remove sklearn/numpy deps

[davidheineman]

Remove sklearn which was being used to compute F1 score. Adds a custom implementation.

This function shows equivalence between the two implementations:

from sklearn.metrics import f1_score as sklearn_f1

def custom_f1_score(y_true, y_pred, pos_label=1):
    y_true = list(y_true)
    y_pred = list(y_pred)
    tp = sum((yt == pos_label) and (yp == pos_label) for yt, yp in zip(y_true, y_pred))
    fp = sum((yt != pos_label) and (yp == pos_label) for yt, yp in zip(y_true, y_pred))
    fn = sum((yt == pos_label) and (yp != pos_label) for yt, yp in zip(y_true, y_pred))

    if tp + fp == 0 or tp + fn == 0:
        return 0.0

    precision = tp / (tp + fp)
    recall = tp / (tp + fn)

    if precision + recall == 0:
        return 0.0

    return 2 * precision * recall / (precision + recall)

# Test
labels = [0, 1, 0, 1, 0, 0, 1]
preds = [0, 1, 1, 1, 0, 0, 0]
preds_no_leading_space = [0, 1, 0, 1, 1, 0, 0]

score_sklearn = sklearn_f1(labels, preds, pos_label=0)
score_custom = custom_f1_score(labels, preds, pos_label=0)

score_sklearn_no_leading = sklearn_f1(labels, preds_no_leading_space, pos_label=0)
score_custom_no_leading = custom_f1_score(labels, preds_no_leading_space, pos_label=0)

def isclose(a, b, rel_tol=1e-9, abs_tol=0.0):
    return abs(a - b) <= max(rel_tol * max(abs(a), abs(b)), abs_tol)

assert isclose(score_custom, score_sklearn)
assert isclose(score_custom_no_leading, score_sklearn_no_leading)

[epwalsh]

Just one note about the CHANGELOG, otherwise LGTM