Gunin09/FEMBA-Tiny-TUAR-BC2

FEMBA-Tiny TUAR BC2 — binary artifact per window

Fine-tuned FEMBA-Tiny for TUH Artifact (TUAR) Goal A: predict whether a 5 s EEG window contains artifact activity (official Binary labels, mild class weights).

Files

File	Description
model.safetensors	FEMBA weights (load into FEMBA with classification_type=bc)
pytorch_lightning.ckpt	Original Lightning checkpoint
config.json	Architecture, normalization, deploy τ
best_tau_val.json	Val-tuned threshold metadata
metrics_summary_test.json	Test GATE-BC1 metrics

Deploy

• Threshold: deploy_tau = 0.31999999999999995 (val-constrained macro-F1 selection)
• Decision: P(ARTIFACT) >= deploy_tau → artifact
• Input: 22-channel bipolar, 5 s @ 200 Hz (1280 samples/ch), quantile norm (see config.json)

Test metrics (τ from val)

Metric	Value
AUROC	0.9033453832685784
AUPR	0.8756412861865515
macro_f1	0.8318813512280141
recall_artifact	0.8524488530688159

Training: mild weights [1.0, 1.2173], frozen encoder, 20 epochs, best val_BinaryAUROC at epoch 19.

Load (example)

from pathlib import Path
from safetensors.torch import load_file
# FEMBA class from BioFoundation repo
from models.FEMBA import FEMBA

model = FEMBA(
    seq_length=1280, num_channels=22, num_classes=2,
    num_blocks=2, embed_dim=35, classification_type="bc",
)
state = load_file("model.safetensors")
model.load_state_dict(state, strict=True)

See FembaTiny-TUAR for preprocessing, inference (run_inference_bc.py), and full result report.