bug report: empy batch when using --max-tokens < 128

[robertodessi]

I noticed that when setting max tokens < 128 in the dummy step before starting the actual training an error occurs.

| training on 1 GPUs
| max tokens per GPU = 25 and max sentences per GPU = None
Traceback (most recent call last):
  File "train.py", line 358, in <module>
    main(args)
  File "train.py", line 78, in main
    trainer.dummy_train_step([dummy_batch])
  File "/home/roberto.dessi/new_fairseq/fairseq/trainer.py", line 326, in dummy_train_step
    self.train_step(dummy_batch, dummy_batch=True)
  File "/home/roberto.dessi/new_fairseq/fairseq/trainer.py", line 176, in train_step
    ignore_grad
  File "/home/roberto.dessi/new_fairseq/fairseq/tasks/fairseq_task.py", line 169, in train_step
    loss, sample_size, logging_output = criterion(model, sample)
  File "/home/roberto.dessi/.virtualenvs/work/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/roberto.dessi/new_fairseq/fairseq/criterions/cross_entropy.py", line 30, in forward
    net_output = model(**sample['net_input'])
TypeError: 'NoneType' object is not subscriptable

The error is caused by https://github.com/robertodessi/fairseq/blob/master/fairseq/data/language_pair_dataset.py#L195

It seems to me that a small value of max_tokens causes that division to be 0 and to create an empty dummy batch that raise the above error.

It works when I replace that line with:

bsz = max(num_tokens // max(src_len, tgt_len), 1)