@qinhanmin2014, here is my preliminary suggestion. The implementation for GeneralNB is slated for the 5 naive Bayes algorithms that sklearn has GaussianNB, BernoulliNB, MultinomialNB, ComplementNB and CategoricalNB.

An sklearn user would expect to use GeneralNB in this manner:

import numpy as np
from sklearn.naive_bayes import GaussianNB, BernoulliNB, 

X = np.array([[1.5, 2.3, 0, 1],
              [2.7, 3.8, 1, 0],
              [1.7, 0.1, 1, 0]])
y = np.array([1,0,2])

generalnb = GeneralNB([(GaussianNB(), [0,1]), (BernoulliNB(), [2,3])])
generalnb.fit(X, y)

In general how I would implement the GeneralNB.fit() function is to run fit() on the respective naive Bayes objects.

For the GeneralNB.predict() function, we could sum the _joint_log_likelihood() for each object, and subtract (n-1) log P(c) from this sum.

Let me know what are your thoughts on this or if I missed something out.

Thanks, I think this is what we want.
Actually there's an annoying blocker #14444, but you can start to draft a PR. It's not likely to be included in the next release so please be patient.

Got it, I will start working on it.

take