Fix a bad merge from D24154657 - second try

Summary: This removes a line that reset the dummy batch in begin_epoch. In actuality the batch is reset in get_{train,valid}_iterator, and this line was resetting it to None unnecessarily.

Reviewed By: joshim5, jhcross

Differential Revision: D24157057

fbshipit-source-id: 59ac68327094ceff70f66d7b471fa810997fe84e

fairseq/trainer.py

@@ -413,9 +413,6 @@ def begin_epoch(self, epoch):
         # task specific setup per epoch
         self.task.begin_epoch(epoch, self.get_model())
 
-        # reset dummy batch
-        self._dummy_batch = None
-
         if self.tpu:
             import torch_xla.core.xla_model as xm