Skip to content

Train state-of-the-art Pytorch Models

PhysioEx provides a fast and customizable way to train, evaluate and save state-of-the-art models for different physiological signal analysis tasks with different physiological signal datasets. This functionality is provided by the train command provided by this repository.

Setup

Before using the train command, you need to set up a virtual environment and install the package in development mode. Here are the steps:

  1. Make sure to have anaconda or miniconda correctly installed in your machine, then start installing a new virtual enviroment 

        conda create -n myenv python==3.10

  2. Now jump into the enviroment and upgrade pip 

        conda activate myenv
    conda install pip
    pip install --upgrade pip

  3. Last but not least install PhysioEx in development mode 

        git clone https://github.com/guidogagl/physioex.git
    cd physioex
    pip install -e .

Experiments

  • chambon2018: This experiment uses the Chambon2018 model for sleep stage classification.
  • tinysleepnet: This experiment uses the TinySleepNet model for sleep stage classification.
  • seqsleepnet: This experiment uses the SeqSleepNet model for sleep stage classification (time-frequency images as input).

To run an experiment, use the -e or --experiment argument followed by the name of the experiment. For example:

train --experiment chambon2018

Dataset-experiment compatibility

SleepPhysioNet Dreem
chambon2018 ✔ ✔️
tinysleepnet ✔ ✔️
seqsleepnet ✔ ✔️
contr_chambon2018 ✔ ✔️
contr_tinysleepnet ✔ ✔️
contr_seqsleepnet ✔ ✔️

Train Command

The train command is used to train models. Here are the available arguments:

  • -e, --experiment: Specify the experiment to run. Expected type: str. Default: "chambon2018".
  • -ckpt, --chekpoint: Specify where to save the checkpoint. Expected type: str. Default: None
  • -d, --dataset: Specify the dataset to use. Expected type: str. Default: "SleepPhysionet".
  • -v, --version: Specify the version of the dataset. Expected type: str. Default: "2018".
  • -c, --use_cache: Specify whether to use cache for the dataset. Expected type: bool. Default: True.
  • -sl, --sequence_lenght: Specify the sequence length for the model. Expected type: int. Default: 3.
  • -me, --max_epoch: Specify the maximum number of epochs for training. Expected type: int. Default: 20.
  • -vci, --val_check_interval: Specify the validation check interval during training. Expected type: int. Default: 300.
  • -bs, --batch_size: Specify the batch size for training. Expected type: int. Default: 32.
  • -nj, --n_jobs: Specify the number of jobs for parallelization. Expected type: int. Default: 10
  • -imb, --imbalance: -me "Specify rather or not to use f1 score instead of accuracy to save the checkpoints. Expected type: bool. Default: False

Experimental Results

Sequence Lenght: 3

results table

results table