🔎 Overview

A minimalistic Python library for EEG/MEG deep learning research, primarely focused on self-supervised learning.

📦 Installation

Either clone this repository and perform a local install accordingly

git clone https://github.com/neurocode-ai/neurocode.git
cd neurocode
poetry install

or install the most recent release from the Python Package Index (PyPI).

pip install neurocode

🚀 Example usage

Below you can see an example adapted for a SSL training workflow using the SimCLR framework.

import torch

from pytorch_metric_learning import losses
from neurocode.datasets import SimulatedDataset, RecordingDataset
from neurocode.samplers import SignalSampler
from neurocode.models import SignalNet
from neurocode.training import SimCLR
from neurocode.datautil import manifold_plot, history_plot

sample_data = SimulatedDataset("sample", seed=7815891891337)
sample_data.read_from_file("MEG/sample/sample_audvis_raw.fif")

# create random extrapolated data from the raw MEG recording,
# you need to provide a location to a forward solution (source space) to use
sample_data.simulate("MEG/sample/sample_audvis-meg-eeg-oct-6-fwd.fif")

dataset = RecordingDataset(sample_data.data(), sample_data.labels(), sfreq=200)
train, valid = dataset.train_valid_split(split=0.75)

samplers = {
  'train': SignalSampler(train.data(), train.labels(), train.info(), ...),
  'valid': SignalSampler(valid.data(), valid.labels(), valid.info(), ...),
}

device = "cuda" if torch.cuda.is_available() else "cpu"
model = SignalNet(...)
optimizer = torch.optim.Adam(model.parameters(), ...)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, ...)
criterion = losses.NTXentLoss()

# train the neural network using the self-supervised learning SimCLR framework,
# save or plot the history to see training loss evolution
simclr = SimCLR(model, device, ...)
history = simclr.fit(samplers, save_model=True)

📋 License

All code is to be held under a general MIT license, please see LICENSE for specific information.