torcheeg/torcheeg

What it solves

TorchEEG simplifies the process of analyzing Electroencephalogram (EEG) signals by providing a plug-and-play library. It removes the need for researchers to handle repetitive technical details—such as data IO, preprocessing, and dataset partitioning—allowing them to focus on their core research goals.

How it works

Built on PyTorch, the library organizes EEG analysis into several modular components:

• Data IO and Datasets: Implements a unified input-output format and provides built-in support for benchmark datasets (e.g., DEAP, SEED, DREAMER, AMIGOS) to ensure quick access and reproducibility.
• Transforms: Offers a wide range of preprocessing tools for feature engineering (e.g., Band Power Spectral Density), general operations (normalization), and specific representations for different model architectures like CNNs (grid conversion) and GNNs (graph conversion).
• Model Selection: Provides specialized data splitting methods tailored for EEG research, including subject-dependent and subject-independent (leave-one-subject-out) partitioning.
• Models: Includes implementations of deep learning architectures based on published papers, such as Convolutional Neural Networks (CNNs), Graph Convolutional Neural Networks (GCNs), and Transformers.

Who it’s for

Researchers and data scientists working with EEG signals who want to quickly reproduce existing work or develop new deep learning-based analysis methods without building the entire pipeline from scratch.

Highlights

• Hybrid Preprocessing: Supports both offline processing (save once, use many times) and online processing (on-the-fly transformations).
• Unified IO: Standardizes how EEG data is stored and read to speed up training.
• Diverse Model Support: Ready-to-use implementations of CNNs, GNNs, and Transformers.
• Specialized Splitting: Built-in support for subject-specific cross-validation common in brain-computer interface research.