This rule raises an issue when an attribute ending with _ is set in the __init__ method of a class inheriting from Scikit-Learn BaseEstimator

Why is this an issue?

On a Scikit-Learn estimator, attributes that have a trailing underscore represent attributes that are estimated. These attributes have to be set in the fit method. Their presence is used to verify if an estimator has been fitted.

from sklearn.neighbors import KNeighborsClassifier

X = [[0], [1], [2], [3]]
y = [0, 0, 1, 1]
knn = KNeighborsClassifier(n_neighbors=1)
knn.fit(X, y)
knn.n_samples_fit_

In the example above the attributes of the KNeighborsClassifier, n_samples_fit_ is set only after the estimator’s fit method is called. Calling n_samples_fit_ before the estimator is fitted would raise an AttributeError exception.

When implementing a custom estimator by subclassing Scikit-Learn’s BaseEstimator, it is important to follow the above convention and not set attributes with a trailing underscore inside the __init__ method.

How to fix it

To fix this issue, move the attributes with a trailing underscore from the __init__ method to the fit method.

Code examples

Noncompliant code example

from sklearn.base import BaseEstimator

class MyEstimator(BaseEstimator):
    def __init__(self):
        self.estimated_attribute_ = None # Noncompliant: an estimated attribute is set in the __init__ method.

Compliant solution

from sklearn.base import BaseEstimator

class MyEstimator(BaseEstimator):
    def fit(self, X, y):
        self.estimated_attribute_ = some_estimation(X) # Compliant

Resources

Documentation