python numpy implementation of https://www.kaggle.com/wiki/MultiClassLogLoss

multi-class-log-loss-kaggle.py

import numpy as np

def logloss(y_true, prob_pred):
    '''
    https://www.kaggle.com/wiki/MultiClassLogLoss
    '''

    y_true = np.asarray(y_true, dtype=np.uint8)
    prob_pred = np.array(prob_pred, dtype=np.float)
    n_data = len(y_true)

    # rescale pred
    row_sum = np.sum(prob_pred, axis=1, keepdims=True)
    np.clip(row_sum, 1e-15, None, out=row_sum)
    np.divide(prob_pred, row_sum, out=prob_pred)

    # clip extremes
    np.clip(prob_pred, 1e-15, 1-1e-15, out=prob_pred)

    score = -(np.sum(np.log(prob_pred) * y_true) / n_data)
    return score

def test_logloss():
    y_true = [
        [1, 0],
        [1, 0],
        [1, 0],
        [0, 1],
        [0, 1],
        [0, 1],
    ]
    prob_pred = [
        [0.5, 0.5],
        [0.1, 0.9],
        [0.01, 0.99],
        [0.9, 0.1],
        [0.75, 0.25],
        [0.001, 0.999],
    ]
    assert logloss(y_true, prob_pred) == 1.881797068998267