decent_bench.algorithms#

class decent_bench.algorithms.Algorithm[source]#

Bases: ABC, Generic

Base class for decentralized algorithms.

iterations: int#

Number of iterations to run the algorithm for.

abstract property name: str#

Name of the algorithm.

abstractmethod initialize(network: NetworkT) → None[source]#

Initialize the algorithm.

Parameters:

network – provides the agents and topology for this algorithm.

abstractmethod step(network: NetworkT, iteration: int) → None[source]#

Perform one iteration of the algorithm.

Parameters:

  • network – provides the agents and topology for this algorithm.

  • iteration – current iteration number.

abstractmethod cleanup_agents(network: NetworkT) → Iterable[Agent][source]#

Return the agents whose auxiliary variables should be cleared.

Parameters:

network – provides the agents and topology for this algorithm.

cleanup(network: NetworkT) → None[source]#

Clean up the algorithm state by clearing auxiliary variables from agents.

This method is used to free up memory used by auxiliary variables that are not needed after training. Can be overridden to control what gets cleaned up.

Note

Override cleanup_agents() to control which agents are cleaned up.

Parameters:

network – provides the agents and topology for this algorithm.

final run(network: NetworkT, start_iteration: int = 0, progress_callback: Callable[[int], None] | None = None) → None[source]#

Run the algorithm.

This method first calls initialize(), then step() for the specified number of iterations. Optionally call cleanup() after run() to clear auxiliary variables and free up memory.

Parameters:

  • network – provides the agents and topology for this algorithm.

  • start_iteration – iteration number to start from, used when resuming from a checkpoint. If greater than 0, initialize() will be skipped.

  • progress_callback – optional callback to report progress after each iteration.

Raises:

ValueError – if start_iteration is not in [0, iterations]

Warning

Do not override this method. Instead, override initialize() and step() as needed.

Note

The algorithm saves the agents’ states every state_snapshot_period.