Update --restore-file logic (partially fixes #999)

Summary: Pull Request resolved: https://github.com/pytorch/fairseq/pull/1007

Differential Revision: D16762490

Pulled By: myleott

fbshipit-source-id: d67137bcf581887850323d188bb4ea643a35ac9e

fairseq/checkpoint_utils.py

@@ -97,10 +97,10 @@ def load_checkpoint(args, trainer):
     if args.distributed_rank == 0:
         os.makedirs(args.save_dir, exist_ok=True)
 
-    if os.path.isabs(args.restore_file):
-        checkpoint_path = args.restore_file
+    if args.restore_file == 'checkpoint_last.pt':
+        checkpoint_path = os.path.join(args.save_dir, 'checkpoint_last.pt')
     else:
-        checkpoint_path = os.path.join(args.save_dir, args.restore_file)
+        checkpoint_path = args.restore_file
 
     extra_state = trainer.load_checkpoint(
         checkpoint_path,

fairseq/options.py

@@ -361,7 +361,8 @@ def add_checkpoint_args(parser):
     group.add_argument('--save-dir', metavar='DIR', default='checkpoints',
                        help='path to save checkpoints')
     group.add_argument('--restore-file', default='checkpoint_last.pt',
-                       help='filename in save-dir from which to load checkpoint')
+                       help='filename from which to load checkpoint '
+                            '(default: <save-dir>/checkpoint_last.pt')
     group.add_argument('--reset-dataloader', action='store_true',
                        help='if set, does not reload dataloader state from the checkpoint')
     group.add_argument('--reset-lr-scheduler', action='store_true',