Braindecode — Decode raw EEG, ECoG and MEG with deep learning#

EEG · ECoG · MEG · EMG · iEEG · Open source since 2017

Decode raw brain signals with deep learning.

A PyTorch-native toolbox for end-to-end neural decoding. Load a dataset, pick a published model, train. Same scikit-learn API you already know.

Get started → Tutorials API reference GitHub

For neuroscientists who want to work with deep learning, and deep learning researchers who want to work with neurophysiological data.

v1.5 · BSD-3-Clause Python 3.11+ Works with MNE-Python

2017First release

80⁺Contributors

PyTorch+ scikit-learn · + MNE

Built & maintained at

+ 80 community contributors

By the numbers

65+ models 20+ augmentations 150+ datasets via MOABB 700+ datasets via EEGDash 35+ tutorials 80+ contributors 894K downloads 9 yrs of releases

open source since 2017

65+

Published model architectures

20+

EEG-specific augmentations

150+

Public datasets via MOABB

700+

BIDS datasets via EEGDash

Why braindecode

Everything you need to go from signal to prediction.

Built to plug into the EEG ecosystem you already use: every MOABB dataset, every MNE-Python preprocessing function, every scikit-learn training loop, plus 700+ BIDS datasets via EEGDash. One library, no lock-in.

Decode raw electrophysiology

End-to-end models go straight from raw EEG/ECoG/MEG to predictions. No hand-crafted features required.

Models

PyTorch-native, sklearn-friendly

Built on PyTorch and skorch with a Lightning-friendly API. Drop into any training loop you already know.

API reference

Every MOABB dataset, drop-in

All 150+ MOABB datasets work out of the box, with MOABBDataset(...). Plus TUH, HBN, Sleep Physionet, BCI Competition IV, BIDS, and 700+ datasets via sibling project EEGDash.

Datasets

Preprocessing & augmentations

Fully compatible with every MNE preprocessing function and with EEGPrep, plus exponential standardization and 20+ EEG augmentations.

Augmentation

Curated model zoo

EEGNeX, ConvNets, ATCNet, EEGConformer, foundation models. 60+ architectures reproduced from the original papers.

Browse the zoo

Plays well with MNE-Python

Native interop with mne.io.Raw and mne.Epochs. Reuse the analysis tools you already know for QC and visualization.

MNE bridge

Quickstart

Train a model in 15 lines.

From raw motor-imagery EEG to a trained classifier. Each step is a stand-alone module; swap any piece without touching the others.

01 Load the dataset Pull BCI Competition IV-2a via MOABB.
02 Preprocess & window Bandpass-filter, then cut motor-imagery trials.
03 Pick a model EEGNeX with 22 channels, 4 classes, 4.5 s @ 250 Hz.
04 Train skorch wrapper with sklearn-style fit().

train_eegnex.py

PyTorch · CUDA

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

# 1. Load BCI Competition IV-2a (subject 3) via MOABB
from braindecode.datasets import MOABBDataset
dataset = MOABBDataset("BNCI2014_001", subject_ids=[3])
# 2. Bandpass-filter and create event-aligned windows
from braindecode.preprocessing import (
    Preprocessor, preprocess, create_windows_from_events,
)
preprocess(dataset, [Preprocessor("filter", l_freq=4., h_freq=38.)])
windows = create_windows_from_events(dataset)
# 3. Instantiate a published architecture
from braindecode.models import EEGNeX
model = EEGNeX(n_chans=22, n_outputs=4, n_times=1125)
# 4. Train with the skorch-based EEGClassifier
from braindecode import EEGClassifier
clf = EEGClassifier(module=model, max_epochs=10)
clf.fit(windows, y=None)

Trains on CPU or GPU · Swap the dataset, model and trainer freely; same API. Open the full tutorial →

Model zoo

A library of published architectures.

Every model is reproduced from its original paper and ships with a consistent constructor. The shared EEGModuleMixin takes any of n_chans, n_outputs, n_times, chs_info, input_window_seconds or sfreq; the missing ones are derived for you.

Loading… Browse the full zoo on the Models page

+

Hugging Face Hub

Pretrained, on the Hub.

Every Braindecode model implements from_pretrained and push_to_hub, the same unified API as transformers. Models serialize as a config-driven JSON + safetensors pair (diffable, audit-friendly), and datasets ride on Zarr so push / pull stream chunks lazily without re-uploading the full archive.

Models

Curated foundation weights

BENDR, EEGPT, Signal-JEPA, LaBraM, BIOT and more. Pretrained checkpoints stored as safetensors alongside a readable JSON config, so fine-tuning is reproducible and weight provenance stays auditable. Curated and benchmarked under OpenEEG-Bench.

model = EEGPT.from_pretrained("braindecode/eegpt") Visit huggingface.co/braindecode →

Datasets

Share & pull EEG datasets

Push WindowsDataset, EEGWindowsDataset or RawDataset objects to the Hub with one call. Storage is Zarr-backed, so subsequent pulls only fetch the chunks you read. Useful for multi-GB sessions over flaky links. EEGDash mirrors 700+ BIDS-ready EEG/MEG datasets the same way.

ds.push_to_hub("alice/bnci2014_001") Browse EEGDash datasets →

Tutorials

What you can decode

Browse all tutorials →

Motor imagery Trialwise decoding on BCI IV-2a

Sleep staging Chambon 2018 on Sleep Physionet

ECoG Finger-flexion regression on BCI IV-4

Foundation FT Fine-tune Signal-JEPA

Hugging Face Load pretrained foundation models

Hugging Face Push & pull datasets via the Hub

XAI Interpretability sanity checks

Self-supervised Relative positioning pretraining

MEG Temporal generalization on MEG

Augmentation EEG augmentation search

Architecture

A unified API, nine modules.

From raw recording to trained checkpoint, the same nine modules cover every task: classification, regression, sleep staging, self-supervised pretraining.

PHASE 01 Data

braindecode.datasets MOABB, BIDS, TUH, HBN, sleep corpora. braindecode.preprocessing MNE-backed filtering, resampling, windowing. braindecode.samplers Balanced & sequence samplers for windows.

PHASE 02 Modeling

braindecode.models 60+ architectures, ConvNets to foundation. braindecode.modules Reusable layers for new architectures. braindecode.augmentation 20+ EEG-specific data augmentations.

PHASE 03 Train & evaluate

braindecode.training Cropped decoding, scoring callbacks. EEGClassifier / EEGRegressor skorch wrappers with .fit() / .predict(). braindecode.visualization Topomaps, gradient plots, interpretability.

In good company

Built and shared with neighbors.

Open neural decoding doesn't live alone. Braindecode plugs into a wider open-source neuroscience stack, ships sibling projects that round out the data story, and is built on top of by other research toolkits.

Sister project

EEGDash

Data-sharing archive with 700+ BIDS-ready EEG/MEG/iEEG/fNIRS/EMG recordings from collaborating labs.

Benchmark partner

MOABB

Reproducible motor-imagery benchmarking. Braindecode wraps every MOABB dataset out of the box.

Foundation

MNE-Python

Braindecode preprocessing builds on MNE; any mne.io.Raw or mne.Epochs works as input.

FAIR ecosystem Meta

Meta Neuro AI

FAIR's Python suite for Neuro-AI: discovery, preprocessing and deep learning across neural and stimulus data. Ships NeuralTrain, which depends on Braindecode.