OpenNMT/OpenNMT-tf

Multi-decoder setup for multi-task learning

Open

#130 opened on May 22, 2018

View on GitHub
 (5 comments) (0 reactions) (0 assignees)Python (1,428 stars) (394 forks)batch import
enhancementhelp wanted

Description

Similarly to the ParallelEncoder, a ParallelDecoder setup could allow multi-task learning. This should not be too hard to implement but we need to take care of some details:

  • support separate values for the decoding parameters (beam_width, length_penalty, etc.),
  • parts of SequenceToSequence assume a single output head (e.g. loss computation, reverse vocabulary lookup, exported outputs for model serving, etc. which should be moved in the decoder itself)

Contributor guide

Tech stack
pythontensorflow
Domain
machine learningai
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
PythonTensorFlowsequence to sequence modelsOpenNMT tf codebase
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30
Research direction
Examine the existing `ParallelEncoder` implementation for guidance. Focus on the `SequenceToSequence` class to identify all components that assume a single output head, such as loss computation and reverse vocabulary lookup. Review decoder parameter handling to understand how to support separate beam width and length penalty per decoder. Check linked issues and PRs for any prior discussion or partial implementations.