In the RNN classification example, using characters of names to predict the names language, the train function re-zeros the hidden state (and gradient) every epoch. I was wondering why this is done, instead of carrying over the final hidden states of the epoch before?