SincTran

An end-to-end transformer pipeline for imagined speech electroencephalogram classification.

SincTran combines learnable sinc bandpass filter banks (one per physiological EEG band) with a depthwise-separable CNN spatial/temporal encoder and a Transformer with a CLS token. Band contributions are weighted by a learned attention mechanism before spatial mixing, and the Transformer output at the CLS position is passed to an MLP classification head.

Table of Contents

• Repository Structure
• Datasets
• Installation
• Usage
• Configuration
• Reproducibility
• Citation

---

Repository Structure

SINCTRAN/
├── Code/
│   ├── core_model.py        # SincTran architecture + SincBandFilter + TransparentEncoderLayer
│   ├── core_dataset.py      # EEG preprocessing, ZCA whitening, stratified k-fold data loaders
│   ├── core_loaders.py      # Dataset loaders for ASU and BCI Competition datasets
│   ├── core_loss.py         # Cross-entropy loss wrapper
│   ├── core_train.py        # Training loop, evaluation, metric logging, confusion matrices
│   ├── core_utils.py        # Reproducibility, timing, FLOPs, metadata serialisation
│   │
│   ├── main_asu.py          # SincTran training pipeline — ASU dataset
│   ├── main_bci.py          # SincTran training pipeline — BCI Competition Track 3
│   ├── main_driver.py       # Entry point for main experiment runs
│   │
│   ├── sota_asu.py          # SOTA baseline training pipeline — ASU dataset
│   ├── sota_bci.py          # SOTA baseline training pipeline — BCI Competition Track 3
│   ├── sota_driver.py       # Entry point for SOTA benchmark runs
│   │
│   ├── int_module.py        # Interpretation utilities: band attention, attention rollout, GradCAM, UMAP
│   ├── int_asu.py           # Interpretation pipeline — ASU dataset
│   ├── int_bci.py           # Interpretation pipeline — BCI Competition Track 3
│   └── int_driver.py        # Entry point for interpretation runs
│
├── CITATION.cff
├── requirements.txt
└── README.md

All scripts live under Code/ and are intentionally flat within it — all imports resolve within the same directory with no relative-path dependencies. Run all driver scripts from inside Code/.

---

Datasets

Two imagined-speech EEG benchmarks are supported.

ASU Speech Imagery Dataset Four tasks (n1–n4) covering long words, short words, vowels, and mixed word pairs. Raw .mat files are expected under a single root directory; core_loaders.py resolves subfolder layout automatically. ASU Dataset

BCI Competition 2020 Track 3 Five-class imagined-speech dataset (15 subjects). Expects the standard competition folder layout with a Track3_clean.csv label file for the held-out test partition. BCI Dataset

---

Installation

Python 3.10 or later is recommended.

pip install -r requirements.txt

All imports use only the standard library plus the packages listed in requirements.txt. No package installation beyond this step is required.

---

Usage

Each experiment group has a dedicated driver script. Configure the run by editing the RUN_CFG dict at the bottom of the relevant driver, then execute it from inside the Code/ directory.

Main experiments (SincTran)

cd Code
python main_driver.py

Set DATA_PATH and RESULTS_ROOT at the top of main_asu.py or main_bci.py before running.

SOTA baselines

python sota_driver.py

Supported baselines: EEGNet, MSVTNet, ShallowFBCSPNet, CTNet, EEGConformer, ATCNet, EEGNeX (all via braindecode).

Interpretation

python int_driver.py

Requires trained weights saved by a main-experiment run. Set WEIGHTS_ROOT and RESULTS_ROOT in int_asu.py / int_bci.py.

Interpretation outputs per subject/fold:

Output file	Description
band_attn_norm_*.svg	Band-attention weights (L1-normalised), per class
band_attn_lin_*.svg	Raw linear band-attention scores, per class
cls_umap_*.svg	UMAP of CLS-token embeddings
mlp_umap_*.svg	UMAP of MLP first-layer activations
attn_rollout_*.svg	CLS-token attention rollout heatmap (Abnar & Zuidema 2020)
gradcam_*.svg	Spatial Grad-CAM heatmap

Fold-average versions of rollout and GradCAM figures are written automatically after the last fold.

---

Configuration

Each pipeline script (main_asu.py, main_bci.py, sota_asu.py, etc.) contains three self-contained config dicts — DATA_CFG, MODEL_CFG, and TRAIN_CFG — alongside a DATA_PATH / RESULTS_ROOT header block. No external config files are required; all hyperparameters are co-located with the pipeline that uses them.

Driver scripts expose a RUN_CFG dict for selecting dataset, task, subjects, fold count, and epoch budget without touching the pipeline files.

---

Reproducibility

core_utils.set_all_seeds fixes seeds for Python, NumPy, and PyTorch (including cudnn.deterministic). All pipelines call this before data loading. The default seed across all experiments is 37.

Normalisation and ZCA whitening statistics are always fitted on the training split only and applied to validation and test splits, with statistics recomputed per fold.

---

Citation

If you use SincTran in your research, please cite:

@misc{sinctran2026,
  author       = {Arun Balasubramanian and Santhoshkumar Peddi and Debasis Samanta},
  title        = {SincTran: An End-to-End Transformer Pipeline for Imagined Speech EEG Classification},
  year         = {2026},
  doi          = {10.2139/ssrn.6641648},
  url          = {https://github.com/ArunBsmn/SincTran},
  note         = {GitHub repository}
}