Created by: Superjomn

The beam search is a core algorithm for machine translation and other generation models, the old version of PaddlePaddle implement this as a c++ module in RecurrentLayerGroup and lacks flexibility.

In our further design of refactoring, a more flexible beam search is needed, the new one should be a composition of some different ops, and a user is free to change the details of a beam search.

That means, to implement beam search, more operators should be implemented, and we might need to change the framework.

Let's survey TensorFlow and other mature op-based frameworks first.

The current plan is

• survey TF and MXNet
  • determine the design of a beam search, a PR should be provided. (Let's finish this before next Monday)
  • collect all the related operators, create a project and let others join in implementing. (before next Wednesday)
• implementing operators
  • all others are free to join and develop parallelly.