Added ONNX support

[jacobf18]

PR Description

Added ONNX support to iclabel. I changed the function run_iclabel to include a library parameter. This parameter's default value is set to use PyTorch. I added unit tests to check the network outputs against MATLAB. ONNX is faster to load and more lightweight than PyTorch.

The options for the library parameter are pytorch and onnx so far. It throws a ValueError for any other input.

Closes #125.

Merge checklist

Maintainer, please confirm the following before merging:

• All comments resolved
• This is not your own PR
• All CIs are happy
• PR title starts with [MRG]
• whats_new.rst is updated
• New contributors have been added to CITATION.cff
• PR description includes phrase "closes <#issue-number>"

[mscheltienne]

Thanks @jacobf18 ! That looks good to me.

As you said, we are going to need to add some logic to load the libraries. I'm also +1 to make Qt an optional dependency. Overall, +1 to merge as is and clean-up, add library selection logic, and so on.. in a second PR.

[mscheltienne]

Any reason to change this?

[adam2392]

If we can make the dependencies optional, then we should nest the import statements for torch and onnx inside the functions that use them.

Perhaps we should add a helper function that does the import and checks that at least one of them is installed?

[adam2392]

Is there a way we can make both optional dependencies, since the user only needs to install one?

[mscheltienne]

Yes, we need to add some selection logic and get rid of the import at the top-level. Probably requires also 2 separate files, one for pytorch and one for onnx since the classes defined inherit from torch.nn directly. Maybe some logic like psychopy has here in an __init__.
Anyway, I think that could be done in a second PR which restructure the iclabel module before a 0.5 release.

[adam2392]

Alright. Sounds good to me. LGTM with the exception of CI. Does it just need a restart possibly?

Thanks @jacobf18!

[mscheltienne]

No the ubuntu ones have been broken for a while because of some xvfb stuff, it will be a good occasion to fix them as well..

[mscheltienne]

@jacobf18 Do you want to give a shot to a logic to set pytorch and onnx as optional?
Ideally:

• import mne_icalabel and from mne_icalabel import label_components would work regardless of the installation status of the backends.
• import mne_icalabel.iclabel and from mne_icalabel.iclabel import label_components would fail if neither backend is available.

A couple of ideas: we should move the network part to a mne_icalabel.iclabel.network module, with 3 files:

• __init__.py which attempts to load both in a similar fashion to psychopy with the drivers for the parallel port
• pytorch.py with the pytorch backend
• onnx.py with the onnx backend

Or maybe an empty __init__.py and some helper function which handles the loading of the backends.
The tests might also require a decorator @requires_onnx or @requires_torch.

---

Let me know if you want to give it a shot, else I'll do it in the coming days. Once it's sorted out, we can release a 0.5 version with this addition, and maybe get this correctly packaged on conda-forge for windows.

[mscheltienne]

Merging and picking this up today with another PR to add the selection of the backend. Let's not stall this :)