Skip to main content

TorchEEG is a library built on PyTorch for EEG signal analysis. TorchEEG aims to provide a plug-and-play EEG analysis tool, so that researchers can quickly reproduce EEG analysis work and start new EEG analysis research without paying attention to technical details unrelated to the research focus.

Project description

PyPI Version Docs Status Downloads Join the Chat

Documentation | TorchEEG Examples

TorchEEG is a library built on PyTorch for EEG signal analysis. TorchEEG aims to provide a plug-and-play EEG analysis tool, so that researchers can quickly reproduce EEG analysis work and start new EEG analysis research without paying attention to technical details unrelated to the research focus.

TorchEEG specifies a unified data input-output format (IO) and implement commonly used EEG databases, allowing users to quickly access benchmark datasets and define new custom datasets. The datasets that have been defined so far include emotion recognition and so on. According to papers published in the field of EEG analysis, TorchEEG provides data preprocessing methods commonly used for EEG signals, and provides plug-and-play API for both offline and online pre-proocessing. Offline processing allow users to process once and use any times, speeding up the training process. Online processing allows users to save time when creating new data processing methods. TorchEEG also provides deep learning models following published papers for EEG analysis, including convolutional neural networks, graph convolutional neural networks, and Transformers.


TorchEEG depends on PyTorch, please complete the installation of PyTorch according to the system, CUDA version and other information:

# Conda
# please refer to
# e.g. CPU version
conda install pytorch==1.11.0 torchvision torchaudio cpuonly -c pytorch
# e.g. GPU version
conda install pytorch==1.11.0 torchvision torchaudio cudatoolkit=11.3 -c pytorch

# Pip
# please refer to
# e.g. CPU version
pip install torch==1.11.0+cpu torchvision==0.12.0+cpu torchaudio==0.11.0 --extra-index-url
# e.g. GPU version
pip install torch==1.11.0+cu102 torchvision==0.12.0+cu102 torchaudio==0.11.0 --extra-index-url


Since version v1.0.10, torcheeg supports installing with conda! You can simply install TorchEEG using Anaconda, just run the following command:

conda install -c tczhangzhi -c conda-forge torcheeg


TorchEEG allows pip-based installation, please use the following command:

pip install torcheeg


In case you want to experiment with the latest TorchEEG features which are not fully released yet, please run the following command to install from the main branch on github:

pip install git+


TorchEEG provides plugins related to graph algorithms for converting EEG in datasets into graph structures and analyzing them using graph neural networks. This part of the implementation relies on PyG.

If you do not use graph-related algorithms, you can skip this part of the installation.

# Conda
# please refer to
conda install pyg -c pyg

# Pip
# please refer to
# e.g. CPU version
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f
# e.g. GPU version
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f

More About TorchEEG

At a granular level, PyTorch is a library that consists of the following components:



A set of unified input and output API is used to store the processing results of various EEG databases for more efficient and convenient use.

torcheeg.da tasets

The packaged benchmark dataset implementation provides a multi-process preprocessing interface.

torcheeg.transf orms

Extensive EEG preprocessing methods help users extract features, construct EEG signal representations, and connect to commonly used deep learning libraries.


Extensive dataset partitioning methods for users to experiment with different settings.

torchee g.models

Extensive baseline method reproduction.

Implemented Modules

We list currently supported datasets, transforms, data splitting, and deep learning models by category.

Datasets: All datasets rely on a set of efficient IO APIs,, to store data preprocessing results on disk and read them quickly during training. Data preprocessing and storage support multiprocessing (speed up!).

Transforms: TorchEEG provides extensive data transformation tools to help users build EEG data representations suitable for a variety of task formulation and a variety of model structures.

Data Splitting: In current research in the field of EEG analysis, there are various settings based on different considerations for data partitioning. Please choose a reasonable data division method according to the research focus:

Models: Coming soon after pushing to align with the official implementation or description. If the current version of CNNs, GNNs and Transformers is to be used, please refer to the implementation in torcheeg.models.


In this quick tour, we highlight the ease of starting an EEG analysis research with only modifying a few lines of PyTorch tutorial.

The torcheeg.datasets module contains dataset classes for many real-world EEG datasets. In this tutorial, we use the DEAP dataset. We first go to the official website to apply for data download permission according to the introduction of DEAP dataset, and download the dataset. Next, we need to specify the download location of the dataset in the root_path parameter. For the DEAP dataset, we specify the path to the data_preprocessed_python folder, e.g. ./tmp_in/data_preprocessed_python.

from torcheeg.datasets import DEAPDataset
from torcheeg.datasets.constants.emotion_recognition.deap import DEAP_CHANNEL_LOCATION_DICT

dataset = DEAPDataset(io_path=f'./tmp_out/deap',
                      ]), num_worker=4)

The DEAPDataset API further contains three parameters: online_transform, offline_transform, and label_transform, which are used to modify samples and labels, respectively.

Here, offline_transform will only be called once when the dataset is initialized to preprocess all samples in the dataset, and the processed dataset will be stored in io_path to avoid time-consuming repeated transformations in subsequent use. If offline preprocessing is a computationally intensive operation, we also recommend setting multi-CPU parallelism for offline_transform, e.g., set num_worker to 4.

online_transform is used to transform samples on the fly. Please use online_transform if you don’t want to wait for the preprocessing of the entire dataset (suitable for scenarios where new transform algorithms are designed) or expect data transformation with randomness each time a sample is indexed.

Next, we need to divide the dataset into a training set and a test set. In the field of EEG analysis, commonly used data partitioning methods include k-fold cross-validation and leave-one-out cross-validation. In this tutorial, we use k-fold cross-validation on the entire dataset (KFold) as an example of dataset splitting.

from torcheeg.model_selection import KFold

k_fold = KFold(n_splits=10,

Let’s define a simple but effective CNN model according to CCNN:

class CNN(torch.nn.Module):
    def __init__(self, in_channels=4, num_classes=3):
        self.conv1 = nn.Sequential(
            nn.ZeroPad2d((1, 2, 1, 2)),
            nn.Conv2d(in_channels, 64, kernel_size=4, stride=1),
        self.conv2 = nn.Sequential(
            nn.ZeroPad2d((1, 2, 1, 2)),
            nn.Conv2d(64, 128, kernel_size=4, stride=1),
        self.conv3 = nn.Sequential(
            nn.ZeroPad2d((1, 2, 1, 2)),
            nn.Conv2d(128, 256, kernel_size=4, stride=1),
        self.conv4 = nn.Sequential(
            nn.ZeroPad2d((1, 2, 1, 2)),
            nn.Conv2d(256, 64, kernel_size=4, stride=1),

        self.lin1 = nn.Linear(9 * 9 * 64, 1024)
        self.lin2 = nn.Linear(1024, num_classes)

    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)

        x = x.flatten(start_dim=1)
        x = self.lin1(x)
        x = self.lin2(x)
        return x

Specify the device and loss function used during training and test.

device = "cuda" if torch.cuda.is_available() else "cpu"
loss_fn = nn.CrossEntropyLoss()
batch_size = 64

The training and validation scripts for the model are taken from the PyTorch tutorial without much modification. Usually, the value of batch contains two parts; the first part refers to the result of online_transform, which generally corresponds to the Tensor sequence representing EEG signals. The second part refers to the result of label_transform, a sequence of integers representing the label.

def train(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)
    for batch_idx, batch in enumerate(dataloader):
        X = batch[0].to(device)
        y = batch[1].to(device)

        # Compute prediction error
        pred = model(X)
        loss = loss_fn(pred, y)

        # Backpropagation

        if batch_idx % 100 == 0:
            loss, current = loss.item(), batch_idx * len(X)
            print(f"loss: {loss:>7f}  [{current:>5d}/{size:>5d}]")

def valid(dataloader, model, loss_fn):
    size = len(dataloader.dataset)
    num_batches = len(dataloader)
    val_loss, correct = 0, 0
    with torch.no_grad():
        for batch in dataloader:
            X = batch[0].to(device)
            y = batch[1].to(device)

            pred = model(X)
            val_loss += loss_fn(pred, y).item()
            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
    val_loss /= num_batches
    correct /= size
    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {val_loss:>8f} \n")

Traverse k folds and train the model separately for testing. It is worth noting that, in general, we need to specify shuffle=True for the DataLoader of the training data set to avoid the deviation of the model training caused by consecutive labels of the same category.

for i, (train_dataset, val_dataset) in enumerate(k_fold.split(dataset)):

    model = CNN().to(device)
    optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)

    train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
    val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)

    epochs = 50
    for t in range(epochs):
        print(f"Epoch {t+1}\n-------------------------------")
        train(train_loader, model, loss_fn, optimizer)
        valid(val_loader, model, loss_fn)

For more specific usage of each module, please refer to the documentation.

Releases and Contributing

TorchEEG is currently in beta; Please let us know if you encounter a bug by filing an issue. We also appreciate all contributions.

If you would like to contribute new datasets, deep learning methods, and extensions to the core, please first open an issue and then send a PR. If you are planning to contribute back bug fixes, please do so without any further discussion.


TorchEEG has a MIT license, as found in the LICENSE file.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

torcheeg-1.0.11.tar.gz (161.1 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page