import numpy as np
from sklearn.metrics import f1_score as skf1_score

def f1_score(y_true, y_pred, average):
    n_labels = len(set(y_true) | set(y_pred))
    true_sum = np.bincount(y_true, minlength=n_labels)
    pred_sum = np.bincount(y_pred, minlength=n_labels)
    tp = np.bincount(y_true[y_true == y_pred], minlength=n_labels)
    if average == "binary":
        tp = np.array([tp[1]])
        true_sum = np.array([true_sum[1]])
        pred_sum = np.array([pred_sum[1]])
    elif average == "micro":
        tp = np.array([np.sum(tp)])
        true_sum = np.array([np.sum(true_sum)])
        pred_sum = np.array([np.sum(pred_sum)])
    precision = np.zeros(len(pred_sum))
    mask = pred_sum != 0
    precision[mask] = tp[mask] / pred_sum[mask]
    recall = np.zeros(len(true_sum))
    mask = true_sum != 0
    recall[mask] = tp[mask] / true_sum[mask]
    denom = precision + recall
    denom[denom == 0.] = 1
    fscore = 2 * precision * recall / denom
    if average == "weighted":
        fscore = np.average(fscore, weights=true_sum)
    elif average is not None:
        fscore = np.mean(fscore)
    return fscore

# binary
for i in range(10):
    rng = np.random.RandomState(i)
    y_true = rng.randint(2, size=10)
    y_pred = rng.randint(2, size=10)
    score1 = f1_score(y_true, y_pred, average="binary")
    score2 = skf1_score(y_true, y_pred, average="binary")
    assert np.isclose(score1, score2)