dataeval.protocols.EvaluationSchedule¶

class dataeval.protocols.EvaluationSchedule¶

Protocol for determining evaluation points in sufficiency analysis.

Implementations determine at which dataset sizes to train and evaluate the model during sufficiency analysis.

Examples

Custom scheduler evaluating at 0%, 50%, 100% of the dataset

>>> class MidpointSchedule:
...     def get_steps(self, dataset_length: int) -> np.typing.NDArray[np.intp]:
...         return np.array([0, dataset_length // 2, dataset_length - 1], dtype=np.intp)

get_steps(dataset_length)¶

Compute evaluation points for given dataset length.

Parameters:¶

dataset_length : int¶

Total length of training dataset

Returns:¶

Array of dataset sizes at which to evaluate, must be monotonically increasing and within [1, dataset_length]

Return type:¶

NDArray[np.intp]