INNER CODE UNIT · Python

KendallVariants

linzhiqiu/t2v_metrics · dataset.py:69

def KendallVariants(
    gold_scores,
    metric_scores,
    variant: str = 'acc23',
    epsilon: float = 0.0,
) -> Tuple[float, float]:
    """Lightweight, optionally factored versions of variants on Kendall's Tau.

    This function calculates the sufficient statistics for tau in two different
    ways, either using a Fenwick Tree (`_FenwickTreeSufficientStatistics`) when
    `epsilon` is 0 or NumPy matrices (`_MatrixSufficientStatistics`) otherwise.
    Note that the latter implementation has an O(n^2) space requirement, which
    can be significant for long vectors.

    This implementation makes several changes to the SciPy implementation of
    Kendall's tau:
    1) For the Fenwick tree version, the cython function for computing discordant
        pairs is replaced by inline python. This works up to 2x faster for small

View source record →