braindecode/plot_dance_event_detection

DANCE — BI2014a P300 event-detection checkpoint

Pretrained weights for the braindecode.models.DANCE model, used by the tutorial "From window labels to events: asynchronous EEG decoding with DANCE".

DANCE detects a set of (start, end, class) events directly from long, unaligned EEG windows, without being told where an event starts. This checkpoint was trained on the Brain Invaders BI2014a P300 dataset (flash detection: class 1 = non-target, class 2 = target, class 0 = background).

Results

Cross-subject, held-out subject 3 (never seen in training):

Metric	Value
F1-event (IoU > 0.5 + class match)	0.495
F1-sample (per-token macro)	0.372

On the first held-out window the model predicts 51 events (ground truth: 53), mean duration ≈ 1.1 s, with confidences 0.75–0.99 — i.e. sharp, correctly localized P300 flashes.

How to load

The architecture must be built exactly as in the tutorial (the state dict is saved with these hyper-parameters):

import torch
from huggingface_hub import hf_hub_download
from braindecode.models import DANCE

model = DANCE(
    n_outputs=3,
    n_chans=len(chs_info),   # 16 for BI2014a
    chs_info=chs_info,
    n_times=4096,            # 32 s @ 128 Hz
    sfreq=128.0,
    input_window_seconds=32.0,
)
model.load_state_dict(
    torch.load(hf_hub_download("braindecode/plot_dance_event_detection", "model.pt"))
)

Training

• Data: BI2014a, subjects 1, 2, 4–9 for training; subject 3 held out.
• Preprocessing (paper recipe): pick EEG, band-pass 0.1–100 Hz, resample to 128 Hz, per-channel robust scaling clamped to [-16, 16].
• Windows: 32 s fixed length, up to 150 events/window, num_latents = 256.
• Optimiser: Adam, constant lr 5e-4, batch size 16, 120 epochs, best-by-held-out-F1-event checkpoint.
• Loss: braindecode.training.DanceLoss (matched-only IoU normalization).

Reproduce with train_checkpoint.py in this repository.