def test_f1_score(aggregator: StatsAggregator):
a_precision = 4 / 5
a_recall = 4 / 6
assert np.isclose(aggregator.f1_score("a"), 2 * a_precision * a_recall,
a_precision + a_recall)
b_precision = 1 / 3
b_recall = 1 / 2
assert np.isclose(aggregator.f1_score("b"), 2 * b_precision * b_recall,
b_precision + b_recall)
def test_empty_batch():
aggregator = StatsAggregator(["a", "b"])
assert aggregator.accuracy() == 0
assert aggregator.precision("a") == 0
assert aggregator.precision("b") == 0
assert aggregator.recall("a") == 0
assert aggregator.recall("b") == 0
assert aggregator.f1_score("a") == 0
assert aggregator.f1_score("b") == 0