GeoNIRS-Base (pretrained encoder)
Weights of GeoNIRS-Base, the geometry-aware fNIRS foundation model of Li et al. (2026), for
braindecode.models.GeoNIRS.
They are the encoder of checkpoint/geonirs_base.pt of
GeoNIRS/GeoNIRS
at commit 81f8995, pretrained on 8 datasets (383.30 h), converted by convert.py.
from braindecode.models import GeoNIRS
model = GeoNIRS.from_pretrained(
"braindecode/geonirs-base-pretrained",
chs_info=raw.info["chs"],
n_outputs=2,
n_times=60,
)
features = model(x, return_features=True)["features"]
The classification head (final_layer) is not part of the release: it reads
2 * n_latent_channels * n_patches * feature_dim features, so from_pretrained builds it for
your n_times and n_outputs. Fine-tune or linear-probe before use.
Intended input
The inputs of the GeoNIRS paper (Appendix B.1): HbO and HbR in µmol/L resampled to 5 Hz, each
chromophore z-scored per recording over all its valid channels and time points, in windows of at
most 150 samples (30 s, the 15 rows of the temporal position embedding). chs_info gives each
HbO/HbR pair the positions of its source and detector (loc[3:9], where MNE stores them); a NaN
marks a missing sample.
Changes from the original release
convert.py unwraps the {"model", "config"} payload, builds the model from its config, drops
the two pretraining decoders (hbo_to_hbr_decoder, hbr_to_hbo_decoder: 122 of the 333
tensors), renames 6 tensors of the channel compressor and writes the encoder with save_pretrained.
The renames are plain prefix replacements: braindecode builds the compressor's feed-forward block from
a LayerNorm (ffn_norm) and its shared FeedForwardBlock (Linear, activation, dropout, Linear).
| original key | this repository | keys |
|---|---|---|
channel_compressor.ffn.0.{weight,bias} (LayerNorm) |
channel_compressor.ffn_norm.{weight,bias} |
2 |
channel_compressor.ffn.1.{weight,bias} (Linear) |
channel_compressor.ffn.0.{weight,bias} |
2 |
channel_compressor.ffn.4.{weight,bias} (Linear) |
channel_compressor.ffn.3.{weight,bias} |
2 |
The other 205 keys keep their names. config.json lists projection_dims=[128, 96], the inner
widths of the chromophore projections: they run over (embed_dim, *projection_dims, feature_dim), which is
the projection_dimensions [192, 128, 96, 64] of the original config. It also lists, at their defaults, the
numbers the original code fixes: geometry_mlp_ratio and compressor_mlp_ratio (2.0), geometry_scale
(0.02546058255683924, that is 25.46 mm), time_embedding_init_std and latent_init_std (0.02) and
residual_init_scale (0.1).
No value is changed: each of the 211 tensors is bitwise equal to its tensor in the original file (the six
renamed ones under their new names). This layout replaces revision a483f904, which holds the same
tensors under the authors' names.
Verification
GeoNIRS.from_pretrained on these files (braindecode 32850d0b) against braindecode's GeoNIRS on
revision a483f904 (master before the renames, 252a0172), on CPU and the same seeded inputs: the pinned
input, a batch with NaN samples and a ch_mask, pool="mean_std" at 60 samples, MNE-ordered chs_info
and an explicit geometry. 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 1e-5 or more in the features.
Evaluation
Linear probe of the frozen encoder on DS42 (action mode, ds005935) and DS19 (pain intensity,
ds005777), with the folds and seeds of the paper. "Authors' pipeline" is GeoNIRS@81f8995 run end
to end; "braindecode" is GeoNIRS with a re-implementation of the authors' probe. Both build the
same windows (1628 and 1190), on which GeoNIRS gives the authors' encoder features exactly.
| Task | Metric | Paper | Authors' pipeline | braindecode |
|---|---|---|---|---|
| DS42 | macro-F1 | 0.7018 | 0.7000 | 0.7039 |
| DS42 | AUROC | 0.7840 | 0.7840 | 0.7840 |
| DS19 | macro-F1 | 0.5598 | 0.5599 | 0.5304 |
| DS19 | AUROC | 0.5683 | 0.5687 | 0.5549 |
License and citation
The source repository has no licence file; the weights are by the GeoNIRS authors. Please cite:
@article{li2026geonirs,
author = {Li, Zongsheng and Zhu, Shengyuan and Zhang, Zimo and Li, Hanyao and Wang, Shengming and Liu, Quanying},
title = {{GeoNIRS}: Geometry-Aware Pretraining for Heterogeneous {fNIRS} Representation Learning},
journal = {arXiv preprint arXiv:2610.17845},
year = {2026},
}
Conversion
convert.py (in this repository) was run with braindecode at 32850d0b:
python convert.py geonirs_base.pt <out_dir>
| file | sha256 |
|---|---|
model.safetensors |
b60b1a7851ee762881ca4d0467a52c4383c9fa38642e28e0138d541df50dbb63 |
config.json |
d3254e64e487566cd1f33fc3fe833973ddd08201066ca11e6329d1ea4ba7e0e5 |
convert.py |
6e0e62f7f26b00143fed34aef3e6f95961598971df0fa66013654cf55747f8a5 |
original checkpoint/geonirs_base.pt (GeoNIRS/GeoNIRS@81f8995, Git LFS) |
61e4d72221dd99225ac405124515c070a2ef0467df83a01f05e918e0941cd06a |
- Downloads last month
- 35