How to decode? about code2seq HOT 4 CLOSED

Hi @ZhichaoOuyang ,
Thank you for your interest in code2seq!

Yes, the decoding phase generates words one-by-one. It actually decodes exactly the same as seq2seq models (explained here for example: https://youtu.be/IxQtK2SjWWM),
with the only difference that the model computes attention over paths rather than over words.

In every decoding step, the model computes a different weighted average (=attention) of the input paths, and use the weighted average to select the next word.

Let me know if you have any more questions.
Best,
Uri

from code2seq.

Therefore, when decoding, do you still weight the representation z_i of every path and the corresponding attention \alpha_i of each path to get a vector c_t, and then predict the word one by one just like standard NMT models?

In other words, your model formula is the same as the encoder-decoder framework formula described in section 3.1 of your paper, except that z represents the path rather than input source words, right?

from code2seq.

Yes, exactly!

from code2seq.

thanks, I got it.

from code2seq.