Project Page | Paper |
This is the official implementation of the paper "GANSER: A Self-supervised Data Augmentation Framework for EEG-based Emotion Recognition". NOTE: For privacy reasons, we have removed pre-processed data.
In order to help reviewers to reproduce the experimental results of this paper easily, we will open-source our study following this roadmap:
- open-source all the experimental logs and notebooks for reported comparison experiments, ablation studies, and visualization experiments.
- refactor the main training and evaluation scripts, trained parameter files, and documents for out-of-box evaluation.
Please download the DEAP dataset [1], and modify the path to the downloaded DEAP dataset DATASET_FOLD_DIR in train_proposed_pretrain.py and finetune_proposed_pretrain.py. The preprocessors would automatically generate the cache to ./dataset. For the next time when you need to use these pre-processed results, the preprocessors will automatically load them from locally cached files.
[1] S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, and I. Patras, “DEAP: A database for emotion analysis; using physiological signals,” IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18–31, 2011.
To obtain our designed model for generating augmented samples, you need to use the following script to train the proposed adversarial augmentation network:
CUDA_VISIBLE_DEVICES=0 python train_proposed_pretrain.py
Then, you can utilize the following script to train our proposed classifier purely on the real samples. Notably, to speed up the training of five-fold cross-validation, we deploy five-fold training and test at the same time on five GPUs, with kfold_finetune_proposed_pretrain.py.
python kfold_train_backbone.py
You can manually run the script CUDA_VISIBLE_DEVICES=0 python kfold_train_backbone.py --fold n (here, n ranges from zero to four) one by one, of course, if you do not have enough devices.
Finally, as we described in the paper, we propose a multi-factor training network for fine-tuning the classifier with augmented samples. You can run the following script to deploy five-fold training and test at the same time on five GPUs:
python kfold_finetune_proposed_pretrain.py
You can manually run the script CUDA_VISIBLE_DEVICES=0 python finetune_proposed_pretrain.py --fold n (here, n ranges from zero to four) one by one, of course, if you do not have enough devices.
If you have any questions, please feel free to open an issue.