braindecode/emg2qwerty-generic

emg2qwerty — generic baseline (TDS-Conv-CTC, 108-subject pretrained)

Pre-remapped braindecode-compatible copy of the upstream generic.ckpt from the emg2qwerty release (Sivakumar et al., NeurIPS 2024 D&B Track).

from braindecode.models import EMG2QwertyNet

model = EMG2QwertyNet.from_pretrained("braindecode/emg2qwerty-generic")

Source

Upstream repository: https://github.com/facebookresearch/emg2qwerty Paper: Sivakumar V, Seely J, Du A, Bittner S, Berenzweig A, Bolarinwa A, Gramfort A, Mandel M. emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography. Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2024.

Modification from upstream

Upstream models/generic.ckpt is a PyTorch-Lightning checkpoint of emg2qwerty.lightning.TDSConvCTCModule, whose inner nn.Sequential exposes the classifier head as item 4 (model.4.{weight,bias}). braindecode's EMG2QwertyNet exposes the same head as a named attribute (final_layer.{weight,bias}).

The remap is a two-key rename, applied once and saved here:

Upstream key	braindecode key
model.4.weight	final_layer.weight
model.4.bias	final_layer.bias

All 49 other keys (BatchNorm, MLP, TDS conv blocks) match verbatim — both modules expose the backbone as self.model = nn.Sequential(...), so the keys already share the model.<index>. prefix and need no rename. Weights are otherwise unchanged from upstream.

Conversion is reproducible from neuralbench-repo/scripts/convert_emg2qwerty_checkpoint.py.

Verification

BatchNorm running statistics on the first layer match upstream:

Stat	This checkpoint	Expected
model.0.batch_norm.running_mean.mean	0.511	≈ 0.51
model.0.batch_norm.running_var.mean	1.146	≈ 1.15

Forward pass on a 1×32×8000 random input returns shape (1, 373, 99) — the 4 s @ 2 kHz window after the TDS encoder + CTC head.

License

CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International), inherited from the upstream emg2qwerty release.

• Attribution (BY) — cite Sivakumar et al. (2024) and link this repository when the weights are used or redistributed.
• NonCommercial (NC) — not licensed for commercial use.
• ShareAlike (SA) — derivatives must be released under the same CC BY-NC-SA 4.0 license.

braindecode itself is BSD-3-Clause; that license applies to the code, not to these weights. The weights are governed solely by CC BY-NC-SA 4.0.

Intended use

Drop-in pretrained backbone for the emg/qwerty CTC keystroke-decoding task in NeuralBench, or any other research workflow consuming braindecode.models.EMG2QwertyNet.

Per the source paper (table 4): zero-shot val/CER ≈ 16 % on a held-out subject; further fine-tuning typically reduces CER to ≈ 10 % on a personalized split.

Citation

@inproceedings{sivakumar2024emg2qwerty,
  title     = {emg2qwerty: A Large Dataset with Baselines for
               Touch Typing using Surface Electromyography},
  author    = {Sivakumar, Viswanath and Seely, Jeffrey and Du,
               Alan and Bittner, Sean and Berenzweig, Adam and
               Bolarinwa, Anuoluwapo and Gramfort, Alexandre and
               Mandel, Michael},
  booktitle = {Advances in Neural Information Processing Systems
               (NeurIPS), Datasets and Benchmarks Track},
  year      = {2024},
  url       = {https://github.com/facebookresearch/emg2qwerty},
}