budzianowski / pytorch-beam-search-decoding Goto Github PK

How to use the value 'decoded_batch' in function 'def beam_decode' when I test my own model?

When decoding a sentence of length 5 and beamsize=10, and we entered the while loop. After the first while loop, we end up with a queue with length 10. Now we move to T=1. According to the code, we have to run 10 while loops to decode for T=1. So after decoding for T=1, we will end up with a queue of length 100 instead of 10 cause there is no topK selection in the queue. After decoding for T=2, we will have a queue with length 1000. This queue will explode quickly. Did I miss something here?

I had this problem. I found out that this happens when in the PriorityQueue(PQ) two elements have the same value then, the algorithm moves to the element itself. Because the element in the PQ is an object it generates this Error.

The solution was that I add in the class a special function to handle this by choosing the element with the highest probability., as follows. If this seems ok to you, you can add this lines in your code. Or I can make a pull request.

#function in the class BeamSearchNode
def __lt__(self, other):
        return self.prob < other.prob

Hi,

Thank's for this work. It's really usefull.

Why you defined the score function as (line 66)??

self.logp / float(self.leng - 1 + 1e-6) + alpha * reward

Can you explain this definition?

Thank's,
Best

Why do you use 'number_required' in your code? When topk=1, the beam search algorithm only generates one sentence!

Could you please, describe more beam search decoding ?