Reported at SO:

classifier = OneVsRestClassifier(SVC(class_weight='auto'))
classifier.fit(X1, y1)
y2 = classifier.predict(X2)

raises an exception because SVC.predict_proba doesn't actually work unless probability=True is passed.

Suggested fix: refactor the SVM code to introduce a new class, PlattSVC, that functions like SVC(probability=True) except that it uses probabilities for predict and decision_function also. Remove predict_proba from SVC. That would immediately fix the problem that SVC.predict_proba is inconsistent with decision_function and predict.