enunlg: an Extensible NLG library for Python

Goals:

• provide an environment for our team to build their neural network programming skills using Pytorch
• develop a modular codebase that students can use for projects and that researchers can use for baselines and as the basis for new NNLG systems
• serve as a starting point for collaborations on pipeline NNLG, WebNLG shared tasks, etc

We are starting by re-implementing existing seq2seq models (TGen, SC-LSTM, Neural Checklist) and adapting the implementations to each other.

Design Principles:

• 'engineering' experiments should be trivial (trying different random seeds, hidden layers, kinds of nonlinearities or reccurent cells, types of padding, etc)
• 'research' experiments should be easy (knowing which part of the code needs to be modified to incorporate a new architecture, kind of attention, large pre-trained model, language, domain, or something else)

Project Structure

We are managing dependencies and virtual environments using poetry. Testing is done with pytest.

We will primarily use READMEs for documentation, though maybe the built-in wiki in GitHub will be helpful at some point. We are using GitHub issues to track bugs and plan features and projects.

Getting Started

0. Install poetry into your normal Python environment. poetry will manage dependencies and create a virtual environment for you.

   $ pip install poetry
   

   Note: if you are on MacOS you may need to use homebrew and run brew install poetry instead.

1. Clone this repository.

   $ git clone git@github.com:NapierNLP/enunlg.git
   

2. Edit enlg/pyproject.toml's entry for torch to choose the version matching your environment (or find your own version of torch to use!)

3. cd to the root of the cloned repository and run poetry install, which will install everything in pyproject.toml unless you have a poetry.lock file already

   $ cd enlg
   $ poetry env use 3.9
   $ poetry install
   

Note: if you run into problems with the version of Python available on your system, it is often easy to install another version of Python using your package manager. On Fedora we use the following command with dnf, but the equivalent for Ubuntu would use apt-get and I think you could also use brew on MacOS. You will need to tell poetry to use this version of Python then, with poetry env use 3.9.

   $ sudo dnf install python39

4. Download the E2E Challenge dataset so you have something to test.

   $ ./scripts/fetch_e2e.bash
   

If you don't want to run a random bash script, the URL for fetching the data and the intended directory structure is documented in the script used above.

5. Run poetry shell in your terminal to activate the environment, after which you can run python script/tgen.py (for example) and it will use the virtual environment instance of Python with all the correct dependencies.

   $ poetry shell
   (enlg-SOMECHARS-py3.9) $ python scripts/tgen.py
   

Note: SOMECHARS above represents a sequence of characters which is a hash generated by poetry provide a unique location for the virtual environment it creates.

This model takes about 1400 minutes (23 hours, 10 minutes) to train on an Intel i7-4790K rated at 4.4 GHz (runing Fedora 36 w/Linux kernel 5.19)

Dev environments

• dev: running Python 3.9 on Fedora 33-36. Editing in PyCharm.