pnpl 0.2.0

Load and process brain datasets for deep learning

🍍PNPL Brain Data Deep Learning Library

The current primary use of the PNPL library is for the LibriBrain competition. Click here to learn more and get started!

Welcome to PNPL — a Python toolkit for loading and processing brain datasets for deep learning. The package now ships four MEG dataset loaders (LibriBrain, MEG-MASC, Armeni 2022, MOUS) plus a composable preprocessing pipeline and shared task abstractions.

Features

• Friendly dataset APIs backed by real MEG recordings
• Composable preprocessing pipeline (bads+headpos+sss+notch+bp+ds, etc.)
• On-demand download from Hugging Face (LibriBrain), OSF (MEG-MASC), Radboud WebDAV (Armeni, MOUS), and OpenNeuro (LittlePrince)
• Task-based API: pick a task object, get (x, y) (or (x, y, info)) windows
• Works with PyTorch DataLoader out of the box
• Clean namespace and lazy imports to keep startup fast

Installation

pip install pnpl

This installs the package and its core dependencies.

Usage

A common entry point uses a task object:

from pnpl.datasets import LibriBrain
from pnpl.tasks import SpeechDetection

dataset = LibriBrain(
    data_path="./data/LibriBrain",
    task=SpeechDetection(tmin=0.0, tmax=0.5),
    partition="train",
)

sample_data, label = dataset[0]
print(sample_data.shape, label.shape)

Dataset-specific wrapper classes are also available:

from pnpl.datasets import LibriBrainSpeech, LibriBrainPhoneme

speech_ds = LibriBrainSpeech(data_path="./data/LibriBrain", partition="train")
phoneme_ds = LibriBrainPhoneme(data_path="./data/LibriBrain", partition="train")

The same task-based pattern works for the other corpora:

from pnpl.datasets import Gwilliams2022, Armeni2022, Schoffelen2019
from pnpl.tasks.gwilliams2022 import PhonemeClassification

meg_masc = Gwilliams2022(
    data_path="./data/meg_masc",
    task=PhonemeClassification(tmin=-0.2, tmax=0.6),
    include_subjects=["01"], include_sessions=["0"], include_tasks=["0"],
    preprocessing="notch+bp+ds",
)

For the full LibriBrain release (deep sub-0 across 9 Sherlock books + TIMIT + MOCHA-TIMIT + 30 Moth podcasts, plus 32 broad subjects on Sherlock1 ses-11/ses-12), use LibriBrain100:

from pnpl.datasets import LibriBrain100
from pnpl.tasks import SpeechDetection

ds = LibriBrain100(
    data_path="./data/LibriBrain100",
    task=SpeechDetection(tmin=0.0, tmax=0.5),
    partition="train",
    subjects="deep",       # or "broad", "all", 0, [1, 2, 3], range(1, 33)
    corpus="sherlock",     # or "timit", "mocha", "podcasts", "all"
)

Included Datasets

Class	Source	Auth
LibriBrain (+ LibriBrainSpeech/Phoneme/Word/Sentence)	Hugging Face pnpl/LibriBrain	none
LibriBrain100 (+ LibriBrain100Speech/Phoneme/Word)	HF pnpl/LibriBrain ∪ pnpl/LibriBrain2 (deep + broad release)	none
Gwilliams2022 (MEG-MASC)	OSF ag3kj	none
Armeni2022	Radboud DSC_3011085.05_995_v1	Radboud credentials
Schoffelen2019 (MOUS)	Radboud DSC_3011020.09_236_v1	Radboud credentials
Pallier2025 (LittlePrince Listen)	OpenNeuro ds007523	none

For the Radboud-hosted datasets, set RADBOUD_USERNAME and RADBOUD_PASSWORD (an approved data-sharing agreement is required before access is granted).

Support

In case of any questions or problems, please get in touch through our Discord server.

Quickstart

Load a single run of the LibriBrain Speech dataset and iterate samples:

from pnpl.datasets.libribrain2025 import constants
from pnpl.datasets import LibriBrainSpeech

ds = LibriBrainSpeech(
    data_path="./data/LibriBrain",
    preprocessing_str="bads+headpos+sss+notch+bp+ds",
    include_run_keys=[constants.RUN_KEYS[0]],  # pick a single run
    tmin=0.0,
    tmax=0.2,
    standardize=True,
    include_info=True,
)

print(len(ds), "samples")
x, y, info = ds[0]
print(x.shape, y.shape, info["dataset"])  # (channels,time), (time,), "libribrain2025"

Documentation

We publish documentation with Jupyter Book and GitHub Pages.

• Local preview: pip install -r docs/requirements.txt && jupyter-book build docs/ then open docs/_build/html/index.html.
• GitHub Pages: when made public, enable Pages via repo settings to publish automatically from the existing workflow.

The docs cover:

• Per-dataset pages (docs/libribrain.md, docs/gwilliams2022.md, docs/armeni2022.md, docs/schoffelen2019.md)
• The preprocessing pipeline (docs/preprocessing.md) and tasks (docs/tasks.md)
• Tutorials for the LibriBrain competition tracks

Contributing

We welcome contributions from the community!

• Read the Contributor Guide in docs/contributing.md for setup, coding style, and PR workflow.
• Open issues for bugs and enhancements with clear, minimal repros when possible.
• Tests: add/update pytest tests for any feature or fix.

Quick dev setup:

git clone https://github.com/neural-processing-lab/pnpl.git
cd pnpl
python -m venv .venv && source .venv/bin/activate
pip install -e .
pip install pytest
pytest -q

Support and Questions

• Check the FAQ at docs/faq.md.
• If something is unclear in the docs, please open a documentation issue.

License

BSD‑3‑Clause. See LICENSE for details.