.Iterable[flwr.app.Message]) → flwr.app.ArrayRecord | None: Forward per-client telemetry, aggregate, then report the round aggregated.

configure_evaluate(server_round: int, arrays: flwr.app.ArrayRecord, config: flwr.common.ConfigRecord, grid: flwr.serverapp.Grid) → collections.abc.Iterable[flwr.app.Message]: Dispatch the round’s evaluate tasks; owns the round events when nothing trains.

aggregate_evaluate(server_round: int, replies: collections.abc.Iterable[flwr.app.Message]) → flwr.app.MetricRecord | None: Forward per-client telemetry, aggregate, and close evaluation-only rounds.