Backport "Fix handling of sequence_length w/ beam search decoder"

[guillaumekln]

System information

• OS Platform and Distribution: Ubuntu 16.04
• TensorFlow version and how it was installed: binary
• TensorFlow-Addons version and how it was installed: binary
• Python version: 3.6
• Is GPU used? No

Describe the bug

For some reasons, patches are still pushed to tf.contrib.seq2seq, see:

tensorflow/tensorflow@8bbc039

The bug is described in the commit message.

Code to reproduce the issue

None.

Other info / logs

@qlzh727 as you refactored the seq2seq module, do you confirm this bug is also applicable to the addons version?

[qlzh727]

Thanks for the notice, let me take a look for the change in contrib.

As a side note, since addons hasn't been port into google internal yet, eng within google is still using existing contrib version of seq2seq.

[qlzh727]

I think tensorflow/tensorflow@8bbc039 should be mirrored to addons. let me send a PR quickly

[ebrevdo]

@qlzh727 thanks so much for handling this!