VR Cybersickness EEG Dataset - Beyond Subjectivity

This repository provides the framework and scripts for processing, training, and evaluating models related to the VR Cybersickness EEG dataset. The codebase is designed to handle different input modalities, train neural networks, and evaluate performance for classification or regression tasks.

Features

• Modular Design: Supports various input types (e.g., kinematic, power-spectral-difference).
• Flexible Task Selection: Allows for both classification and regression tasks.
• Customizable Training: Integrated with options for early stopping, learning rate scheduling, and performance logging.
• Reproducibility: Seed setting ensures reproducible results across experiments.

Requirements

The repository is built with Python and requires the following dependencies:

• torch (PyTorch)
• torchvision
• numpy
• argparse
• torchutils (for training utilities)

Install the required dependencies using:

pip install -r requirements.txt

Usage

Setting Up

Clone the repository and navigate to the project directory:

git clone https://github.com/eth-siplab/VR_Cybersickness_EEG_Dataset-Beyond_Subjectivity
cd VR_Cybersickness_EEG_Dataset-Beyond_Subjectivity

Running the Script

The main entry point is the main.py script, which can be executed with the following options:

python main.py --patient <PATIENT_ID> \
               --seed <SEED_VALUE> \
               --task <TASK_TYPE> \
               [--wandb] \
               [--input-type <INPUT_TYPE>] \
               [--logprefix <LOG_DIRECTORY>] \
               [--output] \
               [--no-cuda] \
               [--no-save-model] \
               [--no-load-model]

Arguments:

• --patient: Patient ID (required).
• --seed: Random seed for reproducibility (required).
• --num-epochs: Number of training epochs (default: 200).
• --batch-size: Batch size for training (default: 8).
• --task: Task type (classification or regression) (required).
• --wandb: Enable Weights & Biases logging (optional).
• --input-type: Input data type (default: multi-segment).
  • Options: kinematic, power-spectral-difference, power-spectral-no-eeg, power-spectral-no-kinematic.
• --logprefix: Prefix for log and checkpoint files.
• --output: Enable output logging to console.
• --no-cuda: Use CPU instead of GPU.
• --no-save-model: Skip saving the model checkpoint.
• --no-load-model: Skip loading pre-trained model checkpoints.

Example:

python main.py --patient 0012 --seed 42 --task classification --input-type kinematic --logprefix ./logs --output

Key Components

loader

Handles loading of training, validation, and test datasets. Supports multiple data types and formats.

networks

Defines the architecture and configurations for training neural networks based on the selected task and input type.

trainer

Integrates model training, validation, and testing workflows. Includes:

• Early Stopping: Monitors validation loss to prevent overfitting.
• Average Score Logger: Tracks performance metrics during training.

set_seed

Utility to ensure reproducibility by setting seeds for Python, NumPy, and PyTorch.

Logging

Logs are written to both console and files (if --logprefix is specified). Log files include model checkpoints and detailed training/validation/testing performance.

License

This project is licensed under the MIT License.

Acknowledgments

This repository is part of ongoing research at ETH SIPLab. Special thanks to the contributors and developers of torchutils and related libraries.