1. The title of the paper is: "SFT-Net:Spatial–Frequency–Temporal Network based on attention Mechanism for Detecting Driver Fatigue From EEG Signals"

2. The original address of the paper is: https://ieeexplore.ieee.org/document/10149185

3. The GITHUB address of this project is: https://github.com/wangkejie97/SFT-Net

​ In this folder path, use the CMD command to the terminal, (if it is a virtual environment, please switch first), and then execute the following code to install the dependency package.

pip install -r requirements.txt

It contains the following five packages.

• numpy==1.19.5
• scikit_learn==1.0.2
• scipy==1.5.4
• torch==1.9.0
• visdom==0.1.8.9

• DE_3D_Feature.py : Convert raw EEG data of 23 subjects to 3D features.
• DE_4D_Feature.py : Convert 3D features into 4D features according to the 2D topographic map (refer to the paper).
• dataloader : Divide the four-dimensional features and dataset labels into training set (4/5) and test set (1/5) according to the custom five-fold cross-validation.
• train : training and testing, the training curve can be displayed in real time on the web page through visdom.
• myNet : the defined SFT-Net model.
• "./processedData/" : used to store the converted 3D features and 4D features.
• "./pth/" : used to store the model with the highest accuracy in the nth fold training.

1. Open "SFT-Net/DE_3D_Feature", at line 92, replace with the actual data set path in your computer, then run the py file, it will be in "4D-A-DSC- The "data_3d.npy" file is generated under LSTM/processedData".

2. Open "SFT-Net/DE_4D_Feature" and run it directly. After completion, the "data_4d.npy" file will be generated under "SFT-Net/processedData".

3. Open "SFT-Net/dataloader", you can adjust the number of folds in the five-fold cross-validation for verification, set batch_size, or set a random number seed.

4. Open "SFT-Net/train", before starting the training, please open the cmd command line, (if using a virtual environment, please switch first), enter

   python -m visdom.server
   

Then, open the website in the prompt for real-time visualization. You can adjust the learning rate or Epoch yourself.

• Attention visualization can be obtained by spaAtten, freqAtten output by the model network.
• Requires visdom to be turned on while training.