MegNet 3.12 or above python compatibility.

[moustafalmanla]

We are currently working on integrating MEGnet into MEGqc at the ANCP Lab, University of Oldenburg.

Our MEGqc environment has been updated and tested for compatibility with Python versions up to 3.14. However, we noticed that MEGnet currently specifies support only for Python versions below 3.12, which creates a compatibility constraint for the integration.

We would like to ask whether there are any known technical limitations preventing MEGnet from supporting newer Python versions (3.12–3.14), or whether the current restriction is primarily due to untested compatibility.

any dependencies that currently block support for newer Python versions,
whether there are ongoing plans to extend compatibility,
and whether you would recommend a specific approach for testing or adapting MEGnet for Python 3.14 compatibility.

Thank you for your time and for your work on MEGnet. We look forward to your guidance.

Best regards,
Moustafa Almanla
ANCP Lab

[jstout211]

Hi Moustafa,

The install was pinned below 3.12 because of the compatibility with Tensorflow. It appears that tensorflow is now compatible with py 3.12 and 3.13. I will change the dependency file and do some testing - but currently it still may only get to 3.13 due to the Tensorflow's incompatibility with 3.14. Give me a few days to test, update pypi, and get back to you.

Thanks,
Jeff

[moustafalmanla]

Great, looking forward to hear again from you.

[moustafalmanla]

Dear Jeff,

Any updates on the compatibility?

Thanks,
Moustafa

[jstout211]

Sorry for the delay on this, but these last weeks have been pretty busy. I looked at upgrading the package, but it does not seem as trivial as updating the tensorflow and python version. The newer versions of tensorflow require a v3 keras model (and the model weights are currently saved under v2). The instructions appear straightforward to convert over, but so far it has not seemed to work. I will try some more tomorrow to see if I can convert this to the new format. Thanks for the patience.

[jstout211]

@moustafalmanla -
I pushed the keras V3 model weights to huggingface and started a new branch for the updates of the code. It appears to download and read in the weights. The ICA code runs and it appears to run the classification.

• I still need to update the github actions to test on the updated python
• Github actions does not seem to be finding the correct hugging face model ...
• I want to run a test against a larger known dataset to confirm nothing weird happened with the weight transfer

My local pytest seems to pass (w/ py 3.13). I need to do some further testing, but given those pass I should be able to repackage and update pypi before the end of the week.

[moustafalmanla]

Okay thank you so much for this.

Sorry for the delayed response, last few days been hectic.

Take as much time as you need no rush.
Let me know if I can help with anything.

[jstout211]

The package has been uploaded to pypi as version 0.3.2
It is currently tested against python 3.12 and 3.13 -- github actions pass on a matrix with mne 1.10, 1.11, 1.12.
I will run the final tests soon, but I dont expect those to change.

[jstout211]

@moustafalmanla -
I modified the installer and keras to use pytorch instead of tensorflow (this version is 0.3.3 and is on pypi now). This allowed python to be increased to 3.14. Github actions passes for the matrix of mne>=1.10 and python>3.10. The GH actions tests the classifier against 15 datasets (saved as pre compiled numpy arrays). The preprocessing pipeline testing is performed on one dataset for direct np.all assertions to confirm no changes in the outputs, but most of the upstream testing for consistency is expected to be performed by the MNE group.

I still want to run this on our internal training batch and confirm we are getting the same outputs, but it seems like everything so far is ok.

[moustafalmanla]

Great news,

thank you so much