EMG-GPT

Inference weights for EMG-GPT: Predictive Pretraining on Residual-Quantized EMG Tokens for Hand Pose Estimation.

Paper · Code · Inference guide

Ettore Magni · Rolandos Alexandros Potamias · Stefanos Zafeiriou · Konstantinos Barmpas

The selected GPT checkpoints were pretrained with CUDA on an NVIDIA GH200 120 GB GPU. The inference package accepts CPU or CUDA.

Checkpoints

Directory	Task	GPT initialization	Selected pose checkpoint
regression/	Pose estimation without an initial pose	Step 280,000	Pose warm-start → full fine-tuning, step 2,000
tracking/	Pose estimation with a boundary pose per window	Step 400,000	Pose warm-start → full fine-tuning, step 5,000

Each directory contains config.json, model.safetensors, codebooks.safetensors and manifest.json: the adapted GPT backbone, pose head and codebooks needed for inference. The loader verifies their hashes.

The shared tokenizer is downloaded separately from ntinosbarmpas/NeuroRVQ at revision d944b87f44ae0ba2923b2f10d0518f23f6803b76. The helper below downloads and verifies pretrained_models/tokenizers/NeuroRVQ_EMG_tokenizer_v1.pt.

Regression quick start

Use Python 3.11 or newer (tested on 3.11–3.14). Clone the code repository and install '.[download]' following its README, then run:

hf download ettoremagni/EMG-GPT \
  --revision 812d159b4e7a4fb1c95da865f4f1e2635fa6522f \
  --include "regression/*" --include "LICENSE" --local-dir weights
emg-gpt-download-tokenizer --output weights/NeuroRVQ_EMG_tokenizer_v1.pt
emg-gpt-predict \
  --model-dir weights/regression \
  --tokenizer weights/NeuroRVQ_EMG_tokenizer_v1.pt \
  --input recording.npz --output prediction.npz --device cpu

Use --device cuda with a CUDA-enabled PyTorch installation for NVIDIA GPU inference. MPS is unsupported. Use the current release from the code repository; its guide includes a runnable synthetic smoke check.

Input NPZ files need raw 2-kHz emg ([samples,16]) in the native emg2pose amplitude scale, scalar sampling_rate_hz=2000 and Unicode channel_names (c1 through c16, in column order). A full window needs at least 13,119 samples. Predictions are joint angles in radians, [windows,250,20] at 50 Hz, with timestamps and a coverage mask. For Tracking, replace regression/* with tracking/*, keeping LICENSE included. Supply one boundary pose per window; an all-NaN row explicitly skips it, retaining timestamps with NaN angles and no coverage. The inference guide covers HDF5 input, the Python API and alignment.

Scope and verification

This is offline inference: the frontend filters the entire recording and resamples to 1 kHz before tokenization. Outputs exclude warm-up, inter-window gaps and a trailing remainder. Other sensors or amplitude scales are unvalidated.

The code repository includes integration tests using these weights and synthetic reference outputs from the original research implementation. CPU compatibility has also been checked on real recordings. These checks are separate from the paper's benchmark; CUDA numerical parity remains unverified. See the inference guide for the test commands.

Citation

@misc{magni2026emggpt,
  title={EMG-GPT: Predictive Pretraining on Residual-Quantized EMG Tokens for Hand Pose Estimation},
  author={Ettore Magni and Rolandos Alexandros Potamias and Stefanos Zafeiriou and Konstantinos Barmpas},
  year={2026},
  eprint={2610.05235},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2610.05235}
}

License and attribution

EMG-GPT learned weights (regression/model.safetensors and tracking/model.safetensors) are licensed under CC BY-NC-SA 4.0. Attribute the EMG-GPT authors and paper cited above.

The NeuroRVQ tokenizer and the bundled codebooks.safetensors are separate upstream assets and retain their CC BY-NC 4.0 terms; the EMG-GPT grant does not relicense them. The inference code also remains CC BY-NC 4.0. See the third-party notices for provenance and component licenses.