A manager for facilitating multiple in-progress evaluations.

This manager performs three responsbilities:

  1. Prepares, starts and tracks new evaluation loops. This involves creating a new evaluation process and state manager for that process, adding the new process to the list of tracked inprocess evaluations, and creating a new asyncio.Task to run the evaluation loop.
  2. Record evaluations that have finished. This removes the evaluation from the list of in-progresss evaluations.
  3. If the program has restarted, load the most recent state of in-progress evaluations and restart each of the evaluations.

This class uses N + 1 tff.program.ProgramStateManagers to enable resumable evaluations.

  • The first state managers is for this class itself, and manages the list of in-progress evaluations via two tensor objects. Tensor objects must be used (rather than Python lists) because tff.program.FileProgramStateManager does not support state objects that change Python structure across versions (e.g. to load the next version, we must known its shape, but after a restart we don't know). Alternatively, we can use tensor or ndarray objects with shape [None] to support changing shapes of structure's leaf elements.
  • The next N state managers manage the cross-round metric aggregation for each evaluation process started. One for each evaluation process.

data_source A tff.program.FederatedDataSource that the manager will use to create iterators for evaluation loops.
aggregated_metrics_manager A tff.program.ReleaseManager for releasing the total aggregated metrics at the end of the evaluation loop.
create_state_manager_fn A callable that returns a tff.program.FileProgramStateManager that will be used to create the overall evaluation manager's state manager, and each per evaluation loop state manager that will enable resuming and checkpointing.
create_process_fn A callable that returns a 2-tuple of tff.learning.templates.LearningProcess and tff.program.ReleaseManager for the per-evaluation round metrics releasing that will used be to start each evaluation loop.
cohort_size An integer denoting the size of each evaluation round to select from the iterator created from data_source.
duration The datetime.timedelta duration to run each evaluation loop.

aggregated_metrics_manager A manager for releasing metrics at the end of each evaluation loop.
cohort_size The size of each evaluation round to select from the iterator.
create_process_fn A callable that returns a process and manager each evaluation loop.
create_state_manager_fn A callable that returns a program state manager each evaluation loop.
data_source A data source used to create iterators each evaluation loop.
duration The duration to run each evaluation loop.



View source

Removes evaluation for train_round from the internal state manager.

train_round The integer round number of the training round that has finished evaluation.

RuntimeError If train_round was not currently being tracked as an in-progress evaluation.


View source

Load the most recent state and restart in-progress evaluations.


View source

Starts a new evaluation loop for the incoming model_weights.


View source

Creates an awaitable that blocks until all evaluations are finished.