An implementation of the sigf toolkit for randomization tests in Python 3

sigf.py

#!/usr/bin/env python3

__author__ = 'Dmitry Ustalov'
__credits__ = 'Sebastian Padó'
__license__ = 'MIT'

# This is an MIT-licensed implementation of the sigf toolkit for randomization tests:
# https://nlpado.de/~sebastian/software/sigf.shtml

import random
import sys
from statistics import mean


def input_counts(f):
    return [int(line.strip()) for line in f]


def input_tp_fp_fn(f):
    result = []

    for line in f:
        line = line.strip()
        if line: result.append(tuple(int(count) for count in line.split(' ', 2)))

    return result


def f1_score(model):
    tp = sum(obs[0] for obs in model)
    tp_fp = sum(obs[1] for obs in model)
    tp_fn = sum(obs[2] for obs in model)

    if tp == 0 or tp_fp == 0 or tp_fn == 0: return 0.

    precision, recall = tp / float(tp_fp), tp / float(tp_fn)
    return 2 * precision * recall / (precision + recall)


def randomized_test(model1, model2, score, trials, getrandbits_func):
    print('# score(model1) = %f' % score(model1), file=sys.stderr)
    print('# score(model2) = %f' % score(model2), file=sys.stderr)

    diff = abs(score(model1) - score(model2))
    print('# abs(diff) = %f' % diff, file=sys.stderr)

    uncommon = [i for i in range(len(model1)) if model1[i] != model2[i]]

    better = 0

    for _ in range(trials):
        model1_local, model2_local = list(model1), list(model2)

        for i in uncommon:
            if getrandbits_func(1) == 1:
                model1_local[i], model2_local[i] = model2[i], model1[i]

        assert len(model1_local) == len(model2_local) == len(model1) == len(model2)

        diff_local = abs(score(model1_local) - score(model2_local))

        if diff_local >= diff:
            better += 1

    p = (better + 1) / (trials + 1)
    return p


# Every element of SCORES is a pair of input-reading function and scoring function.
SCORES = {
    'mean': (input_counts, mean),
    'f1': (input_tp_fp_fn, f1_score)
}


def main():
    import argparse

    parser = argparse.ArgumentParser()
    parser.add_argument('--seed', '-s', type=int)
    parser.add_argument('--score', choices=SCORES.keys(), default='mean')
    parser.add_argument('--trials', '-n', type=int, default=10 ** 5)
    parser.add_argument('model1', type=argparse.FileType('r'))
    parser.add_argument('model2', type=argparse.FileType('r'))
    args = parser.parse_args()

    if args.seed is None:
        getrandbits_func = random.getrandbits
    else:
        rng = random.Random(args.seed)
        getrandbits_func = rng.getrandbits

    reader, score = SCORES[args.score]

    model1, model2 = reader(args.model1), reader(args.model2)
    assert len(model1) == len(model2)

    p = randomized_test(model1, model2, score, args.trials, getrandbits_func)
    print('p-value = %f' % p)


if '__main__' == __name__:
    main()

Input files have the following format: tp tp+fp tp+fn.

$ cat exampleFScore/model1
1 2 1
2 2 2
1 2 2

$ cat exampleFScore/model2
1 2 1
0 2 2
0 2 2

$ ./sigf.py --score=f1 -n 100000 exampleFScore/model{1,2}
# score(model1) = 0.727273
# score(model2) = 0.181818
# abs(diff) = 0.545455
p-value = 0.498935