Maybe we can use scipy.special.expit instead, seems that it's much faster than our implementation.