Add small comments for MonolingualDataset and TokenBlockDataset

Summary: Pull Request resolved: https://github.com/pytorch/fairseq/pull/669

Differential Revision: D15114160

Pulled By: myleott

fbshipit-source-id: 64f4a8154c8931ddbbe459d4d4a54c46680ad6b6

fairseq/data/monolingual_dataset.py

@@ -78,6 +78,14 @@ class MonolingualDataset(FairseqDataset):
 
     def __getitem__(self, index):
         if self.targets is not None:
+            # *future_target* is the original sentence
+            # *source* is shifted right by 1 (maybe left-padded with eos)
+            # *past_target* is shifted right by 2 (left-padded as needed)
+            #
+            # Left-to-right language models should condition on *source* and
+            # predict *future_target*.
+            # Right-to-left language models should condition on *source* and
+            # predict *past_target*.
             source, future_target, past_target = self.dataset[index]
             source, target = self._make_source_target(source, future_target, past_target)
         else:

fairseq/data/token_block_dataset.py

@@ -112,8 +112,8 @@ class TokenBlockDataset(FairseqDataset):
 
         if self.include_targets:
             # *target* is the original sentence (=item)
-            # *source* is rotated left by 1 (maybe left-padded with eos)
-            # *past_target* is rotated left by 2 (left-padded as needed)
+            # *source* is shifted right by 1 (maybe left-padded with eos)
+            # *past_target* is shifted right by 2 (left-padded as needed)
             if s == 0:
                 source = torch.cat([item.new([self.eos]), buffer[0:e - 1]])
                 past_target = torch.cat([item.new([self.pad, self.eos]), buffer[0:e - 2]])