[ENH] Refactor code to enable a benchmark that "matches" ICLabel on simulations and test datasets

[adam2392]

PR Description

@mscheltienne this branch should contain a working version of what we suspect is as "close as we can get" to a matching ICLabel Network.

It doesn't match 100% as of right now, which is made explicit in the unit tests. It is also still a bit rough, but I've laid out the entire project now as a structured open-source python repository.

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
• PR description includes phrase "closes <#issue-number>"

[mscheltienne]

Amazing! I'll give it a quick try next week, I'm curious to see how close you got :)

[mscheltienne]

Quick update: our intern should have started on the 1st of March. But he did not yet get his work permit (mandatory in Switzerland), thus his start has been delayed and we hope he will be on board next week.

---

I finally took the time to test it! It looks very promising but seems to favor a lot 'Other'.
Am I correct in assuming that the classes outputted are in the order:

classes = ['Brain', 'Muscle', 'Eye', 'Heart', 'Line Noise', 'Channel Noise', 'Other']

I ran it on this file quickly:

from matplotlib import pyplot as plt
import mne
import numpy as np

from mne_icalabel.ica_label import ica_eeg_features
from mne_icalabel.ica_net import run_iclabel

#%% Load
raw = mne.io.read_raw_fif('ica-test-raw.fif', preload=True)

#%% ICA
ica = mne.preprocessing.ICA(method='picard', max_iter='auto')
ica.fit(raw, picks='eeg')

#%% Features
features = ica_eeg_features(raw, ica)
topo = features[0].astype(np.float32)
psds = features[1].astype(np.float32)
autocorr = features[2].astype(np.float32)
labels = run_iclabel(topo, psds, autocorr)

classes = ['Brain', 'Muscle', 'Eye', 'Heart', 'Line Noise', 'Channel Noise', 
           'Other']

#%% Plot
f, axes = plt.subplots(1, 7, sharex=True, sharey=True, figsize=(20, 5))
for k, ax in enumerate(axes):
    ax.bar(np.arange(0, labels.shape[0]), labels[:, k])
    ax.set_title(classes[k])

for i, label in enumerate(labels):
    k = np.argmax(label)
    axes[k].bar(i, label[k], color='crimson')
    
f.tight_layout()

#%% ICA plots
ica.plot_components()
ica.plot_sources(raw)

Eye and heartbeat are correct. The last time we talked, you told me it failed at classifying even simple blinks, did you find what caused this?