neurorvq-tokenizer-ppg-pretrained

Weights of the NeuroRVQ-PPG tokenizer (four-scale encoder, residual vector quantizers, decoder) for braindecode.models.NeuroRVQTokenizer (modality="ppg"), converted from the authors' release. All weights are pretrained; model(x) returns the per-window z-scored target and reconstruction, model.tokenize(x) the codes.

from braindecode.models import NeuroRVQTokenizer
model = NeuroRVQTokenizer.from_pretrained("braindecode/neurorvq-tokenizer-ppg-pretrained", chs_info=raw.info["chs"])

Channel names must come from the model's PPG channel list (see the class docstring). The channels and window in config.json (1 PPG channel (ppg_c1), 960 samples at 100 Hz) are only a default; pass chs_info (or channel_names and n_chans) and n_times. Input is expected at that sampling rate, preprocessed as in the authors' examples; the model does not preprocess.

Source and conversion

• Source: ntinosbarmpas/NeuroRVQ at revision d944b87f44ae0ba2923b2f10d0518f23f6803b76, file pretrained_models/tokenizers/NeuroRVQ_PPG_tokenizer_v1.pt (sha256 8221b752e645d0819d02d81f036f90d434fe2bc34e708c78a263c480e1009e79), CC BY-NC 4.0. Code: KonstantinosBarmpas/NeuroRVQ.
• convert_neurorvq_checkpoint.py (in this repository) renames the Transformer feed-forward mlp.fc1/mlp.fc2 keys to braindecode's mlp.0/mlp.2, drops the pretraining-only tensors (mask_token, norm_pre, head_pre_*) and writes config.json, model.safetensors and pytorch_model.bin with save_pretrained.
• The converted model's outputs equal the authors' module loading the original file (max-abs difference 0.0 on random inputs).
• Requires a braindecode version newer than 1.8.1.

Citation

@misc{neurorvq,
  title         = {NeuroRVQ: Multi-Scale Biosignal Tokenization for Generative Foundation Models},
  author        = {Konstantinos Barmpas and Na Lee and Dimitrios Chalatsis and William Raftery and
                   Yannis Panagakis and Dimitrios A. Adamos and Nikolaos Laskaris and
                   Alexandros Koliousis and Dario Farina and Stefanos Zafeiriou},
  year          = {2026},
  eprint        = {2510.13068},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2510.13068},
}

@article{aristimunha2025braindecode,
  title   = {Braindecode: a deep learning library for raw electrophysiological data},
  author  = {Aristimunha, Bruno and others},
  journal = {Zenodo},
  year    = {2025},
  doi     = {10.5281/zenodo.17699192},
}

License

CC BY-NC 4.0 (Attribution-NonCommercial), as the original NeuroRVQ release: non-commercial use only, with attribution to Barmpas et al. and the source repository above.