fix a bug when resuming training from the last epoch (#1275)

Summary:
If the training stopped in the middle of the last epoch, and then it was resumed from checkpoint, it will not continue the training because `epoch_itr.epoch < max_epoch` is not satisfied. This PR fixed the issue.
Pull Request resolved: https://github.com/pytorch/fairseq/pull/1275

Differential Revision: D18483945

Pulled By: myleott

fbshipit-source-id: 80df6f73fa17606a79a28e8328bb4c577f504683

train.py

@@ -76,7 +76,12 @@ def main(args, init_distributed=False):
     train_meter = StopwatchMeter()
     train_meter.start()
     valid_subsets = args.valid_subset.split(',')
-    while lr > args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates() < max_update:
+    while (
+        lr > args.min_lr
+        and (epoch_itr.epoch < max_epoch or (epoch_itr.epoch == max_epoch
+            and epoch_itr._next_epoch_itr is not None))
+        and trainer.get_num_updates() < max_update
+    ):
         # train for one epoch
         train(args, trainer, task, epoch_itr)