[CI] Enable CI and setup all logistics

[adam2392]

First step for #11

Summary for CI

We should keep CI and setup and installation workflow relatively similar. From my exp in OSS, this alleviates developer burden and overall technical debt.

This PR adds additional checks to the CI that are based off best practices:

• circleCI will build documentation for every PR and commit to this repo
• GH actions has a job to link output of circleCI to be linked to the checks
• GH actions will test for style using poetry
• GH actions will test for building the overall package
• GH actions will run unit tests
• GH actions will upload code coverage from unit tests
• GH actions will also automatically release on pypi given a tag is made on main branch

TODO for CI:

• GH actions should also weekly test integration suite. I.e. this will eventually encompass networkx, dowhy, pywhy_graphs and more.
• ideally poetry doesn't install all dependencies for every single job it runs

Summary for repo docs

I've added more details to the README based on my experience. What would be nice is to extend the main figure in dowhy's README to encompass causal discovery. Otw, I think short READMEs are the way to go and then linking to the documentation.

Summary for code documentation

I've uploaded boiler plate documentation that will leverage the pydata sphinx layout. Inside we should track: contributors, and PRs. The API is also auto-rendered here via sphinx.

cc: @robertness

[darthtrevino]

LGTM; only comment is setup.py is no longer necessary

[darthtrevino]

We don't need setup.py with Poetry, on package build it will auto-generate a setup.py with metadata from pyproject.toml.

[adam2392]

Don't we if a developer git clones and wants to run pip install -e .? python-poetry/poetry#34

[darthtrevino]

Does poetry install -E <extra_name> -E <other_extra_nome> not work for you?

[adam2392]

No that only installs dev dependencies. Currently in python-poetry/poetry#34 it is said that one cannot run the equivalent of pip install -e ..

[adam2392]

@darthtrevino do you know why the circleCI is not catching the install for matplotlib?

Also do you know why poetry bugs out when installing numpy for python v3.10?

These are the only 2 issues left before this is good to merge.

[robertness]

Assuming "structure learning" is a bigger scope than "causal discovery" Let's narrow scope down to "Python library for causal discovery (causal structure learning)."

[robertness]

Suggested rephrasing

dodiscovery is a Python library for causal discovery (causal structure learning). Causal discovery is useful when you aren't sure what specific causal questions you have in your domain and want to start by exploring causal relationships in the data. dodiscovery also supports finding the causal structure you need to target specific causal inferences (e.g., cause effect inference).

I think it is best to avoid framing global causal discovery necessarily as the first step of a cause inference pipeline. Causal discovery tends to focus on getting the "right" graph given the data you have, and most of causal effect inference methods focus on getting the "right" causal effect in a way that is robust to not knowing all of the graph or observing all of the data. I think we'll address this with learning ADMG's, providing Markov blanket discovery, etc. But we shouldn't overpromise.

[adam2392]

Done.

[robertness]

Requested changes to README, otherwise LGTM

[robertness]

Suggested rephrasing

dodiscovery is a Python library for causal discovery (causal structure learning). Causal discovery is useful when you aren't sure what specific causal questions you have in your domain and want to start by exploring causal relationships in the data. dodiscovery also supports finding the causal structure you need to target specific causal inferences (e.g., cause effect inference).

I think it is best to avoid framing global causal discovery necessarily as the first step of a cause inference pipeline. Causal discovery tends to focus on getting the "right" graph given the data you have, and most of causal effect inference methods focus on getting the "right" causal effect in a way that is robust to not knowing all of the graph or observing all of the data. I think we'll address this with learning ADMG's, providing Markov blanket discovery, etc. But we shouldn't overpromise.

[adam2392]

Gonna try then what was suggested in discord.

[adam2392]

Just tryna wrangle with the circleCI docs building

[adam2392]

Cross-linking with poetry issue: python-poetry/poetry#6177

It seems there is an overall flaw in poetry that does not allow one to specify separate dev-dependencies. Therefore dev-dependencies must all be installed for any dev step (e.g. matplotlib must be installed to do style checking).