Skip to content

a ducttape workflow for neural machine translation

Notifications You must be signed in to change notification settings

shuoyangd/tape4nmt

Repository files navigation

tape4nmt cloud build status

What is that?

tape4nmt is a DuctTape workflow created to replace your bash scripts for running NMT experiments. DuctTape is a workflow management system written by Jonathon Clark. In general, if you are writing bash scripts to run any experiments, you may want to consider using DuctTape instead. It organizes your experiment scripts such that:

  1. you are much less likely to do stupid things by e.g. deleting your best model by accident, use the wrong data for the experiment, the list goes on;
  2. you can easily run, delete, and/or re-run batch experiments, re-using any partial results possible, with correctness guarantee;
  3. you can easily re-use part of your modularized experiment script (called "workflow") to create new workflows.

I barely get rid of EMS. Why creating another workflow for MT?

Face the truth: MT researchers are dealing with pipeline, even we are having a end-to-end system. For example, this is a minimalistic NMT system building pipeline with BPE:

Minimalistic Workflow

Doing this by hand, from my own experiences in several evaluation campaigns, is very error-prone and inefficient. Hence the need for a experiment management system.

tape4nmt also fixes a lot of problems EMS has. For example, EMS consolidated workflow management and workflow definition into one gigantic perl script, with some regular expressions failing about every 6 months. tape4nmt however, uses DuctTape for workflow management and only take cares of workflow definition. It also keeps data/workflow separation in mind throughout its design. Besides, hopefully you'll also find tape4nmt has a better structure for maintainence.

tape4nmt aims to support all major NMT toolkits on the market. We currently support fairseq, sockeye, and OpenNMT-py, with marian support in the near future. Also, if your favorite NMT toolkit hasn't been supported yet, you should only need to create one extra tape file containing toolkit-specific operations (binarize data, train, and decode). (If you actually did, please consider contributing your effort!)

I'm sold. How do I start to use it?

The workflow is pre-configured to build an IWSLT German-English system with fairseq. Follow this tutorial to get started.

Note that this workflow is tested on the sun grid manager configured on CLSP grid out-of-box, so you may need to slightly tweak action_flags and resource_flags in the main.tconf file to be able to use this on other grid that uses sun grid manager. Or in the worse case, you could substitute all submitter with "bash" to run everything locally, or define your own submitter for grid manager of your choice!

Hey I found your workflow does not do xxx...

This workflow is not supposed to include everything at the first place! I created this repository mainly in hope that people could benefit from running a basic workflow to learn the basics, and then modify it to do whatever they would like to do.

In order to modify this repository properly, you need to learn how to write DuctTape workflow. DuctTape itself contains a pretty comprehensive tutorial. Nathan Schneider has a better one, albeit seems incomplete.

What are the limitations I should keep in mind while using this?

While we try our best to make the workflow intuitive and easy to use, it should be kept in mind that our upstream project DuctTape hasn't been developed for quite a while. As a consequence, there are some loose ends that haven't been properly tied, e.g., this and this, which means there a moderate amounts of work-arounds we have to make (e.g., all the dummy tasks we've made) to achieve what we aim to do. I plan to make a short-list of failure cases that you can hopefully avoid should you modify the workflow to achieve your goal.

What if I want to contribute?

Pull requests are always appreciated!

About

a ducttape workflow for neural machine translation

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published