Actually if you are still having the problem you describe on this other thread, that would explain things. In other words the problem is that your train_naive_bayes function is that one that is broken and it needs to work in order to produce the logprior and loglikelihood values that the next function uses.