braindecode/mirepnet-pretrained

MIRepNet

Braindecode-format re-host of the official MIRepNet checkpoint released by Liu et al. The checkpoint can be loaded directly through Braindecode's standard Hugging Face integration:

from braindecode.models import MIRepNet

model = MIRepNet.from_pretrained("braindecode/mirepnet-pretrained")

# Fine-tuning for another task replaces the released three-class head.
model = MIRepNet.from_pretrained(
    "braindecode/mirepnet-pretrained",
    n_outputs=4,
)

Released configuration

• 45 channels in the order stored in mirepnet_channels.json
• 1,000 samples at 250 Hz
• 256-dimensional embedding
• 6 Transformer blocks with 8 heads
• 3-output supervised pretraining head

The paper's 8--30 Hz filtering, resampling, channel-template preparation, and Euclidean alignment are preprocessing steps and are not performed by the model.

Provenance and conversion

• Official code: https://github.com/staraink/MIRepNet at revision edb80d7605f75ba8b72b417a124cc9db07385f72
• Official checkpoint: https://huggingface.co/starself/MIRepNet at revision 9bac0439c0d3e9ffdb40ca675d61a51b439a446e
• Source file: MIRepNet.pth, SHA-256 432288958007e344a5a84a9ffe9d0e5e5c0cb616aef86c85522375a3f4da9aaf

All 109 downstream tensors were converted. The 34 pretraining-only tensors (mask_token, decoder.*, and the upstream embedding.chan_embed.weight, which is not used by the released forward pass) were intentionally omitted. Against the official implementation, maximum absolute error was 2.38e-7 for pooled features and 1.19e-7 for logits. The conversion is reproducible with convert_mirepnet_checkpoint.py.

The source code and checkpoint are distributed under the MIT License. The original copyright notice is preserved in LICENSE.

Training data

The data that shaped these weights. The ids are the datasets: field of this card, which braindecode reads into model.card_data.datasets when you call from_pretrained (MOABB class names for MOABB datasets, OpenNeuro / NEMAR accessions, Hub ids otherwise). nm/on ids are datasets on NEMAR (on = NEMAR's copy of the OpenNeuro dataset with the same number). Rows marked private: are data the authors did not release; they are not in datasets:.

datasets: id	Dataset	Stage	Modality
nm000171	BNCI 2014-002 (Steyrl et al., right hand vs both feet MI)	pretraining	EEG
on004362	PhysioNet EEG Motor Movement/Imagery Dataset (BCI2000)	pretraining	EEG
nm000250	Dreyer et al. 2023 large left/right hand MI dataset (87 subjects; A+B+C)	pretraining	EEG
nm000146	Yi et al. 2014 simple/compound limb MI (Weibo2014)	pretraining	EEG
nm000115	Zhou et al. 2016 three-class MI	pretraining	EEG
nm000338	Lee et al. 2019 OpenBMI, motor-imagery part	pretraining	EEG
nm000245	Cho et al. 2017 left/right hand MI (GigaDB)	pretraining	EEG

• Hybrid pretraining: masked token reconstruction + supervised MI classification (the released 3-output head = left hand / right hand / feet union of pretraining labels; order undocumented). Downstream datasets (BNCI2014_001, BNCI2015_001, BNCI2014_004, AlexMI) were NOT used for these weights.
• Sources: arXiv:2507.20254v1, Section IV-A Datasets; https://github.com/staraink/MIRepNet README @ edb80d7.

Limitations

The official repository does not document the semantic ordering of the three pretraining-head outputs. Replace the head with n_outputs=... and fine-tune it for downstream use unless that label mapping has been independently verified. Dataset licenses are separate from the checkpoint's MIT license.

Citation

@article{LIU2026115966,
  title = {MIRepNet: A pipeline and pre-trained model for EEG-based motor imagery classification},
  journal = {Knowledge-Based Systems},
  volume = {343},
  pages = {115966},
  year = {2026},
  issn = {0950-7051},
  doi = {10.1016/j.knosys.2026.115966},
  url = {https://www.sciencedirect.com/science/article/pii/S0950705126006921},
  author = {Dingkun Liu and Zhu Chen and Jingwei Luo and Shijie Lian and Yuheng Chen and Shaojie Hou and Xiaolian Zhu and Dongrui Wu}
}

Cite Braindecode

If you use this braindecode port in scientific work, cite the software on Zenodo, doi:10.5281/zenodo.17699192, and the original model and datasets cited above.

@software{braindecode,
  author = {Aristimunha, Bruno and
            Guetschel, Pierre and
            Wimpff, Martin and
            Gemein, Lukas and
            Rommel, Cedric and
            Banville, Hubert and
            Sliwowski, Maciej and
            Wilson, Daniel and
            Brandt, Simon and
            Gnassounou, Théo and
            Paillard, Joseph and
            {Junqueira Lopes}, Bruna and
            Sedlar, Sara and
            Moreau, Thomas and
            Chevallier, Sylvain and
            Gramfort, Alexandre and
            Schirrmeister, Robin Tibor},
  title = {Braindecode: toolbox for decoding raw electrophysiological brain data
           with deep learning models},
  url = {https://github.com/braindecode/braindecode},
  doi = {10.5281/zenodo.17699192},
  publisher = {Zenodo},
  license = {BSD-3-Clause},
}