tinymyo-epn612 / README.md
bruAristimunha's picture
TinyMyo converted from PulpBio/TinyMyo at bc14906
ec76952 verified
|
Raw History Blame Contribute Delete
2.18 kB
metadata
license: cc-by-nd-4.0
library_name: braindecode
base_model: PulpBio/TinyMyo
tags:
  - braindecode
  - pytorch
  - safetensors
  - emg
  - semg
  - tinymyo
  - foundation-model

TinyMyo (epn612)

Weights of TinyMyo (Fasulo et al., 2026), the encoder fine-tuned for EPN-612 gesture classification (8 channels, 6 classes), for braindecode.models.TinyMyo. They are EPN612/EPN_finetune_5sec.safetensors of PulpBio/TinyMyo (revision bc14906), converted to braindecode's parameter names by convert.py.

from braindecode.models import TinyMyo

model = TinyMyo.from_pretrained("braindecode/tinymyo-epn612", n_chans=8, n_outputs=6, n_times=1000)

The 6-class head is the authors' fine-tuned classifier.

Intended input

Surface EMG windows of 1000 samples (n_times=1000). The authors pretrain on 20-450 Hz band-passed signals with a 50 Hz notch, min-max normalized per channel; in fine-tuning they band-pass at 20-90 Hz and z-score each channel.

Changes from the original release

convert.py changes the format only:

  • the model. prefix stripped and the parameter names mapped to braindecode's;
  • the patch projection Conv2d(1, D, (1, patch_size)) stored as the equivalent Linear(patch_size, D) weight;
  • the classifier renamed from model_head.classifier to final_layer;
  • the pretraining-only mask token, reconstruction decoder and regressor not included.

No value is changed: braindecode's TinyMyo with these weights gives the outputs of the original file bit for bit on CPU in float32.

License and citation

CC BY-ND 4.0, as the original release; the weights are by the TinyMyo authors (ETH Zurich, PULP platform). Please cite:

@article{fasulo2026tinymyo,
  author  = {Fasulo, Matteo and Spacone, Giusy and Ingolfsson, Thorir Mar and Li, Yawei and Benini, Luca and Cossettini, Andrea},
  title   = {{TinyMyo}: a Tiny Foundation Model for Flexible {EMG} Signal Processing at the Edge},
  journal = {IEEE Transactions on Biomedical Engineering},
  year    = {2026},
  doi     = {10.1109/TBME.2026.3736157},
}