---
language: en
license: mit
tags:
  - eeg
  - foundation-model
  - BCI
  - neuroscience
  - burn
  - rust
  - safetensors
base_model: eegdino/EEG-DINO
library_name: burn
pipeline_tag: feature-extraction
---

# EEG-DINO (safetensors, Burn-compatible)

Converted weights for the [EEG-DINO](https://github.com/miraclefish/EEG-DINO) foundation model, packaged as safetensors for use with the [`eegdino`](https://crates.io/crates/eegdino) Rust inference crate.

**Derived from:** [eegdino/EEG-DINO](https://huggingface.co/eegdino/EEG-DINO)

## Files

| File | Model | Params | d_model | Heads | Layers | Size |
|------|-------|--------|---------|-------|--------|------|
| `eeg_dino_small.safetensors` | Small | 4.6 M | 200 | 8 | 12 | 17 MB |
| `eeg_dino_medium.safetensors` | Medium | 33 M | 512 | 16 | 16 | 129 MB |
| `eeg_dino_large.safetensors` | Large | 201 M | 1 024 | 16 | 24 | 770 MB |

## What changed from the original weights

The original PyTorch `.pt` checkpoints from [eegdino/EEG-DINO](https://huggingface.co/eegdino/EEG-DINO) were converted with the following transformations:

1. **Filtered** --- only student encoder weights are kept; teacher, projector, and loss tensors are discarded
2. **Renamed** --- `module.student.` prefix stripped; Sequential indices mapped to descriptive names (e.g. `proj_in.0` → `proj_in.conv1`)
3. **Transposed** --- linear weight matrices converted from PyTorch `[out, in]` to Burn `[in, out]` layout
4. **Format** --- saved as float32 safetensors (no bf16)

The conversion script is at [`scripts/convert_weights.py`](https://github.com/eugenehp/eegdino-rs/blob/main/scripts/convert_weights.py).

## Usage with Rust

```toml
# Cargo.toml
[dependencies]
eegdino = "0.1"
burn = { version = "0.20", features = ["ndarray"] }
```

```rust
use eegdino_rs::prelude::*;
use burn::backend::NdArray;

type B = NdArray;

let encoder = EegDinoEncoder::<B>::builder()
    .weights("eeg_dino_small.safetensors")
    .device(Default::default())
    .build()
    .unwrap();

// 19 channels, 10 seconds @ 200 Hz
let signal = vec![0.0f32; 19 * 2000];
let result = encoder.encode_raw(&signal, 1, 19, 2000).unwrap();
// result.shape == [1, 191, 200]
```

## Download

```bash
# All weights
hf download eugenehp/eegdino --local-dir weights

# Single model
hf download eugenehp/eegdino eeg_dino_small.safetensors --local-dir weights
```

## Numerical parity

These converted weights produce outputs with NRMSE < 1e-6 compared to the original PyTorch model on identical inputs:

| Model | Max abs error | NRMSE |
|-------|---------------|-------|
| Small | 8.5e-7 | 5.5e-7 |
| Medium | 2.1e-6 | 8.8e-7 |
| Large | 4.8e-6 | 5.9e-7 |

## Citation

If you use these weights, please cite the original EEG-DINO paper:

```bibtex
@article{jiang2024eegdino,
    title={Learning EEG Foundation Models via Hierarchical Self-Distillation},
    author={Jiang, Yuqi and others},
    year={2024}
}
```

## Links

- **Original model:** [eegdino/EEG-DINO](https://huggingface.co/eegdino/EEG-DINO)
- **Original code:** [github.com/miraclefish/EEG-DINO](https://github.com/miraclefish/EEG-DINO)
- **Rust crate:** [crates.io/crates/eegdino](https://crates.io/crates/eegdino)
- **Rust repo:** [github.com/eugenehp/eegdino-rs](https://github.com/eugenehp/eegdino-rs)

## License

MIT (conversion and crate). Original model weights are subject to the [EEG-DINO license](https://huggingface.co/eegdino/EEG-DINO).