How to override Sklearn module function

Question

I'm using sklearn.metrics.cohen_kappa_score to evaluate my module. The function weights can be None , 'linear' or 'quadratic' I would like to override the function in order to be able to send custom weights matrix. how can it be done?

def cohen_kappa_score(y1, y2, *, labels=None, weights=None,
                      sample_weight=None):
    confusion = confusion_matrix(y1, y2, labels=labels,
                                 sample_weight=sample_weight)
    n_classes = confusion.shape[0]
    sum0 = np.sum(confusion, axis=0)
    sum1 = np.sum(confusion, axis=1)
    expected = np.outer(sum0, sum1) / np.sum(sum0)

    if type(w_mat) != np.ndarray: # <------------------------- line I want to add
        if weights is None:
            w_mat = np.ones([n_classes, n_classes], dtype=int)
            w_mat.flat[:: n_classes + 1] = 0
        elif weights == "linear" or weights == "quadratic":
            w_mat = np.zeros([n_classes, n_classes], dtype=int)
            w_mat += np.arange(n_classes)
            if weights == "linear":
                w_mat = np.abs(w_mat - w_mat.T)
            else:
                w_mat = (w_mat - w_mat.T) ** 2   ​
       ​else:
           ​raise ValueError("Unknown kappa weighting type.")
   ​
​k = np.sum(w_mat * confusion) / np.sum(w_mat * expected)
​return 1 - k

Answer 1

You can either go with make_scorer as per showed by @Antoine in the other answer, or you can override the function itself:

import numpy as np
import sklearn.metrics as sm
from sklearn.metrics import confusion_matrix


def custom_cohen_kappa_score(y1, y2, *, labels=None, weights=None, sample_weight=None):
    print("This is the custom function")
    confusion = confusion_matrix(y1, y2, labels=labels,
                                 sample_weight=sample_weight)
    n_classes = confusion.shape[0]
    sum0 = np.sum(confusion, axis=0)
    sum1 = np.sum(confusion, axis=1)
    expected = np.outer(sum0, sum1) / np.sum(sum0)

    if weights is None:
        w_mat = np.ones([n_classes, n_classes], dtype=int)
        w_mat.flat[:: n_classes + 1] = 0
    elif weights == "linear" or weights == "quadratic":
        w_mat = np.zeros([n_classes, n_classes], dtype=int)
        w_mat += np.arange(n_classes)
        if weights == "linear":
            w_mat = np.abs(w_mat - w_mat.T)
        else:
            w_mat = (w_mat - w_mat.T) ** 2
    else:
        raise ValueError("Unknown kappa weighting type.")

    k = np.sum(w_mat * confusion) / np.sum(w_mat * expected)
    return 1 - k


# override it
sm.cohen_kappa_score = custom_cohen_kappa_score

# Test: Here every time `cohen_kappa_score` is called, 
# the custom one will be invoked instead!

from sklearn.metrics import cohen_kappa_score

y_true = [2, 0, 2, 2, 0, 1]
y_pred = [0, 0, 2, 2, 0, 2]

print(cohen_kappa_score(y_true, y_pred))

Output

This is the custom function
0.4285714285714286

Answer 2

The best option would be to encapsulate your own scoring function using sklearn.metrics.make_scorer in order to use it for GridSearchCV and cross_val_score.

as follow:

from sklearn.metrics import make_scorer

weighted_cohen_kappa_score = make_scorer(custom_cohen_kappa,
                                         greater_is_better=True,
                                         needs_proba=False,
                                         needs_threshold=False
)

Where custom_cohen_kappa is you custom scoring function defined in your question.