search

Conchylicultor / DeepQA

My tensorflow implementation of "A neural conversational model", a Deep learning based chatbot
Apache License 2.0
2.93k stars 1.17k forks source link

If we use dynamic rnn encorder do we still need to use the bucket model? #119

Open shamanez opened 7 years ago

Conchylicultor commented 7 years ago

Dynamic RNN allows the model to have variable sentence lengths​. Bucketing allow to efficiently minimize the padding within each batch. Those are two different complementary things.

shamanez commented 7 years ago

Oh I get it. So if we use dynamic RNN encoder without bucketing dynamic rnn function will padded the input to a fixed size right?

Conchylicultor commented 7 years ago

The inputs will be padded to the longest sentence of the current batch. I think that with recent versions of tensorflow, there was a way to automatically handle padding and bucketing.