This is the PyTorch implementation of TSception V2 using DEAP dataset in our paper:
Yi Ding, Neethu Robinson, Su Zhang, Qiuhao Zeng, Cuntai Guan, "TSception: Capturing Temporal Dynamics and Spatial Asymmetry from EEG for Emotion Recognition", accepted as a regular paper in IEEE Transactions on Affective Computing, PDF available at IEEE Xplore
It is an end-to-end multi-scale convolutional neural network to classify EEG signals. Previous version of TSception(IJCNN'20) can be found at website
Please go to the working directory by:
$ cd ./code
Please create an anaconda virtual environment by:
$ conda create --name TSception python=3.8
Activate the virtual environment by:
$ conda activate TSception
Install the requirements by:
$ pip3 install -r requirements.txt
Please download the DEAP dataset at website. Please place the "data_preprocessed_python" folder at the same location as the script (./code/).
To run the code for the arousal dimension, please type the following command in the terminal:
$ python3 main-DEAP.py --data-path './data_preprocessed_python/' --label-type 'A'
To run the experiments for the valance please set the --label-type 'V'. The results will be saved into "result.txt" located in the same place as the script.
We highly suggest running the code on a Ubuntu 18.04 or above machine using anaconda with the provided requirements to reproduce the results. You can also download the saved model at website to reproduce the results in the paper. After extracting the downloaded "save.zip", please place it at the same location of the scripts, as (./code/save/). Please run the code by:
$ python3 main-DEAP.py --data-path './data_preprocessed_python/' --label-type 'A' --reproduce
If you are interested to apply TSception to other datasets, you can use generate_TS_channel_order() in utils.py to generate the suitable channel order for TSception, and reorder your data on channel dimension before feeding the data to TSception as what we did in reorder_channel() in prepare_data_DEAP.py
from utils import generate_TS_channel_order
from networks import TSception
original_order = ['Fp1', 'AF3', 'F3', 'F7', 'FC5', 'FC1', 'C3', 'T7', 'CP5', 'CP1', 'P3', 'P7', 'PO3', 'O1', 'Oz', 'Pz', 'Fp2', 'AF4', 'Fz', 'F4', 'F8', 'FC6', 'FC2', 'Cz', 'C4', 'T8', 'CP6', 'CP2', 'P4', 'P8', 'PO4', 'O2'] # DEAP is used as an example
TS_order = generate_TS_channel_order(original_order) # generate proper channel orders for the asymmetric spatial layer in TSception
data = torch.randn(1, 1, 32, 512) # (batch_size=1, cnn_channel=1, EEG_channel=32, data_points=512)
idx = []
for chan in TS_order:
idx.append(original_order.index(chan))
data = data[:, :, idx, :] # (batch_size=1, cnn_channel=1, EEG_channel=28, data_points=512) Some channels are not selected, hence EEG channel becomes 28.
TS = TSception(
num_classes=2,
input_size=(1, 28, 512),
sampling_rate=128,
num_T=15, # num_T controls the number of temporal filters
num_S=15, # num_S controls the size of hidden embedding due to the global average pooling. Please increase it if you need a larger model capacity, e.g., subject-independent case
hidden=32,
dropout_rate=0.5
)
preds = TS(data)The author would like to thank Su Zhang, Quihao Zeng, and Tushar Chouhan for checking the code
| Permissions | Limitations | Conditions |
|---|---|---|
| ✅ Modification | ❌ Commercial use | |
| ✅ Distribution | ||
| ✅ Private use |
Please cite our paper if you use our code in your work:
@ARTICLE{9762054,
author={Ding, Yi and Robinson, Neethu and Zhang, Su and Zeng, Qiuhao and Guan, Cuntai},
journal={IEEE Transactions on Affective Computing},
title={TSception: Capturing Temporal Dynamics and Spatial Asymmetry From EEG for Emotion Recognition},
year={2023},
volume={14},
number={3},
pages={2238-2250},
doi={10.1109/TAFFC.2022.3169001}}