In this project we:
- Download and process all physionet ECG files for quick dataloading.
- Deploy pytorch models on these datasets (courtesy of TSAI:https://github.com/timeseriesAI/tsai)
- Any pytorch model can be used here, dataloaders can also be modified.
- Create custom signal processing functions to transform datasets before they are given to the model.
- You add your own in transformation_funcs.py
This project makes heavy use of the tsai library (version 0.3.2), which requires pytorch. We recommend use of the conda environment.
You can clone our conda environment with conda create --name <env>
- conda activate
- install the requirements with pip install -r pip-requirements.txt
- run this shell script to download physionet files: get_data.sh
- run parse_data.ipynb notebook to convert the datasets to numpy arrays (this helps with very quick dataloading)
- run inception.ipynb to see how to make a dataloaders and run one of TSAI's models on the data (you can use any pytorch compatible model):
- run transformation.ipynb to play around with transformations
- run experiment_analysis_*.ipynb to see how we analyzed experiment results