Although the code looks nice, I found this code isn't intact. If someone is trying to use it, please be careful.

1. maxlen: it shouldn't be the fixed-length as it defined as 15. In the paper, section 4.3 Training Proceduce, author uses the other method which is corresponding the homogeneous_data.py in the paper's code.
2. Beam search: this code doesn't implement beam search.

So far, I only found out these things and I decide to use the paper's original code instead of this. Anyway, I still appreciate the tensorflow attempt. The code is clear and beautiful. Thanks for sharing.