This repository provides the official implementation of MVCNet: Multi-View Contrastive Network for Motor Imagery Classification.
MVCNet is a dual-branch framework that integrates multi-view data augmentation, CNN–Transformer parallel modeling, and supervised contrastive learning to improve representation learning and decoding generalizability of EEG-based MI Classification.
The codebase is organized as follows:
MVCNet/
│
├── MVCNet_CO.py # Main script for the Chronological Order (CO) scenario
├── MVCNet_CV.py # Main script for the Cross-Validation (CV) scenario
├── MVCNet_LOSO.py # Main script for the Leave-One-Subject-Out (LOSO) scenario
│
├── models/ # Implementations of MVCNet and baseline models
│ ├── IFNet.py
│ ├── Conformer.py
│ └── ...
│
├── data/ # datasets
│ ├── BNCI2014001
│ ├── Zhou2016
│ └── ...
│
├── utils/ # Utility functions
│ ├── data_augment.py # Data augmentation (e.g., time, frequency, spatial)
│ ├── contrastive_loss.py # Contrastive loss definitions
│ ├── network.py # encoder, decoder, etc
│ └── ...
│
└── README.md
MVCNet supports three standard MI decoding paradigms:
- CO (Chronological Order): Within-subject, time-based data split
- CV (Cross-Validation): Within-subject, stratified 5-fold validation. The data partitions were structured chronologically while maintaining class-balance, following FBCNet.
- LOSO (Leave-One-Subject-Out): Cross-subject generalization evaluation
Classification Accuracy (%) ± Std on Five MI Datasets under CO setting:
| Dataset | EEGNet | SCNN | DCNN | FBCNet | ADFCNN | EEGConformer | IFNet | MVCNet (Ours) |
|---|---|---|---|---|---|---|---|---|
| BNCI2014001 | 69.05 ± 1.00 | 73.57 ± 2.36 | 59.29 ± 1.64 | 68.97 ± 1.26 | 73.73 ± 2.26 | 78.57 ± 0.66 | 77.94 ± 0.93 | 83.17 ± 0.74 |
| Zhou2016 | 80.13 ± 3.35 | 75.03 ± 6.15 | 78.03 ± 2.37 | 63.33 ± 2.29 | 71.42 ± 1.95 | 73.87 ± 4.51 | 81.70 ± 2.08 | 84.11 ± 2.68 |
| Blankertz2007 | 78.79 ± 2.78 | 76.71 ± 2.09 | 70.00 ± 4.12 | 75.93 ± 1.31 | 76.07 ± 1.41 | 82.29 ± 2.62 | 84.00 ± 0.57 | 87.07 ± 0.52 |
| BNCI2014002 | 66.07 ± 2.76 | 79.07 ± 1.96 | 64.07 ± 2.70 | 69.50 ± 0.95 | 73.00 ± 1.95 | 76.21 ± 1.46 | 78.29 ± 1.68 | 81.29 ± 2.31 |
| BNCI2015001 | 75.58 ± 1.69 | 83.71 ± 1.34 | 71.08 ± 1.82 | 74.92 ± 0.97 | 78.75 ± 0.62 | 82.63 ± 0.54 | 83.83 ± 0.90 | 85.67 ± 0.55 |
| Average | 73.92 | 77.62 | 68.49 | 70.53 | 74.59 | 78.71 | 81.15 | 84.26 |
Classification Accuracy (%) ± Std on Five MI Datasets under CO Setting:
| Dataset | MVCNet | MVCNet (cvc) | MVCNet (cmc) |
|---|---|---|---|
| BNCI2014001 | 83.17 ± 0.74 | 82.70 ± 1.05 | 82.46 ± 0.46 |
| Zhou2016 | 84.11 ± 2.68 | 83.18 ± 3.07 | 82.22 ± 2.61 |
| Blankertz2007 | 87.07 ± 0.52 | 85.71 ± 1.30 | 86.21 ± 1.42 |
| BNCI2014002 | 81.29 ± 2.31 | 81.29 ± 1.82 | 81.21 ± 1.92 |
| BNCI2015001 | 85.67 ± 0.55 | 85.67 ± 0.08 | 85.42 ± 0.32 |
| Average | 84.26 | 83.71 | 83.51 |
- MVCNet: using both CVC and CMC contrastive modules.
- MVCNet (cvc): only using the cross-view contrasting module.
- MVCNet (cmc): only using the cross-model contrasting module.
For reproducibility, the preprocessed EEG dataset BNCI2014001 can be accessed at (put X.npy into /data/BNCI2014001/):
🔗 https://pan.baidu.com/s/19osNsaDnNliQTXxiK3ncOA (提取码: pdtg)
All datasets can be downloaded from the Mother Of All BCI Benchmarks (MOABB)
If you find this work helpful, please consider citing our paper:
@article{wang2025mvcnet,
title={MVCNet: Multi-View Contrastive Network for Motor Imagery Classification},
author={Ziwei Wang and Siyang Li and Xiaoqing Chen and Wei Li and Dongrui Wu},
journal={arXiv preprint arXiv:2502.17482},
year={2025}
}
Special thanks to Jiaheng for providing the source code of IFNet, which served as a valuable foundation for our implementation.
We appreciate your interest and patience. Feel free to raise issues or pull requests for questions or improvements.