zxl260 / SST-EmotionNet

SST-EmotionNet: Spatial-Spectral-Temporal based Attention 3D Dense Network for EEG Emotion Recognition

SST-EmotionNet

SST-EmotionNet: Spatial-Spectral-Temporal based Attention 3D Dense Network for EEG Emotion Recognition

The SST-EmotionNet consists of the spatial-spectral stream and spatial-temporal stream. Each stream is made up of several Attention based 3D Dense Blocks (A3DB) and Transition Layers.

These are the source code of SST-EmotionNet.

Dataset

We evaluate our model on SEED and SEED-IV datasets, which are available at: http://bcmi.sjtu.edu.cn/~seed/downloads.html. The SEED dataset contains subjects' EEG signals when they were watching films clips. The film clips are carefully selected so as to induce different types of emotion, which are positive, negative, and neutral ones. The SEED-IV is an evolution of the original SEED dataset. The category number of emotion change to four: happy, sad, fear, and neutral. The extracted differential entropy (DE) features of the EEG signals in these datasets are used.

Requirements

• Python 3.7.7

• CUDA 10.1

• CuDNN 7.6.5

• numpy==1.16.2

• scipy==1.4.1

• tensorflow_gpu==2.1.0

• Keras==2.3.1

Usage

• Configuration

  We provide a sample configuration file SEED.ini for SEED dataset.

  • input_width denotes the width of 2D map.
  • specInput_length and temInput_length denote how many 2D maps are stacked in the 3D spatial-spectral representation and 3D spatial-temporal representation, respectively.
  • depth_spec and depth_tem denote the number of layers in spatial-spectral stream and spatial-temporal stream.
  • nb_dense_block denotes the number of A3DBs to add to end.
  • gr_spec and gr_tem denote the growth rate of spatial-spectral stream and spatial-temporal stream.

• Training

  Run run.py with -c parameter, which refers to the path of the configuration file for training. For instance, the model is trained by:

    python run.py -c ./config/SEED.ini
  

References

@inproceedings{jia2020sst,
  title={SST-EmotionNet: Spatial-spectral-temporal based attention 3D dense network for EEG emotion recognition},
  author={Jia, Ziyu and Lin, Youfang and Cai, Xiyang and Chen, Haobin and Gou, Haijun and Wang, Jing},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={2909--2917},
  year={2020}
}