MtReNIRS pretrained encoder
The pretrained MtReNIRS encoder by Chae et al. (2026), for
braindecode.models.MtReNIRS.
model.safetensors and config.json hold the encoder of mtrenirs_10s.pt, which is
models/weights/mtrenirs_10s.pt of microsoft/MtReNIRS
at commit 16c3082,
converted by convert.py.
from braindecode.models import MtReNIRS
# raw: MNE Raw of HbO/HbR at 10 Hz with the Standard 10-05 slot positions in loc[:3]
model = MtReNIRS.from_pretrained(
"braindecode/mtrenirs-pretrained",
chs_info=raw.info["chs"],
n_outputs=3,
n_times=100,
)
features = model(x, return_features=True)["features"]
The classification head (final_layer) is not part of the release: its input width depends on
the montage, so from_pretrained builds it for your chs_info and n_outputs. Fine-tune or
linear-probe before use.
Input the weights expect
10 Hz HbO/HbR in 100-sample (10 s) windows, each channel z-scored over its run and clipped at 6, with channel positions in Standard 10-05 space. The authors' model card describes the intended and out-of-scope uses.
Changes from the original release
convert.py unwraps the checkpoint's {"config", "model"} payload, renames 6 keys of the token
embedding and writes the encoder with save_pretrained. The renames are plain prefix replacements:
braindecode names the two layers directly instead of wrapping each in a one-child Sequential.
| original key | this repository | keys |
|---|---|---|
embed.signal.proj.{weight,bias} |
embed.signal_proj.{weight,bias} |
2 |
embed.spatial.mlp.{0,2}.{weight,bias} |
embed.spatial_mlp.{0,2}.{weight,bias} |
4 |
The other 98 keys keep their names. config.json no longer lists channel_strategy and
channel_strategy_kwargs, which MtReNIRS has not taken since braindecode#1688.
No value is changed: each of the 104 tensors is bitwise equal to its tensor in mtrenirs_10s.pt (the six
renamed ones under their new names). The original file, mtrenirs_10s.pt (sha256
0960d02f9437d86f0f1a670eace8905d58b33f3394e58d5c17a6451bbcb9484d), is kept in this repository unchanged;
revision bbc68ce5 is the mirror without the conversion, and revision 7c5e6053 holds the
same tensors under the authors' names.
Verification
MtReNIRS.from_pretrained on these files (braindecode 32850d0b) against braindecode's MtReNIRS on
revision 7c5e6053 (master before the renames, 252a0172), on CPU and the same seeded inputs:
the pinned input at 100 and at 300 samples, pool="channel_temporal_mean" with a NaN window and a
ch_mask, pool="flatten" and explicit coordinates. In float32 and in float64 the features, the
outputs (same head), the tokens and their validity, and the features and outputs in train mode under one
seed (the dropout draws) are bitwise equal: 60 of 60 comparisons, max abs difference 0.0. A change of 1e-3
in one value of each renamed tensor shows as a difference of 4e-4 or more in the features.
Licence
MIT, Copyright (c) Microsoft Corporation. LICENSE.txt is the source repository's licence file,
unchanged.
Citation
@article{microsoft2026mtrenirs,
author = {Chae, Tony and Avramidis, Kleanthis and Emmanouilidou, Dimitra and Tashev, Ivan J.},
title = {{MtReNIRS}: A Self-Supervised Foundation Model for Transferable Neurovascular Representations},
journal = {arXiv preprint arXiv:2610.36159},
year = {2026},
}
Conversion
convert.py (in this repository) was run with braindecode at 32850d0b:
python convert.py mtrenirs_10s.pt <out_dir>
| file | sha256 |
|---|---|
model.safetensors |
e73ec45447b21213b7dfa861ac80ebf7b6ac362a115e3b465bfde054e22198e4 |
config.json |
29d2bc9b2f67fb495f43f1e14a33daa4b47f327f43f78b506015e7b68151e5f4 |
convert.py |
7913624b7c276a4dca45bc39a638552fdca259a47227d0ebf595085cdb127bfe |
original models/weights/mtrenirs_10s.pt (microsoft/MtReNIRS@16c3082) |
0960d02f9437d86f0f1a670eace8905d58b33f3394e58d5c17a6451bbcb9484d |
- Downloads last month
- 35