PARSE

This is the implementation of PARSE: Pairwise Alignment of Representations in Semi-Supervised EEG Learning for Emotion Recognition in PyTorch (Version 1.11.0).

Install

pip install parse-pytorch==1.10.2

This repository contains the source code of our paper, using the following datasets:

• SEED: 15 subjects participated in experiments with videos as emotion stimuli (three emotions: positive/negative/neutral) and EEG was recorded with 62 channels at a sampling rate of 1000Hz.

• SEED-IV: 15 subjects participated experiments with videos as emotion stimuli (four emotions: happy/sad/neutral/fear). 62 EEG recordings were collected with a sampling frequency of 1000Hz.

• SEED-V: 16 participants were involved in experiments with videos as emotion stimuli (five emotions: happy/sad/disgust/neutral/fear). 62 EEG recordings were collected with a sampling frequency of 1000Hz.

• AMIGOS: 37 participants completed experiments with both 16 short and 4 long video clips as stimuli. Continuous valence and arousal scores in the range of [-1, 1] were assigned by annotators to each 20-second video snippet. The continuous scores were then transformed into binary classes (positive and negative) with a threshold of 0. A total of 14 EEG channels were recorded at a sampling rate of 128Hz.

Prerequisites

Please follow the steps below in order to be able to train our models:

1 - Install Requirements

pip3 install -r ./requirements.txt

2 - Download the dataset, then load data, perform pre-processing, and perform feature extraction. In SEED-series datasets, you can use the features (e.g. differential entropy) that have been released by the dataset provider. Neither the preprocessing and feature extraction scripts nor the extracted features have been shared by the AMIGOS dataset provider. Therefore, the data processing and feature extraction code while strictly following the official dataset description and the original publication in which the dataset was published.

3 - Save the preprocessed data and EEG into separate folders (e.g., '/train/de/' and '/train/psd/'). Move EEG features and corresponding labels to the address shown in here.

4 - Usage: Train PARSE by 10 labeled data of SEED-V dataset:

CUDA_VISIBLE_DEVICES=0 python3 ./PARSE/main.py --manualSeed 0 --dataset SEED-V --method PARSE --n-labeled 10 --batch-size 8 --alpha 0.25

Train MixMatch by 25 labeled data of AMIGOS dataset:

CUDA_VISIBLE_DEVICES=0 python3 ./PARSE/main.py --manualSeed 0 --dataset AMIGOS --method AdaMatch --n-labeled 25 --batch-size 64 --alpha 0.75 --threshold 0.6

5 - Pretrained Model: We provide the pre-trained models (1.4 GB) for SEED-V dataset as examples which can be downloaded from Google Drive. The downloaded folder contains pre-trained models of each 3-fold experiment for all 16 participants in all 6 few-labeled scenarios. Save the downloaded files to this address, and run the evaluation code using bash file:

bash ./PARSE/local_run.sh

Document Description

• \library\model: model architecture
• \library\optmization: unsupervised loss weight, weight optimization, learning rate decay, etc.
• \library\train_loop: training step for our proposed PARSE and other three holistic semi-supervised methods (MixMatch, FixMatch and AdaMatch) for EEG representation learning.
• \main: implementation of experiment set-up for several recent state-of-the-art SSL methods (MixMatch, FixMatch, Adamatch) and our proposed PARSE for all the four datasets.

If you find this material useful, please cite the following article:

Citation

@article{zhang2022parse,
  title={Parse: Pairwise alignment of representations in semi-supervised eeg learning for emotion recognition},
  author={Zhang, Guangyi and Davoodnia, Vandad and Etemad, Ali},
  journal={IEEE Transactions on Affective Computing},
  volume={13},
  number={4},
  pages={2185--2200},
  year={2022},
  publisher={IEEE}
}

Contact

Should you have any questions, please feel free to contact me at guangyi.zhang@queensu.ca.