GSoC 2021 - Refactoring mne.connectivity into mne-connectivity repository

[adam2392]

This issue is meant to track the progress of refactoring existing functionality inside mne-python's "connectivity" sub-module into an entirely new repository.

• Create new repository, mne-connectivity, inside mne.tools (I think I have access)
• Get CI and basic GH action tools setup, based on mne-bids and mne-python CI
• Refactor existing code into the mne-connectivity repo and include a validation notebook that validates that code in this refactored setting matches what is here

To track detailed progress inside mne-connectivity, such as getting CI up and running, I'll point to issues in that specific repository.

Higher level goals

I think @agramfort spoke about this a bit with me and I agree: there is really no "great way" of saying which algorithm or model is better in certain instances. There are many "proposed connectivity algorithms", whether they are time-domain, or free-domain based. So the metric moving forward in this new repo is to at least standardize what you should expect on a dataset.

Given 1-3 datasets, running a specific algorithm, one should expect certain performance or values. The current goal is to leverage openneuro and pull datasets for doing this in CI.

Reference Issues

mne-tools/mne-connectivity#150 mne-tools/mne-connectivity#50 mne-tools/mne-connectivity#17

[larsoner]

@adam2392 I read this only after making my comment in the other repo, which seems to echo almost 1:1 what you wrote here -- so sounds good to me :)

The current goal is to leverage openneuro and pull datasets for doing this in CI.

I would start smaller and try using toy or simulated datasets -- preferably there are some of these in the reference papers. Once that works, examples that build on CircleCI to give outputs for real-data inputs are great, but I wouldn't start from these if possible because debugging them will be harder both in terms of eyeballing errors and in terms of CI time (they are going to require a big download and a long processing time compared to a toy example).

[adam2392]

@drammock @larsoner and @britta-wstnr and I were discussing today and they suggested I ask you:

What do we do with all the mne.connectivity examples inside mne-python?

• should we have a link directly to the mne-connectivity package?
• what’s the best way to advertise this?

Any preferences there?

[drammock]

What do we do with all the mne.connectivity examples inside mne-python?

I can think of 2 broad options:

1. move the examples to the doc site of mne-connectivity
2. leave them on the MNE-Python site, but import the needed functions from mne-connectivity

I think of those I prefer option 1. Each examples category already includes a short text description, so it would be easy to edit that to say something like "connectivity analysis algorithms have been moved to a separate package, mne-connectivity. See the examples there." (and include a link).

[larsoner]

Agreed, my inclination is to have the "Connectivity" section of the examples now just be a single sentence pointing to the mne-connectivity examples page. And the when you click on the front page icon that advertises connectivity, it should take you to the mne-connectivity examples page.

[britta-wstnr]

Yes, that was also my thought, +1 for option 1. I think it is important though that connectivity analysis in some ways stays advertised on the main MNE-Python homepage, as it is something many users look for.
This will also make it easier and cleaner to add more examples later in the connectivity module itself.

[wmvanvliet]

This sounds like something I could merge the conpy module into.