CTNet-HF trained on BNCI2014-001

This repository contains a trained Hugging Face Transformers implementation of CTNet, the convolutional Transformer architecture introduced by Zhao et al. (2024).

The checkpoint was trained as one pooled model on all BNCI2014-001 training sessions: session 0train from subjects 1-9. It expects 22 EEG channels, 1,000 samples at 250 Hz, and microvolt values in the channel order recorded in preprocessor_config.json.

• Implementation: https://github.com/bkozyrskiy/ctnet-hf
• Original CTNet code: https://github.com/snailpt/CTNet
• Dataset interface: MOABB BNCI2014_001

Important evaluation caveat

Session 1test was evaluated during training and used to select the checkpoint. The selected checkpoint is epoch 400, chosen by mean subject accuracy.

The scores below are therefore checkpoint-selection metrics on the nominal test sessions, not unbiased held-out test metrics.

Performance on BNCI2014-001 session 1test

Subject	Accuracy	Cohen's kappa
1	81.94%	0.7593
2	57.99%	0.4398
3	89.58%	0.8611
4	76.74%	0.6898
5	69.10%	0.5880
6	58.33%	0.4444
7	84.72%	0.7963
8	78.82%	0.7176
9	66.32%	0.5509
Mean subject	73.73%	0.6497

Pooled accuracy across all session 1test trials: 73.73%.

Training protocol

• Dataset: BNCI2014-001.
• Subjects: 1-9.
• Training data: all trials from session 0train, 2,592 total trials.
• Checkpoint-selection data: all trials from session 1test, 2,592 total trials.
• Input window: 1,000 samples, 250 Hz, approximately 4 seconds.
• Labels: feet, left_hand, right_hand, tongue.
• Normalization: one global Z-score fitted on training-session trials only.
• Architecture: CTNet paper configuration.
• Optimizer: Adam, learning rate 0.001, betas (0.5, 0.999).
• Batch size: 72.
• Epochs: 1,000.
• Augmentation: subject-aware segmentation-and-reconstruction, 8 segments, factor 3.
• Seed: 0.

See training_metadata.json and training_history.csv for the full run record.

Load

from transformers import AutoFeatureExtractor, AutoModelForSequenceClassification

repo_id = "likan-blk/ctnet-hf"

processor = AutoFeatureExtractor.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(
    repo_id,
    trust_remote_code=True,
).eval()

inputs = processor(eeg_trial, return_tensors="pt")  # eeg_trial: (22, 1000), microvolts
logits = model(**inputs).logits

Because this repository contains custom modeling and preprocessing code, inspect the files before enabling trust_remote_code. For reproducible use, pin revision to a Hub commit hash.

Intended use

This checkpoint is intended for research and reproducibility experiments with BNCI2014-001-style motor-imagery EEG. It is not a medical device, clinical tool, safety-critical classifier, or claim of generalization to arbitrary EEG recordings.