---
library_name: braindecode
license: apache-2.0
tags:
- Brant
- braindecode
- ieeg
- seeg
- foundation-model
- model_hub_mixin
- pytorch_model_hub_mixin
---

# Brant — Foundation Model for Intracranial Neural Signals

Pretrained weights for [`braindecode.models.Brant`](https://braindecode.org/stable/generated/braindecode.models.Brant.html),
a faithful braindecode port of **Brant** (Zhang et al., NeurIPS 2023), a foundation
model for intracranial (sEEG/iEEG) recordings.

- Paper: [Brant: Foundation Model for Intracranial Neural Signal](https://proceedings.neurips.cc/paper_files/paper/2023/hash/535915d26859036410b0533804cee788-Abstract-Conference.html) (NeurIPS 2023)
- Original code & weights: [yzz673/Brant](https://github.com/yzz673/Brant) · [Daoze/Brant](https://huggingface.co/Daoze/Brant)
- braindecode docs: https://braindecode.org

## Provenance & license

These weights are the **official pretrained checkpoint** released by the original
authors (temporal + spatial encoders), converted into the braindecode `Brant`
module (state dict mapped 1:1; the two mask-token embeddings used only for the
masked-autoencoding pretraining objective are dropped). The original release is
under the **Apache-2.0** license, which this repository preserves.

> **Disclaimer (from the original authors).** The pre-training data for Brant was
> collected during routine treatment of epilepsy patients and is intended solely
> for medical or research use. These pre-trained weights are released only for
> medical or research purposes and must not be subjected to any form of misuse.

## Model configuration

This checkpoint uses the paper's configuration (~508M parameters):

| | |
|---|---|
| `patch_size` | 1500 (6 s at 250 Hz) |
| `embed_dim` | 2048 |
| `ffn_dim` | 3072 |
| `temporal_n_layers` | 12 |
| `spatial_n_layers` | 5 |
| `n_heads` | 16 |
| `n_freq_bands` | 8 |
| `n_times` | 22500 (15 patches, 90 s at 250 Hz) |
| `sfreq` | 250 Hz |

The signal is expected at **250 Hz**. The learnable temporal positional encoding
is fixed to 15 patches, so keep `n_times=22500`; you may freely change `n_chans`
(channels are pooled) and `n_outputs` (the classification head is task-specific
and randomly initialized — fine-tune it on your downstream task).

## Usage

```python
from braindecode.models import Brant

# Encoders load the pretrained weights; the classification head is (re)initialized
# for your task via n_outputs.
model = Brant.from_pretrained("braindecode/brant-pretrained", n_outputs=2)
```

## Citation

```bibtex
@inproceedings{zhang2023brant,
  title     = {Brant: Foundation Model for Intracranial Neural Signal},
  author    = {Zhang, Daoze and Yuan, Zhizhang and Yang, Yang and Chen, Junru and Wang, Jingjing and Li, Yafeng},
  booktitle = {Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS)},
  year      = {2023}
}
```