BestCheckpointKeeper

class BestCheckpointKeeper(result_tracker: ResultTracker, metric_selection: MetricSelection)

Bases: CheckpointKeeper

Keep checkpoints for steps that achieved the best value for a metric.

Attributes Summary

metric_selection	the metric selection
result_tracker	the result tracker which receives updates on metrics since the same tracker instance needs to receive results from the training loop, we do require a pre-instantiated one rather than offering to provide hints, too

Methods Summary

__call__(steps)

Iterate over the steps for which checkpoints should be kept.

Attributes Documentation

Parameters:

• result_tracker (ResultTracker)

• metric_selection (MetricSelection)

metric_selection: MetricSelection = <dataclasses._MISSING_TYPE object>

the metric selection

result_tracker: ResultTracker = <dataclasses._MISSING_TYPE object>

the result tracker which receives updates on metrics since the same tracker instance needs to receive results from the training loop, we do require a pre-instantiated one rather than offering to provide hints, too

Methods Documentation

__call__(steps: Sequence[int]) → Iterator[int]

Iterate over the steps for which checkpoints should be kept.

Parameters:

steps (Sequence[int]) – the sorted list of steps at which checkpoints were written.

Yields:

the steps for which checkpoints should be kept

Return type:

Iterator[int]

# noqa:DAR302