otsuts / tl-seediv Goto Github PK

This project explores the performance of various domain adaptation methods for electroencephalography (EEG) signal classification in a subject-dependent setting. The primary objective is to enhance the generalization capability of the models when dealing with unseen subjects.

We have implemented the following domain adaptation methods:

1. Adversarial Discriminative Domain Adaptation (ADDA)
2. Domain-Adversarial Training of Neural Networks (DANN)
3. Meta-Learning for Domain Generalization (MLDG)
4. MixStyle

We construct and test the project on Windows 11 and Linux serve, the python version is 3.10. The required packages have been listed in the requirements.txt, you can use the followed commands to install them.

pip install -r requirements.txt
pip install -i https://test.pypi.org/simple/ tllib==0.4

1. Clone the repository:

git clone https://github.com/Otsuts/TL-SEEDIV.git
cd TL-SEEDIV

2. Install the required dependencies:

pip install -r requirements.txt
pip install -i https://test.pypi.org/simple/ tllib==0.4

3. Run the main.py script with the desired options:

python main.py --model <model_name> --log <log_folder> --phase <phase> --learn_rate <learning_rate> --batch_size <batch_size> --num_epoch <num_epochs> --weight_decay <weight_decay> --patience <patience> --num_class <num_classes> --trade_off <trade_off> --pretrain_epoch <pretrain_epochs> --pretrain_lr <pretrain_learning_rate> --num_support <num_support> --num_domain <num_domains>

As we have use the config.yaml to save the best hyperparameters, if you want to use our saved hyperparameters. Please uncomment the followed code in the main.py

args = dict_to_namespace(config[args.model])

Then you can run the scipt simply using:

python main.py --model <model_name>

Replace <model_name> with the desired domain adaptation method (dann, adda, mixstyle, or mldg). Other options can be adjusted as needed.

4. Run the project

The script will output the subject-dependent accuracy for the selected domain adaptation method.

You can comment/uncomment the followed code to close/open the tsne visualization

tsne_visualize(self.test_data, label_pred, f'TL Prediction Labels ({self.args.model})')

Our experiments demonstrate the effectiveness of various domain adaptation methods for subject-dependent EEG signal classification. DANN outperforms all the other methods, achieving the highest classification accuracy. MLDG and MixStyle show similar performance, while ADDA achieves slightly lower accuracy.