Merge internal changes

Changelog:
- `f472d141`: Support tied embeddings in LSTM encoder/decoder
- `89e19d42`: Don't print alignment by default (use `--print-alignment` to re-enable it)
- `d2e2a1d4`: Add Transformer-based language model
- `c2794070`: Add new Transformer configuration for IWSLT
- `2fbfda0d`: Misc changes for pytorch-translate
- Miscellaneous bug fixes

train.py

@@ -32,7 +32,7 @@ def main(args):
     task = tasks.setup_task(args)
 
     # Load dataset splits
-    load_dataset_splits(args, task, ['train', 'valid'])
+    load_dataset_splits(task, ['train', 'valid'])
 
     # Build model and criterion
     model = task.build_model(args)
@@ -316,13 +316,15 @@ def load_checkpoint(args, trainer, epoch_itr):
                 save_checkpoint.best = extra_state['best']
 
 
-def load_dataset_splits(args, task, splits):
+def load_dataset_splits(task, splits):
     for split in splits:
+        if split == 'train':
+            task.load_dataset(split, combine=True)
+        else:
             for k in itertools.count():
                 split_k = split + (str(k) if k > 0 else '')
                 try:
-                task.load_dataset(split_k)
-                print('| {} {} {} examples'.format(args.data, split_k, len(task.dataset(split_k))))
+                    task.load_dataset(split_k, combine=False)
                 except FileNotFoundError as e:
                     if k > 0:
                         break