NeuroPose-emg2pose

NeuroPose — emg2pose regression baseline

Figure 6 of Liu et al. (2021), CC BY.

Meta's released regression_neuropose.ckpt from the emg2pose benchmark, rehosted with braindecode's own parameter names so it loads directly:

from braindecode.models import NeuroPose

model = NeuroPose.from_pretrained("braindecode/NeuroPose-emg2pose")

16-channel sEMG at 2 kHz in, 20 joint angles per sample out — (batch, n_chans, n_times) to (batch, n_times, n_outputs).

Provenance

These are not retrained weights. They are the authors' released checkpoint with only the parameter names rewritten: upstream stores every block in a single nn.Sequential, while braindecode splits them across encoder / resnet / decoder. Loading the original .ckpt into braindecode.models.NeuroPose and running both implementations on the same input gives bit-identical outputs (max absolute difference 0.0, 6,354,903 parameters in both).

The Hub checkpoint uses Braindecode-native parameter names; upstream naming is handled once during publication rather than by the runtime model.

Architecture

Liu et al.'s encoder / ResNet / decoder, in the configuration emg2pose used for its published baseline: encoder widths 32/128/256 pooling (10,2) (8,2) (4,4) over (time x electrodes), five residual blocks of three conv groups, decoder 128/32/1 upsampling (10,4) (8,4) (4,2).

License and attribution

Weights are Meta's, released under CC BY-NC-SA 4.0 — non-commercial, share-alike. That license carries over here unchanged. The UmeTrack hand model used for the forward kinematics behind the labels is CC BY-NC 4.0.

• Salter et al. (2024), emg2pose: A Large and Diverse Benchmark for Surface Electromyographic Hand Pose Estimation, NeurIPS Datasets and Benchmarks. arXiv:2412.02725
• Liu, Zhang, Gowda (2021), NeuroPose: 3D Hand Pose Tracking using EMG Wearables, The Web Conference 2021. doi:10.1145/3442381.3449890