Do forward and backward, and return the loss as computed by *criterion* for the given *model* and *sample*. Args: sample (dict): the mini-batch. The format is defined by the :class:`~fairseq.data.FairseqDataset`. model (~fairseq.model
(self, sample, model, criterion, optimizer, ignore_grad=False)