|
|
--- |
|
|
metadata: |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# Multi-Window EEG Models |
|
|
|
|
|
This repository hosts pre-trained PyTorch models (.pth files) for temporal analysis of EEG signals in object category decoding. The models are trained on the [Alljoined/05_125](https://huggingface.co/datasets/Alljoined/05_125) dataset, using COCO 2017 images as stimuli. Each model processes a specific time window post-stimulus onset to capture different stages of visual processing: |
|
|
|
|
|
- **EarlyVisual (50-150ms)**: Early visual features (e.g., edges, basic shapes). AUROC: ~0.59 |
|
|
- **MidFeature (150-250ms)**: Mid-level object parts (e.g., N170-like responses). AUROC: ~0.97 |
|
|
- **LateSemantic (250-350ms)**: Late semantic integration (e.g., N400/P300). AUROC: ~0.67 |
|
|
- **EarlyCombined (50-250ms)**: Combined early + mid processing. AUROC: ~0.97 |
|
|
- **FullWindow (50-350ms)**: Full baseline window. AUROC: ~0.97 |
|
|
|
|
|
These models use a hybrid CNN-Transformer architecture for multi-label classification over 38 COCO categories (animals, vehicles, food, outdoor objects). They detect weak category-specific signals in noisy EEG data. |
|
|
|
|
|
## Usage |
|
|
|
|
|
Load a model with PyTorch: |
|
|
|
|
|
```python |
|
|
import torch |
|
|
|
|
|
# Example: Load MidFeature model |
|
|
checkpoint = torch.load("model_150_250ms_MidFeature.pth", map_location="cpu") |
|
|
model = HybridCNNTransformer(n_timepoints=52) # Exact points from window |
|
|
model.load_state_dict(checkpoint["model_state_dict"]) |
|
|
model.eval() |
|
|
|
|
|
# Inference on EEG window (shape: [1, 64, n_timepoints]) |
|
|
logits = model(eeg_tensor) |
|
|
probs = torch.sigmoid(logits) |
|
|
top_categories = torch.topk(probs, k=20).indices |
