Include the CSBrain #1077

Hi @bruAristimunha — I'd like to take this one on, if that's okay.

Quick context: I recently contributed #1164 (MNE >= 1.13 montage rename shim), #1169 (the new-model PR template/checklist) and #1186 (MSCFormer, benchmarked on BCI IV-2a against CTNet under a shared harness). I'd follow the same recipe here:

• Port the CSBrain architecture (yuchen2199/CSBrain, arXiv:2506.23075) as an EEGModuleMixin model with the standard registrations (models/__init__.py, summary.csv, util.py, api.rst, architecture figure, whats_new), keeping the region grouping for the structured sparse attention configurable rather than hardcoded to one montage.
• Benchmark on BCI IV-2a with the same 5-fold per-subject harness used in Add MSCFormer model (multi-scale convolutional transformer for motor imagery) #1186, so the numbers are directly comparable to the MSCFormer/CTNet results there. I can add 2b as well if that's useful.
• Pretrained weights: the official checkpoint is distributed via Google Drive. I'd be happy to re-host a copy on the HF Hub (à la braindecode/cbramod-pretrained) if you want from_pretrained support in the first PR — or land the architecture + benchmark first and treat weights as a follow-up. Your call.

One flag: the upstream repo currently ships no LICENSE file. I'll open an issue there asking the authors to add one (MIT/Apache-2.0/BSD). If that stalls, the fallback is a reimplementation from the paper's equations (the EEGModuleMixin port rewrites most of the module structure anyway), with the official repo used only as a numerical cross-check.

Thanks!

it would be great!

I can help in the weights!

let's assume bsd or unlincese

Thanks @bruAristimunha — great to hear! The PR is up now: #1196. A few points to sync:

• Port fidelity: I verified the port against the official pretrained checkpoint (CSBrain.pth, 8.9M params, SHA-256 cfcdf9dcb069905e5e02e92fe3d59bd3574e779bf554cc02cb701994a79e99cc) — after a pure key rename, feeding the same weights through the upstream reference implementation and through the port produces identical outputs (max abs diff 0.0), so the released weights load as-is into the new module.
• License: understood — assuming BSD/Unlicense as you suggest. I've also opened Please consider adding an open-source license (MIT/Apache-2.0/BSD) yuchen2199/CSBrain#7 asking the author to add an explicit license; if/when one lands I'll add the adapted-from attribution plus a NOTICE entry on top of the current re-implementation wording.
• Weights re-host: yes please, that would help a lot! The official checkpoint is currently Google-Drive-only. An HF re-host would make the benchmark fully reproducible without the Drive hop. I can hand over the original file plus a key-mapped state_dict in our module's naming (with hashes for both) — just tell me which shape is most convenient for you.
• Benchmark: BCI IV-2a, 5-fold cross-session, same harness as MSCFormer/CTNet in Add MSCFormer model (multi-scale convolutional transformer for motor imagery) #1186. CSBrain (official pretrained init) reaches 0.4384 — below its published fine-tuned 2a result (0.5773) and below MSCFormer here, consistent with the systematic harness shift we documented in Add MSCFormer model (multi-scale convolutional transformer for motor imagery) #1186; the full table + figure are in the PR description.