multi_weighted_logloss.py

gistfile1.txt

def multi_weighted_logloss(y_ohe, y_p):
    """
    @author olivier https://www.kaggle.com/ogrellier
    multi logloss for PLAsTiCC challenge
    """
    classes = [6, 15, 16, 42, 52, 53, 62, 64, 65, 67, 88, 90, 92, 95]
    class_weight = {6: 1, 15: 2, 16: 1, 42: 1, 52: 1, 53: 1, 62: 1, 64: 2, 65: 1, 67: 1, 88: 1, 90: 1, 92: 1, 95: 1}
    # Normalize rows and limit y_preds to 1e-15, 1-1e-15
    y_p = np.clip(a=y_p, a_min=1e-15, a_max=1-1e-15)
    # Transform to log
    y_p_log = np.log(y_p)
    # Get the log for ones, .values is used to drop the index of DataFrames
    # Exclude class 99 for now, since there is no class99 in the training set 
    # we gave a special process for that class
    y_log_ones = np.sum(y_ohe * y_p_log, axis=0)
    # Get the number of positives for each class
    nb_pos = y_ohe.sum(axis=0).astype(float)
    # Weight average and divide by the number of positives
    class_arr = np.array([class_weight[k] for k in sorted(class_weight.keys())])
    y_w = y_log_ones * class_arr / nb_pos    
    loss = - np.sum(y_w) / np.sum(class_arr)
    return loss