bruAristimunha's picture
Pratibha2 encoder converted from MannasAI/pratibha-2 at 863edc9
dea9139 verified
|
Raw History Blame Contribute Delete
3.13 kB
---
license: apache-2.0
library_name: braindecode
base_model: MannasAI/pratibha-2
tags:
- braindecode
- pytorch
- safetensors
- eeg
- pratibha-2
- manas-2
- masked-autoencoder
- foundation-model
datasets:
- TUH
- I-CARE
---
# PRATIBHA-2 (pretrained encoder)
Weights of PRATIBHA-2, formerly MANAS-2 (Kulkarni et al., 2026), a dense-attention EEG
foundation model pretrained with Raw-Band Hybrid masked autoencoding and Constrained
Reconstruction, for
[`braindecode.models.Pratibha2`](https://braindecode.org/stable/generated/braindecode.models.Pratibha2.html).
They are the encoder of [MannasAI/pratibha-2](https://huggingface.co/MannasAI/pratibha-2)
(revision `863edc9`), converted to braindecode's parameter names by `convert.py`. The
class is added in braindecode 2.1.
```python
from braindecode.models import Pratibha2
model = Pratibha2.from_pretrained(
"braindecode/pratibha2-pretrained",
chs_info=raw.info["chs"], # any montage the position table places
n_outputs=4,
n_times=1000,
)
features = model(x, return_features=True)["features"] # (batch, 512)
```
The classification head (`final_layer`, `n_outputs=2` in `config.json`) is not part of the
release: it is PyTorch's default init under `torch.manual_seed(0)`. Passing `n_outputs` to
`from_pretrained` replaces it. Fine-tune or linear-probe before use. `config.json` stores a
19-channel 10-20 montage; pass your own `chs_info`.
## Intended input
- EEG sampled at **200 Hz**, windows of at least 200 samples (1 s patches, 0.1 s overlap).
- Channels are placed by name in the pretraining electrode table (REVE's position bank
plus PRATIBHA-2's own `AFZ` and `IZ`); aliases such as `T3`, `EEG Fp1-REF` and bipolar
`A-B` pairs are resolved by braindecode.
- Each channel is z-scored within each window inside the model, so volts and microvolts
give the same output.
## Training data
Unlabeled clinical EEG from TUH EEG, I-CARE and the authors' internal sources (private:
not released, not in `datasets:`), as described in the original model card and the paper.
## Changes from the original release
Converted by `convert.py` from `model.safetensors` of MannasAI/pratibha-2 at revision
`863edc940d82a02744776a7820641bf1bc32310b`:
- parameter names renamed to braindecode's (`encoder.layers.N.ffn` ->
`encoder.blocks.N.mlp`, `attn.qkv_proj` -> `attn.qkv`, and so on);
- each fused feed-forward input projection `ffn.in_proj` split into its value rows
(`mlp.fc1`) and gate rows (`mlp.fc_gate`);
- the constant `pos_enc.freq_matrix` dropped (the model rebuilds it).
No value was changed: braindecode's `Pratibha2` with these weights gives the features and
tokens of the original weights bit for bit on CPU in float32.
## License and citation
Apache-2.0, as the original release (see `LICENSE`). Weights by MannasAI; please cite:
```bibtex
@article{kulkarni2026manas2,
title = {MANAS-2: Constrained Reconstruction for EEG Foundation Models},
author = {Kulkarni, A. and Mishra, A. R. and Jain, M. and Runwal, P. and Saini, L. and Panwar, S. and Singh, S.},
journal = {arXiv preprint arXiv:2609.13717},
year = {2026}
}
```