metrics.py

import math
import random


def pop_stddev(arr):
    mu = sum(arr) / len(arr)
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / len(arr))


def sample_stddev(arr):
    mu = sum(arr) / len(arr)
    return math.sqrt(sum([(x - mu) ** 2 for x in arr]) / (len(arr) - 1))


def mean_stderr(arr):
    return sample_stddev(arr) / math.sqrt(len(arr))


def acc_all_stderr(items):
    # Only count as correct if all answers are labeled correctly for each question
    question_scoring_dict = {}
    preds = list(zip(*items))[0]
    docs = list(zip(*items))[1]

    for doc, pred in zip(docs, preds):
        question_id = doc["idx"]["question"]
        if question_id not in question_scoring_dict:
            question_scoring_dict[question_id] = []

        gold_label = doc["label"] == 1
        question_scoring_dict[question_id].append(gold_label == pred)

    acc = mean_stderr([int(all(x)) for x in question_scoring_dict.values()])
    return acc


def metric_max_over_ground_truths(metric_fn, prediction, ground_truths):
    """Compute max metric between prediction and each ground truth."""
    scores_for_ground_truths = []
    for ground_truth in ground_truths:
        score = metric_fn(prediction, ground_truth)
        scores_for_ground_truths.append(score)
    return max(scores_for_ground_truths)


class _bootstrap_internal:
    def __init__(self, f, n):
        self.f = f
        self.n = n

    def __call__(self, v):
        i, xs = v
        rnd = random.Random()
        rnd.seed(i)
        res = []
        for _ in range(self.n):
            res.append(self.f(rnd.choices(xs, k=len(xs))))
        return res


def bootstrap_stderr(f, xs, iters):
    import multiprocessing as mp

    pool = mp.Pool(mp.cpu_count())
    # this gives a biased estimate of the stderr (i.e w/ the mean, it gives something
    # equivalent to stderr calculated without Bessel's correction in the stddev.
    # Unfortunately, I haven't been able to figure out what the right correction is
    # to make the bootstrap unbiased - i considered multiplying by sqrt(n/(n-1)) but
    # that would be ad-hoc and I can't prove that that would actually be an unbiased estimator)
    # Thankfully, shouldn't matter because our samples are pretty big usually anyways
    res = []
    chunk_size = min(1000, iters)
    from tqdm import tqdm

    print("bootstrapping for stddev:", f.__name__)
    for bootstrap in tqdm(
        pool.imap(
            _bootstrap_internal(f, chunk_size),
            [(i, xs) for i in range(iters // chunk_size)],
        ),
        total=iters // chunk_size,
    ):
        # sample w replacement
        res.extend(bootstrap)

    pool.close()
    return sample_stddev(res)


def yesno(x):
    if x:
        return "yes"
    else:
        return "no"