🧠 ConvoReleNet – Hybrid Deep Network for Motor Imagery EEG Classification

This repository contains the official implementation of ConvoReleNet, a hybrid deep neural network combining Convolutional, Transformer, and Recurrent (BiLSTM) modules for Motor Imagery EEG (MI-EEG) classification.

The code accompanies the paper:

ConvoReleNet: A Multi-Branch CNN–Transformer–LSTM Hybrid Network with Transfer Learning for Motor Imagery EEG Decoding
(Frontiers in Neuroscience, 2025)

---

🌟 Overview

ConvoReleNet is designed to capture spatial, temporal, and contextual dependencies in EEG signals.
It integrates three complementary modules:

• Convolutional layers → learn localized spectral–spatial features.
• Transformer encoder → models long-range temporal dependencies via self-attention.
• BiLSTM → refines sequential patterns and stabilizes temporal learning.

The network supports:

• Training from scratch or transfer learning between datasets.
• Subject-wise cross-validation (LOSO-like) evaluation.
• Multiple activation variants (ELU, ReLU, Tanh).

---

🧩 Environment Setup

Tested environment:

Library	Version
Python	3.8
PyTorch	1.10
NumPy	≥1.21
SciPy	≥1.7
scikit-learn	≥1.0
MNE	≥0.24
matplotlib	≥3.4

Installation

pip install torch==1.10 numpy scipy scikit-learn mne matplotlib

You can also use a virtual environment:

python3 -m venv convoenv
source convoenv/bin/activate      # Linux / macOS
convoenv\Scripts\activate         # Windows
pip install torch==1.10 numpy scipy scikit-learn mne matplotlib

---

📦 Dataset Preparation

This project uses publicly available Motor Imagery EEG datasets.
Download any or all of the following:

Dataset	Description	Download Link
BCI Competition IV – 2a	4-class MI EEG (22 channels)	Download
BCI Competition IV – 2b	2-class MI EEG (3 channels)	Download
Weibo 2014 Dataset	MI EEG benchmark	Download
PhysioNet EEG Motor Movement/Imagery	Open EEG dataset	Download

After downloading, organize your folder as follows:

/datasets/
   ├── BNCI_IV_2a/
   ├── BNCI_IV_2b/
   ├── Weibo2014/
   └── PhysioNet_MMI/

The code automatically detects datasets from this directory.

---

⚙️ Running the Model

1️⃣ Train From Scratch

python ConvoReleNet_MIEEG.py --dataset IV2a --mode scratch

2️⃣ Transfer Learning (Cross-Dataset)

python ConvoReleNet_MIEEG.py --pretrained IV2b --finetune IV2a --mode transfer

3️⃣ Subject-Wise Cross-Validation

python ConvoReleNet_MIEEG.py --dataset IV2a --cv subjectwise

Add --device cpu if you do not have a GPU.

---

📊 Output

After training, the script generates:

• Accuracy, F1-score, and Cohen’s κ (per subject)
• Confusion matrices
• Training and validation curves
• CSV summary files and plots in the results/ directory

---

🧠 Notes for Beginners

• You do not need prior EEG experience. Follow the steps above step-by-step.
• The model and preprocessing pipelines are fully implemented in PyTorch.
• Training may take ~15–30 min on GPU, or longer on CPU.
• You can change hyperparameters in the script directly (e.g., number of filters, heads, layers).

---

🧾 Citation

If you use this code or derivative work, you must cite the associated paper:

@article{ConvoReleNet2025,
  title={ConvoReleNet: A Multi-Branch CNN–Transformer–LSTM Hybrid Network with Transfer Learning for Motor Imagery EEG Decoding},
  author={Your Name and Co-authors},
  journal={Frontiers in Neuroscience},
  year={2025}
}

---

⚖️ License

This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).

You are free to use, share, and modify the code as long as you give appropriate credit and cite the original publication.

Example acknowledgment:

“This research used the ConvoReleNet framework (Kyzyrkanov et al., 2025) available under CC BY 4.0.”

---

📫 Contact

Author: Your Name
Email: your.email@example.com
Repository: https://github.com/YourUsername/ConvoReleNet-MI-EEG

For questions or collaboration, feel free to open an issue or reach out via email.

---

✅ Summary

• Python 3.8 / Torch 1.10
• Download EEG datasets (links above)
• Run ConvoReleNet_MIEEG.py to reproduce results
• Cite the paper if you use the code

Enjoy experimenting with ConvoReleNet!