BaRISTA, parcels scale
The released BaRISTA encoder of Oganesian, Hashemi and Shanechi
(2025) for the parcels spatial scale,
converted to
braindecode.models.BaRISTA.
Source: pretrained_models/parcels_chans.ckpt at revision
83b27375eba60e9eba9da4e7dd8fb283baace376,
sha256 6c234517d286a8e710b09716dc88c713618670df523cfffb89e4c9073f2657c1. Conversion renames tensors and fuses the released gated
projections into the single projection this port uses; convert_barista_weights.py
in this repository reproduces it. Encoder tokens match the released forward
equations to 3.6e-06 in float32 on CPU.
Usage
Supply the montage of each batch as spatial_indices, an (n_chans,) tensor of Destrieux parcels in [0, 121):
import torch
from braindecode.models import BaRISTA
model = BaRISTA.from_pretrained("braindecode/BaRISTA-parcels", n_chans=64, n_outputs=2)
logits = model(torch.randn(8, 64, 6144), spatial_indices=torch.randint(0, 121, (64,)))
The saved geometry is 64 channels and 6144 samples at 2048 Hz, the
pretraining window. Mean pooling makes the encoder independent of both, so pass
your own n_chans, n_times and n_outputs when loading.
Limitations
These files hold the encoder only, as does the release. Braindecode
initializes the classification head on load, so it needs fine-tuning, as does
the learned read-out of the paper's protocol (pooling="learned"), which was
never released. The check
above covers float32 CPU encoder tokens, not downstream accuracy, GPU kernels or
mixed precision.
Citation
@inproceedings{oganesian2025barista,
title={BaRISTA: Brain Scale Informed Spatiotemporal Representation of Human Intracranial Neural Activity},
author={Oganesian, Lucine L. and Hashemi, Saba and Shanechi, Maryam M.},
booktitle={Advances in Neural Information Processing Systems},
year={2025}
}
License
Copyright (c) 2025 University of Southern California. Educational, research and
non-profit use only; commercial use requires an agreement with the USC Stevens
Center for Innovation. See NOTICE.txt and the original licence.