Add MAPA mapa_vits384 converted from bentang18/MAPA@988efbf3 (Apache-2.0)

NOTICE

@@ -0,0 +1,64 @@
+MAPA
+Copyright 2026 Ben Tang, Zachary Spalding, and Gregory B. Cogan
+
+================================================================================
+What each license covers
+================================================================================
+
+Source code and released model checkpoints are licensed under the Apache License,
+Version 2.0. See LICENSE. The release attachment named LICENSE-WEIGHTS contains
+the same Apache 2.0 license for standalone checkpoint downloads. The checkpoint
+license covers:
+
+    mapa_vits384.pt
+    mapa_vits384_no_region.pt
+    mapa_vits384_no_relpos.pt
+    mapa_vits384_no_priors.pt
+
+================================================================================
+Attribution
+================================================================================
+
+The checkpoints were pretrained on the Brain Treebank dataset, released under
+CC BY 4.0 at https://braintreebank.dev/. The dataset retains its own license.
+Attribution for the pretraining data:
+
+    Christopher Wang, Adam Yaari, Aaditya K Singh, Vighnesh Subramaniam,
+    Dana Rosenfarb, Jan DeWitt, Pranav Misra, Joseph R Madsen, Scellig Stone,
+    Gabriel Kreiman, Boris Katz, Ignacio Cases, and Andrei Barbu.
+    Brain Treebank: Large-scale intracranial recordings from naturalistic
+    language stimuli. Advances in Neural Information Processing Systems 37
+    (NeurIPS 2024), Datasets and Benchmarks Track.
+
+The evaluation follows Neuroprobe, https://github.com/insight-neuro/neuroprobe,
+released under the MIT License, Copyright (c) 2025 Andrii Zahorodnii,
+collaborators and contributors.
+
+================================================================================
+Vendored Neuroprobe material
+================================================================================
+
+mapa/data/_neuroprobe_lite_tables.py includes tables copied from Neuroprobe
+commit c7b955b0a31464f4a5eec3f3bd78ff29841d61ac. The upstream license follows:
+
+MIT License
+
+Copyright (c) 2025 Andrii Zahorodnii, collaborators and contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,72 @@
+---
+license: apache-2.0
+library_name: braindecode
+tags:
+- braindecode
+- pytorch
+- safetensors
+- ieeg
+- seeg
+- mapa
+---
+
+# 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.
+
+## 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`.

config.json

@@ -0,0 +1,19 @@
+{
+  "n_outputs": 2,
+  "n_chans": 4,
+  "chs_info": null,
+  "n_times": 2048,
+  "input_window_seconds": null,
+  "sfreq": 2048,
+  "contact_labels": null,
+  "regions": null,
+  "d_model": 384,
+  "mlp_ratio": 4,
+  "region_embed": true,
+  "space_rope": true,
+  "deep_sup": true,
+  "pooling": "mean",
+  "normalization": "window",
+  "activation": "torch.nn.modules.activation.GELU",
+  "braindecode_version": "1.8.1"
+}

convert_mapa_checkpoint.py

@@ -0,0 +1,58 @@
+"""Convert the released MAPA checkpoint to a braindecode-native file.
+
+Source: https://huggingface.co/bentang18/MAPA at revision
+988efbf31a7d1f38533b848c993a719d6f900b1f (Apache-2.0), file ``mapa_vits384.pt``
+(sha256 2d236089a2f1a3cc2827e3f150c4a2ba14c51bbfaf0ce0888f84b92a6eb25a7a).
+
+Usage::
+
+    python convert_mapa_checkpoint.py OUT_DIR [SOURCE_FILE]
+
+writes ``OUT_DIR/mapa-pretrained`` with ``config.json``, ``model.safetensors``
+and ``pytorch_model.bin`` (``save_pretrained``). Without ``SOURCE_FILE`` the
+file is downloaded.
+
+Key changes: the feed-forward ``encoder.blocks.{i}.mlp.fc1``/``fc2`` become the
+``FeedForwardBlock`` children ``mlp.0``/``mlp.3``; every other key is kept. The
+classification head ``final_layer`` is not pretrained: it is a seeded random
+``nn.Linear`` default init. The stored montage (4 channels, no labels or
+regions) is only a default; pass ``chs_info`` or ``n_chans``, and
+``contact_labels`` and ``regions``, to ``from_pretrained``.
+"""
+
+import hashlib
+import sys
+from pathlib import Path
+
+import torch
+
+from braindecode.models import MAPA
+
+REPO, REVISION = "bentang18/MAPA", "988efbf31a7d1f38533b848c993a719d6f900b1f"
+FILENAME = "mapa_vits384.pt"
+SHA256 = "2d236089a2f1a3cc2827e3f150c4a2ba14c51bbfaf0ce0888f84b92a6eb25a7a"
+
+
+def convert(source, out):
+    if source is None:
+        from huggingface_hub import hf_hub_download
+
+        source = hf_hub_download(REPO, FILENAME, revision=REVISION)
+    assert hashlib.sha256(Path(source).read_bytes()).hexdigest() == SHA256
+    released = torch.load(source, map_location="cpu", weights_only=True)["model"]
+    state = {
+        key.replace(".mlp.fc1.", ".mlp.0.").replace(".mlp.fc2.", ".mlp.3."): value
+        for key, value in released.items()
+    }
+    torch.manual_seed(0)  # the head is a seeded random init
+    model = MAPA(n_outputs=2, n_chans=4, n_times=2048, sfreq=2048)
+    state.update(
+        {k: v for k, v in model.state_dict().items() if k.startswith("final_layer.")}
+    )
+    model.load_state_dict(state, strict=True)
+    model.save_pretrained(out)
+    return model
+
+
+if __name__ == "__main__":
+    convert(sys.argv[2] if len(sys.argv) > 2 else None, Path(sys.argv[1]) / "mapa-pretrained")

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed0626c405ef0b9f1759525ea67e18c45d8943d548e28bcfd0fb6679ae263611
+size 85369608

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8dfe964f048b7fdc54be199403d01349a626a80df3b29bba6a2773a824bcc35
+size 85411883