final_checkpoints

Simulate a checkpoint schedule and return the set of epochs for which a checkpoint remains.

>>> final_checkpoints(50)
[10, 20, 30, 40, 50]
>>> final_checkpoints(50, schedule="explicit", schedule_kwargs=dict(steps=[30, 35]))
[30, 35]
>>> final_checkpoints(
...     50,
...     schedule="union",
...     schedule_kwargs=dict(
...         bases=["every", "explicit"],
...         bases_kwargs=[dict(frequency=15), dict(steps=[7,])],
...     ),
... )
[7, 15, 30, 45]
>>> final_checkpoints(50, keeper="last", keeper_kwargs=dict(keep=2))
[40, 50]
>>> final_checkpoints(50, keeper="modulo", keeper_kwargs=dict(modulo=20))
[20, 40]
>>> final_checkpoints(50, keeper="explicit", keeper_kwargs=dict(keep=[15]))
[]
>>> final_checkpoints(
...     50,
...     keeper="union",
...     keeper_kwargs=dict(
...         bases=["last", "modulo"],
...         bases_kwargs=[None, dict(divisor=20)],
...     ),
... )
[20, 40, 50]

Warning

You cannot easily inspect schedules which depend on training dynamics, e.g., BestCheckpointSchedule.

Parameters:

num_epochs (int) – the number of epochs
schedule (str | CheckpointSchedule | type[CheckpointSchedule] | None) – a checkpoint schedule instance or selection, cf. pykeen.checkpoints.scheduler_resolver
schedule_kwargs (Mapping[str, Any] | None) – additional keyword-based parameters when the schedule needs to instantiated first from a selection, cf. pykeen.checkpoints.scheduler_resolver
keeper (str | CheckpointKeeper | type[CheckpointKeeper] | None) – a checkpoint retention policy instance or selection, cf. pykeen.checkpoints.keeper_resolver None corresponds to keeping everything that was checkpointed.
keeper_kwargs (Mapping[str, Any] | None) – additional keyword-based parameters when the retention policy needs to instantiated first from a selection, cf. pykeen.checkpoints.keeper_resolver

Returns:

a sorted list of epochs at which a checkpoint remains after clean-up.

Return type:

list[int]