--- license: apache-2.0 library_name: braindecode tags: - braindecode - pytorch - safetensors - ieeg - seeg - mapa datasets: - nm000253 --- # mapa-pretrained Weights of the MAPA encoder `mapa_vits384` (d_model 384, 12 blocks, 21,335,424 parameters), for [`braindecode.models.MAPA`](https://braindecode.org/stable/generated/braindecode.models.MAPA.html), converted from the authors' release. The classification head is not pretrained (seeded random init); fine-tune or linear-probe before use. ```python from braindecode.models import MAPA model = MAPA.from_pretrained("braindecode/mapa-pretrained", n_outputs=2, chs_info=raw.info["chs"], regions=regions) ``` Channel names are read as clinical contact labels (`"LA7"` is contact 7 of array `LA`); `regions` are DKT names from `braindecode.models.mapa.MAPA_DKT_REGIONS`. The montage in `config.json` (4 channels, no regions) is only a default. Input is expected at 2048 Hz, or as the session-normalized spectrogram with `normalization="session"`, `sfreq=32`. ## Source and conversion - Source: [bentang18/MAPA](https://huggingface.co/bentang18/MAPA) at revision `988efbf31a7d1f38533b848c993a719d6f900b1f`, file `mapa_vits384.pt` (sha256 `2d236089a2f1a3cc2827e3f150c4a2ba14c51bbfaf0ce0888f84b92a6eb25a7a`), Apache-2.0. The authors' `NOTICE` is copied in this repository. - `convert_mapa_checkpoint.py` (in this repository) renames the feed-forward `encoder.blocks.{i}.mlp.fc1`/`fc2` to `mlp.0`/`mlp.3` (braindecode's `FeedForwardBlock`), keeps every other key, and writes `config.json`, `model.safetensors` and `pytorch_model.bin` with `save_pretrained`. - The converted model's outputs equal braindecode's loading of the original file (max-abs difference 0.0). - Requires a braindecode version newer than 1.8.1. - The ablation checkpoints (`no_region`, `no_relpos`, `no_priors`) are not re-hosted; they remain at the source repository. ## Training data The data that shaped these weights. The ids are the `datasets:` field of this card, which braindecode reads into `model.card_data.datasets` when you call `from_pretrained` (MOABB class names for MOABB datasets, OpenNeuro / NEMAR accessions, Hub ids otherwise). `nm`/`on` ids are datasets on [NEMAR](https://nemar.org) (`on` = NEMAR's copy of the OpenNeuro dataset with the same number). Rows marked `private:` are data the authors did not release; they are not in `datasets:`. | `datasets:` id | Dataset | Stage | Modality | |---|---|---|---| | [`nm000253`](https://nemar.org/dataset/nm000253) | [Brain Treebank: large-scale intracranial recordings from naturalistic language stimuli (Wang et al., NeurIPS 2024 D&B; 10 subjects, 26 movie sessions, ~43.5 h sEEG)](https://braintreebank.dev/) | pretraining | iEEG | - masked autoencoding; 13 of 26 sessions from 7 subjects, 27.87 h; subjects 7 and 10 (and subject 5) excluded from pretraining; Neuroprobe Lite test sessions held out - Sources: arXiv:2609.13507 p.2; arXiv:2609.13507 Appendix A, Table 3; https://huggingface.co/bentang18/MAPA README, Training data. ## Citation ```bibtex @misc{tang2026pretraining, title = {Pretraining for Sample-Efficient Neural Interfaces}, author = {Ben Tang and Zachary Spalding and Gregory B. Cogan}, year = {2026}, eprint = {2609.13507}, archivePrefix = {arXiv}, primaryClass = {cs.LG}, url = {https://arxiv.org/abs/2609.13507}, } @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 Apache-2.0, as the original MAPA release. The checkpoint was pretrained on the Brain Treebank dataset (CC BY 4.0, https://braintreebank.dev/); see `NOTICE`.