Enable sampling (#571)

Summary:
Pull Request resolved: https://github.com/pytorch/fairseq/pull/571

Enable sampling from Fairseq

Reviewed By: akinh

Differential Revision: D13981666

fbshipit-source-id: 2af1bd67701a73a2c76a9255bd8381d6a7518876

fairseq/search.py

@@ -183,7 +183,7 @@ class Sampling(Search):
             lprobs = lprobs[:, ::beam_size, :].contiguous()
 
         # we exclude the first two vocab items, one of which is pad
-        assert self.pad == 1, 'sampling assumes the first two symbols can be ignored'
+        assert self.pad <= 1, 'sampling assumes the first two symbols can be ignored'
         lprobs_nopad = lprobs[:, :, 2:]
 
         # only sample from top-k candidates