Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Easier-to-understand naming for LSTMBuilder #474

@neubig

Description

@neubig

LSTMBuilder currently represents a relatively unorthodox implementation of LSTMs using coupled input and forget gates: http://dynet.readthedocs.io/en/latest/unorthodox.html#lstm-implementation

I think very few people will realize this, and this could have a significant impact on results (for example, the DyNet LSTMBuilder will never have a cell state with abs(c)>1, but vanilla LSTMs can). It'd probably be best to have the basic LSTMBuilder represent something that looks more like what people expect, so I propose the following change:

  • Rename LSTMBuilder to CoupledLSTMBuilder
  • Make LSTMBuilder a typedef for VanillaLSTMBuilder

If there aren't any objections, maybe we'll do this in the upcoming 1.1 release: https://github.com/clab/dynet/projects

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions