braindecode/STEEGFormer-base

1---2library_name: braindecode3tags:4- braindecode5- STEEGFormer6- eeg7- foundation-model8license: mit9---10 11# STEEGFormer (base)12 13ViT-MAE EEG foundation model — braindecode port of **ST-EEGFormer** (base variant).14 15## Provenance16 17- **Weights ported from:** [LiuyinYang1101/STEEGFormer](https://github.com/LiuyinYang1101/STEEGFormer),18  release [`ST-EEGFormer-base`](https://github.com/LiuyinYang1101/STEEGFormer/releases/tag/ST-EEGFormer-base)19  (asset `checkpoint-288.pth`).20- **Upstream license:** MIT. The braindecode wrapper code is BSD-3-Clause.21- The pre-trained encoder is loaded faithfully (numerical equivalence verified:22  pre-encoder bit-exact, post-encoder relative error ~4e-6). The MAE decoder is23  dropped and the classification head is re-initialised.24 25## Architecture26 27| | embed_dim | depth | num_heads | patch_size | channel vocab |28|---|---|---|---|---|---|29| base | 768 | 12 | 12 | 16 | 145 |30 31Channel positions are resolved from electrode names in `chs_info` (145-slot shared montage vocabulary).32 33## Usage34 35```python36from braindecode.models import STEEGFormer37 38model = STEEGFormer.from_pretrained(39    "braindecode/STEEGFormer-base",40    n_outputs=4, n_chans=22, n_times=1000, chs_info=chs_info,41)42# Encoder features:  out = model(x, return_features=True); out["features"]43```44 45## Citation46 47Yang, L., Sun, Q., Li, A. & Van Hulle, M. M. (2026). *Are EEG foundation models48worth it? Comparative evaluation with traditional decoders in diverse BCI tasks.*49ICLR 2026. https://openreview.net/forum?id=5Xwm8e6vbh50