Implementing Strategies

The strategy abstraction enables implementation of fully custom strategies. A strategy is basically the federated learning algorithm that runs on the server. Strategies decide how to sample clients, how to configure clients for traning, how to aggregate updates, and how to evaluate models. Flower provides a few built-in strategies which are based on the same API described below.

The Strategy abstraction

All strategy implementation are derived from the abstract base class flwr.server.strategy.Strategy, both built-in implementations and third party implementations. This means that custom strategy implementations have the exact same capabilities at their disposal as built-in ones.

The strategy abstraction defines a few abstract methods that need to be implemented:

class Strategy(ABC):

    @abstractmethod
    def configure_fit(
        self, rnd: int, weights: Weights, client_manager: ClientManager
    ) -> List[Tuple[ClientProxy, FitIns]]:

    @abstractmethod
    def aggregate_fit(
        self,
        rnd: int,
        results: List[Tuple[ClientProxy, FitRes]],
        failures: List[BaseException],
    ) -> Optional[Weights]:

    @abstractmethod
    def configure_evaluate(
        self, rnd: int, weights: Weights, client_manager: ClientManager
    ) -> List[Tuple[ClientProxy, EvaluateIns]]:

    @abstractmethod
    def aggregate_evaluate(
        self,
        rnd: int,
        results: List[Tuple[ClientProxy, EvaluateRes]],
        failures: List[BaseException],
    ) -> Optional[float]:

    @abstractmethod
    def evaluate(self, weights: Weights) -> Optional[Tuple[float, float]]:

Creating a new strategy means implmenting a new class derived from the abstract base class Strategy which provides implementations for the previously shown abstract methods:

class SotaStrategy(Strategy):

    def configure_fit(self, rnd, weights, client_manager):
        # Your implementation here

    def aggregate_fit(self, rnd, results, failures):
        # Your implementation here

    def configure_evaluate(self, rnd, weights, client_manager):
        # Your implementation here

    def aggregate_evaluate(self, rnd, results, failures):
        # Your implementation here

    def evaluate(self, weights):
        # Your implementation here

The following sections describe each of those methods in detail.

The configure_fit method

coming soon

The aggregate_fit method

coming soon

The configure_evaluate method

coming soon

The aggregate_evaluate method

coming soon

The evaluate method

coming soon

Deprecated methods

The following methods were replaced by updated versions with the same type signature. Migrate to the new versions by renaming them (i.e., remove the on_ prefix):

• on_configure_fit (replaced by configure_fit)

• on_aggregate_fit (replaced by aggregate_fit)

• on_configure_evaluate (replaced by configure_evaluate)

• on_aggregate_evaluate (replaced by aggregate_evaluate)