def _compare_against_f1_score():
y_true = np.random.randint(0, 2, (64, ))
y_pred = np.random.uniform(0, 1, (64, ))
pred_thresholds = [.01, .05, .1, .2, .4, .5, .6, .8, .95, .99]
single_scores = [
f1_score(y_true, y_pred, th) for th in pred_thresholds
]
parallel_scores = f1_score_multi_th(y_true, y_pred, pred_thresholds)
assert np.allclose(single_scores, parallel_scores, atol=1e-15)
def _test_vs_sklearn():
y_true = np.array([1] * 5 + [0] * 27) # imbalanced
np.random.shuffle(y_true)
y_pred = np.random.uniform(0, 1, 32)
test_score = f1_score(y_true, y_pred, pred_threshold=.5)
sklearn_score = sklearn.metrics.f1_score(y_true, y_pred > .5)
adiff = abs(test_score - sklearn_score)
assert (adiff < 1e-10), ("sklearn: {:.15f}\ntest: {:.15f}"
"\nabsdiff: {}".format(
sklearn_score, test_score, adiff))
def _test_no_positive_predictions():
y_true = [0, 0, 1]
y_pred = [0, 0, 0]
assert f1_score(y_true, y_pred) == 0
def _test_no_positive_labels():
y_true = [0] * 6
y_pred = [.1, .2, .3, .65, .7, .8]
assert f1_score(y_true, y_pred) == 0.5
def _test_basic():
y_true = [0, 0, 1, 0, 0, 0, 1, 1]
y_pred = [.01, .93, .42, .61, .15, 0, 1, .5]
assert abs(f1_score(y_true, y_pred) - 1 / 3) < 1e-15
def f05_score(y_true, y_pred, pred_threshold=.5):
return metrics.f1_score(y_true, y_pred, pred_threshold, beta=.5)