Portiloop training

Quick start guide

• clone the repo
• cd to the root of the repo (i.e., the folder where setup.py is)
• pip install with the -e option:

pip install -e .

• download the datasets.zip and the experiments.zip files
• unzip the datasets.zip file and paste its content under Portiloop>portiloop_software>dataset
• unzip the experiments.zip file and paste its content under Portiloop>portiloop_software>experiments

Simulation:

The simulate_Portiloop_1_input_classification.ipynb notebook enables performing inference and analyzing the real-time performance of the neural network in terms of stimulation. This notebook can be executed with jupyter.

Offline inference:

We enable easily using out trained artificial neural network on your own data to detect sleep spindles (note that the data must be collected in the same experimental setting as MODA for this to work, see our paper).

This is easily done by writing your signal in a simple text file, on the model of the example_data_not_annotated.txt file provided in the datasets.zip file.

Importantly, note that your signal needs to be preprocessed with our matlab script to ensure the sampling frequency is 250Hz and to simulate the Portiloop online filtering operations. Please adapt and execute this script.

The output file can then be directly used for inference in our offline_inference notebook.

Training:

Functions used for training are defined in python under the portiloop_software folder. We provide bash scripts examples for SLURM to train the model on HPC systems. Adapt these scripts to your configuration. Your training curves can be visualized in real time easily using wandb (the code is ready, you may adapt it to your project name and entity).