子类化sklearn LinearSVC适配GridSearchCV时遇数值错误排查

Let's cut to the chase: your ValueError about NaN/inf values isn't actually a problem with your data—it's a missing return statement in the methods you overrode in your LinearSVCSub class.

Here's the core issue: sklearn's estimator API requires methods like fit, predict, score, and decision_function to return specific values (e.g., fit returns the estimator instance itself, predict returns the predicted labels). In your code, you call the parent class's method but don't pass its result back to the caller. For example:

def predict(self, X):
    X = self.transform_this(X)
    super(LinearSVCSub, self).predict(X)  # No return here!

This means the method returns None by default. When GridSearchCV tries to use these methods during cross-validation, it gets None instead of valid predictions/scores, which causes downstream errors that manifest as the NaN/inf message you're seeing.

Corrected Code #

Here's the fixed version of your subclass, with the missing return statements added:

from sklearn.datasets import load_breast_cancer
from sklearn.svm import LinearSVC
from sklearn.model_selection import GridSearchCV

RANDOM_STATE = 123

class LinearSVCSub(LinearSVC):
    def __init__(self, penalty='l2', loss='squared_hinge', additional_parameter1=1, additional_parameter2=100, dual=True, tol=0.0001, C=1.0, multi_class='ovr', fit_intercept=True, intercept_scaling=1, class_weight=None, verbose=0, random_state=None, max_iter=1000):
        super(LinearSVCSub, self).__init__(penalty=penalty, loss=loss, dual=dual, tol=tol, C=C, multi_class=multi_class, fit_intercept=fit_intercept, intercept_scaling=intercept_scaling, class_weight=class_weight, verbose=verbose, random_state=random_state, max_iter=max_iter)
        self.additional_parameter1 = additional_parameter1
        self.additional_parameter2 = additional_parameter2

    def fit(self, X, y, sample_weight=None):
        X = self.transform_this(X)
        return super(LinearSVCSub, self).fit(X, y, sample_weight)  # Added return

    def predict(self, X):
        X = self.transform_this(X)
        return super(LinearSVCSub, self).predict(X)  # Added return

    def score(self, X, y, sample_weight=None):
        X = self.transform_this(X)
        return super(LinearSVCSub, self).score(X, y, sample_weight)  # Added return

    def decision_function(self, X):
        X = self.transform_this(X)
        return super(LinearSVCSub, self).decision_function(X)  # Added return

    def transform_this(self, X):
        return X

if __name__ == '__main__':
    data = load_breast_cancer()
    X, y = data.data, data.target
    # Parameter tuning with custom LinearSVC
    param_grid = {'C': [0.00001, 0.0001, 0.0005], 'dual': (True, False), 'random_state': [RANDOM_STATE], 'additional_parameter1': [0.90, 0.80, 0.60, 0.30], 'additional_parameter2': [20, 30]}
    gs_model = GridSearchCV(estimator=LinearSVCSub(), verbose=1, param_grid=param_grid, scoring='roc_auc', n_jobs=-1)
    gs_model.fit(X, y)

Key Notes #

• Always stick to sklearn's estimator API when subclassing: every method that returns something in the parent class must return the same type of value in your subclass. This ensures compatibility with tools like GridSearchCV and Pipeline.
• Your transform_this method is currently a no-op, but when you expand its functionality later, just make sure it returns a valid feature matrix (no NaN, infinity, or values outside the float64 range) and the rest of the pipeline will work as expected.

内容的提问来源于stack exchange，提问作者Hawklaz