m2lines / equationdisco

M²LInES equation discovery package.

🚧 🚧 M²LInES Equation discovery package. 🚧 🚧

This repo will house the M²LInES equation discovery repo, and is under construction.

Who is this package for?

TODO: Populate.

Authors and contributors

The code in this repository was originally developed by Andrew Ross and Pavel Perezhogin, in this repository, for use in the publication: Benchmarking of machine learning ocean parameterizations in an idealized model (published in JAMES, https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2022MS003258).

It is currently being worked on by Jim Denholm and Andrew Ross, along with other members of the ICCS and M²LInES.

License

The work is available under an MIT License; see the license.

equationdisco's Issues

Update README with current state-of-development.

Bring the README up-to-date with the current development and TODOs.

Overview on how to use this package

There is a jupyter notebook in the pyqg_parameterization_benchmarks repo with some guidance on how to use the equation discovery stuff.

• We should move this notebook over and revise it accordingly for this package.
• The result should provide clear end-to-end guidance on how to use this repository.

Breaking the pyqg dependency

The utils.py currently depends on the pyqg repository, but Laure says she would like this equation discovery repository to be completely independent of pyqg .

• Is dropping the pyqg dependency definitely what we want to achieve?
• How can we refactor/redesign this repository such that it is standalone with no pyqg dependency?

Review CI

Would you mind taking a look at the basic GitHub workflows and checking they are okay, please? If so I will create a dev branch.

Distribution of this repository

How should we distribute this repository?

Options:

• Should we release it as a PyPI package?
• Should we release it on a conda channel?
• Should is just stay in a git repo with clear instructions on how to install it?

Refactoring and code quality

After moving the required modules—hybrid_symbolic.py and utils.py—across from pyqg_parameterization_benchmarks, the first useful thing we can do is add some basic pythonic code quality measures:

• Docstrings: make it clear to a general user what each function does and the arguments it requires.
• Type-hinting: allows us to lint the code with mypy in order to make sure we don't pass arguments of the wrong type to any functions.
• Apply Pylint: this means the code will comply with most up-to-date python encouraged practises PEP8, and will present in a standard and readable way.
• Apply black: black is perhaps the most-used Python code formatter and will standardise the code's appearance across the entire repo, making it conform to a standard.
• Apply pydocstyle with numpy convention.

Get demo notebook working

When we are finished refactoring the code ported over from the original repo, the next step will be to get the demo notebook working.

Minimal Testing

The original https://github.com/m2lines/pyqg_parameterization_benchmarks repo does not contain unit tests for the equation discovery module.

A minimal test could be defined to simply import the refactored module and ensure no errors occur.

A secondary functional test could be created by considering the Hybrid Symbolic notebook and using a statistical check to ensure that given the example input used in the notebook an expected value is produced. This would require codifying the check against the visual outputs that are produced in the notebook example.

Addition of some minimal testing would allow the development of more advanced unit tests at a later date and would ensure that a minimal set of tests exists to build a CI pipeline. Eg. #1

Review license

Would you mind taking a look at the license and checking it's okay, please? I'm happy to make any changes you'd like.