fix bug for masking (#752)

Summary:
Pull Request resolved: https://github.com/pytorch/fairseq/pull/752

previously we sample masked tokens with replace=True (default). Because of this, we would mask same tokens multiple times, which will make us mask less tokens finally

Reviewed By: liaimi

Differential Revision: D15403556

fbshipit-source-id: cf12eeb13f9610431136a345de9199ad0292984b

fairseq/data/masked_lm_dataset.py

@@ -152,7 +152,7 @@ class MaskedLMDataset(FairseqDataset):
         masked_sent = np.copy(sentence)
         sent_length = len(sentence)
         mask_num = math.ceil(sent_length * self.masking_ratio)
-        mask = np.random.choice(sent_length, mask_num)
+        mask = np.random.choice(sent_length, mask_num, replace=False)
         target = np.copy(sentence)
 
         for i in range(sent_length):