Misc fixes

Summary: Pull Request resolved: https://github.com/pytorch/fairseq/pull/439

Differential Revision: D13608151

Pulled By: myleott

fbshipit-source-id: 198b84995a6329f8329829cc91184d88f1eab947

fairseq/trainer.py

@@ -162,6 +162,7 @@ class Trainer(object):
             torch.cuda.manual_seed(seed)
 
         self.model.train()
+        self.criterion.train()
         self.zero_grad()
 
         if not dummy_batch:
@@ -286,6 +287,7 @@ class Trainer(object):
         """Do forward pass in evaluation mode."""
         with torch.no_grad():
             self.model.eval()
+            self.criterion.eval()
 
             sample = self._prepare_sample(sample)
             if sample is None:

train.py

@@ -375,8 +375,6 @@ if __name__ == '__main__':
     if args.distributed_init_method is not None:
         # distributed training
         distributed_main(args.device_id, args)
-        args.distributed_rank = distributed_utils.distributed_init(args)
-        main(args)
     elif args.distributed_world_size > 1:
         # fallback for single node with multiple GPUs
         port = random.randint(10000, 20000)