fix bug for masking prob (#758)

Summary:
Pull Request resolved: https://github.com/pytorch/fairseq/pull/758

Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/603

fixed a typo for _mask_block of mlm. This typo will make we never set masked token as random token, which should take 10% of the masked tokens.

Reviewed By: akinh

Differential Revision: D15492315

fbshipit-source-id: 1e03dc862e23a6543e51d7401c74608d366ba62d

fairseq/data/masked_lm_dataset.py

@@ -166,7 +166,7 @@ class MaskedLMDataset(FairseqDataset):
 
                 # replace with random token if probability is less than
                 # masking_prob + random_token_prob (Eg: 0.9)
-                elif rand < (self.masking_ratio + self.random_token_prob):
+                elif rand < (self.masking_prob + self.random_token_prob):
                     # sample random token from dictionary
                     masked_sent[i] = (
                         np.random.randint(