replace torch.div() with torch.floor_divide(); add as_tuple=False to...

        # give padding probability of 0 everywhere

        smooth_dist[:, self.pad_index] = 0

        # masking out padding area (sum of probabilities for padding area = 0)

        padding_positions = torch.nonzero(targets.data == self.pad_index)

        padding_positions = torch.nonzero(targets.data == self.pad_index, as_tuple=False)

        # pylint: disable=len-as-condition

        if len(padding_positions) > 0:

            smooth_dist.index_fill_(0, padding_positions.squeeze(), 0.0)

                b = batch_offset[i]

                if end_condition[i]:

                    is_finished[i].fill_(1)

                finished_hyp = is_finished[i].nonzero().view(-1)

                finished_hyp = is_finished[i].nonzero(as_tuple=False).view(-1)

                # store finished hypotheses for this batch

                for j in finished_hyp:

                    # Check if the prediction has more than one EOS.

                    # If it has more than one EOS, it means that the

                    # prediction should have already been added to

                    # the hypotheses, so you don't have to add them again.

                    if (predictions[i, j, 1:] == eos_index).nonzero().numel() \

                    if (predictions[i, j, 1:] == eos_index).nonzero(as_tuple=False).numel() \

                            < 2:

                        # ignore start_token

                        hypotheses[b].append(

                            break

                        results["scores"][b].append(score)

                        results["predictions"][b].append(pred)

            non_finished = end_condition.eq(False).nonzero().view(-1)

            non_finished = end_condition.eq(False).nonzero(as_tuple=False).view(-1)

            # if all sentences are translated, no need to go further

            # pylint: disable=len-as-condition

            if len(non_finished) == 0: