braindecode/eegpt-pretrained

This model has been pushed to the Hub using the PytorchModelHubMixin integration:

• Code: https://braindecode.org
• Paper: [More Information Needed]
• Docs: https://braindecode.org/stable/generated/braindecode.models.EEGPT.html

Input geometry (config.json)

The encoder weights are the target_encoder of the original EEGPT-large checkpoint eegpt_mcae_58chs_4s_large4E.ckpt, pretrained on 58 channels resampled to 256 Hz in 4-s windows (1024 samples; patches of 64 samples = 250 ms; Wang et al., NeurIPS 2024; BINE022/EEGPT pretrain/engine_pretraining.py, img_size=[58, 256*4]). config.json therefore declares sfreq=256 and n_times=1024. The original linear-probe and fine-tuning scripts in downstream/ (BCIC-IV-2a/2b, KaggleERN, PhysioNet P300, Sleep-EDF) also bring their data to 256 Hz before the encoder; BCIC-IV-2a, recorded at 250 Hz, is interpolated to 1024 samples. n_chans=62 and chs_info list the model's channel vocabulary: the 62 rows of chan_embed, in the original CHANNEL_DICT order. Four of them (AF7, AF8, PO5, PO6) were not in the 58-channel pretraining montage, so pretraining never trained their embeddings. With your own data, pass chs_info, n_times and sfreq to from_pretrained; the channel IDs are rebuilt from the channel names. The default chan_proj_type="conv1d_constraint" adds a new 1x1 projection onto the 19 10-20 channels used by the original BCIC-IV-2a probe. chan_proj_type="none" feeds your channels to the encoder directly. The model does not resample, so resample your data to 256 Hz first.

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
on004362	EEG Motor Movement/Imagery Dataset (EEGMMIDB; 'PhysioMI' in the paper)	pretraining	EEG
nm000172	High-Gamma Dataset (HGD; Schirrmeister et al. 2017; 14 subjects)	pretraining	EEG
Wang2016	Tsinghua SSVEP benchmark dataset (TSU; Wang et al. 2016; 35 subjects, 40 targets) (NEMAR nm000126 holds a subset of it)	pretraining	EEG
SEED	SJTU Emotion EEG Dataset (15 subjects; Preprocessed_EEG release)	pretraining	EEG
M3CV	M3CV: Multi-subject, Multi-session, Multi-task EEG biometrics database (106 subjects; Huang et al. 2022) (NEMAR nm000166 holds a subset of it)	pretraining	EEG

• self-supervised (alignment + masked reconstruction), 58 ch, 256 Hz, 4 s; 90% of each dataset's samples for training, 10% validation; PhysionetMI: all 109 subjects, MI+ME runs; HGD: train and test files of all 14 subjects; TSU: all 35 subjects/40 targets
• Sources: Wang et al. NeurIPS 2024 (proceedings PDF 4540d267...-Paper-Conference.pdf) p.6, Table 1 and text; Wang et al. NeurIPS 2024, Sec. 3 (pretraining strategy); https://github.com/BINE022/EEGPT README (Pretrained Models); https://github.com/BINE022/EEGPT/blob/main/datasets/pretrain/prepare_pretrain_dataset.py.

Pretrained components

The released EEGPT encoder is pretrained; the task probe/head and the default channel projection are newly initialized in braindecode. Train those components for the downstream task rather than treating this as a pretrained classifier.

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},
}