Braindecode

Use this skill when a task names braindecode, or asks for deep learning on EEG, ECoG, MEG, or similar electrophysiological recordings with MNE-shaped objects, windowed datasets, skorch wrappers, or Braindecode model families.

Operating sequence

1. Establish the input signal contract: channels, sampling frequency, units, recording/epoch layout, targets, and whether data are local or need a network-backed dataset.

2. Install PyTorch first, then braindecode; add only the optional extras that the selected workflow needs. The minimal check is:

   import braindecode, torch
   print(braindecode.__version__, torch.__version__)
   print(torch.cuda.is_available())  # acceleration probe only
   

3. Route to exactly one primary workflow below. Workflows commonly compose in this order: datasets and windowing -> preprocessing -> models and training; add augmentation or interpretation only when requested.

4. Keep units and preprocessing identical between training and inference. Never infer a model's final temporal shape from the model name; use its signal parameters and a tiny forward check.

5. Treat MOABB, BIDS/OpenNeuro, TUH, Sleep Physionet, Hugging Face Hub, EEGPrep, and pretrained checkpoints as optional integrations requiring their own dependencies, data, network, credentials, or storage.

Focused routes

• Datasets and windows: Construct datasets from NumPy/MNE objects, attach descriptions and targets, create event/fixed/target-channel windows, split, concatenate, or serialize data. Read datasets-and-windowing.
• Preprocessing: Apply MNE-backed or array-backed preprocessors, filters, resampling, channel operations, scaling, windowing order, parallel execution, or serialized preprocessing. Read preprocessing.
• Models and training: Select/configure a model, infer signal parameters, train EEGClassifier/EEGRegressor, use cropped decoding, score/predict, or load a local/pretrained model. Read models-and-training.
• Augmentation and sampling: Compose signal transforms, use AugmentedDataLoader, or construct sequence, relative-positioning, or self-supervised samplers. Read augmentation-and-sampling.
• Interpretation and visualization: Compute Captum attributions, frequency gradients, topomaps, confusion/metric plots, or sanity checks. Read interpretation-and-visualization.

Shared guardrails

• Use float32 tensors shaped (batch, channels, time) unless a selected model explicitly documents another shape. Preserve channel order and sampling rate.
• Split by subject/session before overlapping windows when evaluating generalization. Do not leak windows from the same recording across splits.
• Keep runtime scripts self-contained and local-data-only by default. Do not run long gallery examples, download datasets, upload private recordings, or log in to a model/data Hub without explicit authorization.
• For missing optional integrations, report the exact extra or package and continue with a local synthetic fixture where behavior is equivalent.
• Read API reference for the verified public surface, troubleshooting for cross-cutting failures, and provenance before deciding whether this graph is stale for a checkout.