Update 'log_probs' to for alignment. We adjust 'log_probs' to disallow p.sentence_boundary_token_id or EOS if emitting it would result in a misaligned output with an unequal number of sentences. Args: log_probs: encoder's log_probs output. source_num_sentences: shape [b
(self, log_probs, source_num_sentences, hyp_num_sentences)
source not stored for this graph (policy: none)
no outgoing calls
no test coverage detected