loss computation in splitcross.py is false

[adjidieng]

Hi,

Your code for adaptive softmax in splitcross.py is false.

(1) For the tail, you are taking log_softmax twice and you are accounting for p(tombstone) twice. You are getting the entropy both in self.log_prob() and inside the forward function.

Thanks.