🧠 DSAINet

An Efficient Dual-Scale Attentive Interaction Network for General EEG Decoding

This repository provides the official implementation of "DSAINet: An Efficient Dual-Scale Attentive Interaction Network for General EEG Decoding".

📖 Overview

The overall architecture can be summarized as below:

• Shared Spatiotemporal Tokenization: Converts raw EEG signals into compact shared spatiotemporal tokens for subsequent modeling.
• Fine/Coarse-Scale Temporal Convolution: Uses two parallel temporal branches with different receptive fields to capture fine- and coarse-scale temporal patterns.
• Intra-Branch Attentive Refinement: Refines branch-wise features by strengthening salient scale-specific token relationships within each branch.
• Inter-Branch Attentive Interaction: Enables explicit bidirectional information exchange between fine- and coarse-scale representations.
• Adaptive Token Aggregation: Aggregates branch outputs into compact representations by assigning larger weights to more informative tokens.

✨ Highlights

• Unified architecture with consistent architecture-related hyperparameters across datasets
• Dual-scale temporal modeling with fine- and coarse-scale branches
• Structured attention for within-scale refinement and cross-scale interaction
• Strict subject-independent evaluation on 10 public EEG datasets
• Strong accuracy-efficiency trade-off with only ~77K parameters
• Interpretable saliency maps and learned attention patterns

📂 Code Structure

DSAINet/
├── config/
│   ├── DSAINet.yaml        # Default configuration
|   ├── ShallowConvNet.yaml
|   ├── DeepConvNet.yaml
|   ├── EEGNet.yaml
|   ├── Conformer.yaml
|   ├── ADFCNN.yaml
|   ├── LMDANet.yaml
|   ├── MSVTNet.yaml
|   ├── CTNet.yaml
|   ├── MSCFormer.yaml
|   ├── DBConformer.yaml
|   ├── TMSANet.yaml
|   ├── MGFormer.yaml
|   └── Deformer.yaml
├── figure/
|   └── main.png            # Network architecture illustration
├── model/
│   ├── DSAINet.py          # Architecture implementation
|   ├── ShallowConvNet.py
|   ├── DeepConvNet.py
|   ├── EEGNet.py
|   ├── Conformer.py
|   ├── ADFCNN.py
|   ├── LMDANet.py
|   ├── MSVTNet.py
|   ├── CTNet.py
|   ├── MSCFormer.py
|   ├── DBConformer.py
|   ├── TMSANet.py
|   ├── MGFormer.py
|   └── Deformer.py
├── preprocess/             # Preprocessing scripts for each dataset
│   ├── BCIC_IV_2a.py
│   ├── BCIC_IV_2b.py
│   ├── Zhou2016.py
│   ├── OpenBMI.py
│   ├── PhysioNet_MI.py
│   ├── Mumtaz2017.py
│   ├── ADFTD.py
│   ├── EEGMat.py
│   ├── Shin2018.py
│   └── Rockhill2021.py
├── utils/
│   ├── load_data.py        # Unified data loading interface
│   ├── load_model.py       # Build model according to config
│   └── util.py             # Logging, seeding, augmentation, etc.
├── train_loso.py           # LOSO evaluation (leave-one-subject-out)
└── train_kfold.py          # K-fold evaluation (subject-wise KFold)

🚀 Quick Start

Important

Please update the dataset path, log directory, and checkpoint save directory to match your local setup before training or evaluation.

DSAINet is evaluated under two subject-independent protocols:

• LOSO (Leave-One-Subject-Out) for datasets with relatively limited subject numbers.
• Subject-level K-Fold Cross-Validation for the remaining datasets.

⚙️ Common arguments

• --config: path to the model configuration file
• --dataset: dataset name
• --batch-size: batch size
• --lr: learning rate
• --epochs: number of training epochs
• --seed: random seed
• --device: GPU device ID

👤 LOSO (Leave-One-Subject-Out)

BCIC-IV-2a

python train_loso.py --config config/DSAINet.yaml --dataset BCIC-IV-2a --batch-size 32 --lr 0.001 --epochs 100 --seed 0 --device 0

BCIC-IV-2b

python train_loso.py --config config/DSAINet.yaml --dataset BCIC-IV-2b --batch-size 32 --lr 0.001 --epochs 100 --seed 0 --device 0

Zhou2016

python train_loso.py --config config/DSAINet.yaml --dataset Zhou2016 --batch-size 32 --lr 0.001 --epochs 100 --seed 0 --device 0

🔁 Subject-level K-Fold Cross-Validation

OpenBMI

python train_kfold.py --config config/DSAINet.yaml --dataset OpenBMI --batch-size 128 --lr 0.001 --epochs 100 --seed 0 --device 0

PhysioNet-MI

python train_kfold.py --config config/DSAINet.yaml --dataset PhysioNet-MI --batch-size 128 --lr 0.001 --epochs 100 --seed 0 --device 0

Mumtaz2017

python train_kfold.py --config config/DSAINet.yaml --dataset Mumtaz2017 --batch-size 128 --lr 0.0001 --epochs 30 --seed 0 --device 0

ADFTD

python train_kfold.py --config config/DSAINet.yaml --dataset ADFTD --batch-size 128 --lr 0.0001 --epochs 30 --seed 0 --device 0

Rockhill2021

python train_kfold.py --config config/DSAINet.yaml --dataset Rockhill2021 --batch-size 32 --lr 0.0001 --epochs 30 --seed 0 --device 0

EEGMat

python train_kfold.py --config config/DSAINet.yaml --dataset EEGMat --batch-size 32 --lr 0.001 --epochs 30 --seed 0 --device 0

Shin2018

python train_kfold.py --config config/DSAINet.yaml --dataset Shin2018 --batch-size 32 --lr 0.001 --epochs 100 --seed 0 --device 0