Crater Labs ODE-RNN

Repository implementing ‘Improved Batching Strategy For Irregular Time-Series ODE’ https://arxiv.org/abs/2207.05708

This includes the implementation of our models, dataloaders, as well as a wrapper for ODE-RNN by Rubanova et al. (referred to as odeint).

Installation

First, install Pytorch and its related packages, and then install requirements.txt.

pip install --upgrade pip wheel
pip3 install torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio==0.10.2+cu113 -f https://download.pytorch.org/whl/torch/ -f https://download.pytorch.org/whl/torchvision -f https://download.pytorch.org/whl/torchaudio/
pip3 install -r requirements.txt

Requirements

• Python 3.6 - 3.9 and CUDA 11.3
• Nvidia NGC PyTorch container 21.10 (nvcr.io/nvidia/pytorch:21.10-py3)
• MIMIC experiments requires minimum of 163 GB of CPU memory

Running

• Ensure you are in the source folder (src/ for our model, src/odeint for odeint)
• To run a single experiment, use train.py for our model, specifying which experiment you want to run with --experiment [syn|mujoco|mimic] flag
• To run multiple experiments recreating the paper, run src/schedule.sh for our model or src/odeint/schedule_int.sh scripts for odeint.
• The test result after training will be recorded in the .out output file within the lightning_logs directory

Contents

All files with _int suffix are for Duvenaud's ODE-RNN model, otherwise are for our own model.

• train.py and odeint/train_int.py are execution files for our model and odeint respectively.
• model.py and odeint/model_int.py are the model files.
• schedule.sh and odeint/schedule_int.sh are scripts to run sample experiments for our model and odeint respectively.
• dataloader_mujoco.py, dataloader_phy.py and dataloader_syn.py are the dataloaders for Mujoco, MIMIC, and synthetic data respectively used to train the models.
• data folder contains the data for MuJoCo experiments.
• params folder contains the default hyperparameters for the experiments.

Data processing for MIMIC-IV

• Request access to MIMIC-IV data and download MIMIC-IV v1.0 files
  https://physionet.org/content/mimiciv/1.0/

• Following the procedure from the repo of the paper "Neural flows: Efficient alternative to neural ODEs" by Biloˇs et al. (https://arxiv.org/abs/2110.13040), run all the notebooks except mimic_prep.ipynb from the following repo. Finally run datamerging.ipynb to yield full_dataset.csv
  https://github.com/mbilos/neural-flows-experiments/tree/master/nfe/experiments/gru_ode_bayes/data_preproc

• Modify get_data.py at line 64 and insert a line to save full_data as a csv, then run it to obtain the csv file
  https://github.com/mbilos/neural-flows-experiments/blob/master/nfe/experiments/gru_ode_bayes/lib/get_data.py

full_data.to_csv('data/MIMIC-IV-intercepted.csv')

• Run the python script src/MIMIC-IV_processing.py from this repo using the csv file obtained above

Reference:

• Marin Biloˇs, Johanna Sommer, Syama Sundar Rangapuram, Tim Januschowski, and Stephan G¨unnemann. Neural flows: Efficient alternative to neural ODEs. Advances in Neural Information Processing Systems, 2021.
• Yulia Rubanova, Ricky T. Q. Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series.