Code for the paper:
Jacob Kelly*, Jesse Bettencourt*, Matthew James Johnson, David Duvenaud. "Learning Differential Equations that are Easy to Solve" arXiv preprint (2020). [arxiv] [bibtex]
*Equal Contribution
Follow installation instructions here.
Follow installation instructions here.
For using the MNIST dataset, follow installation instructions here.
Different scripts are provided for each task and dataset.
python mnist.py --reg r3 --lam 6e-5
python latent_ode.py --reg r3 --lam 1e-2
python ffjord_tabular.py --reg r2 --lam 1e-2
python ffjord_mnist.py --reg r2 --lam 3e-4
tensorflow-datasets
(instructions for installing above) will download the data when called from the training script.
The file physionet_data.py
, adapted from Latent ODEs for Irregularly-Sampled Time Series will download and process the data when called from the training script.
Data must be downloaded following instructions from gpapamak/maf and placed in data/
. Only MINIBOONE
is needed for experiments in the paper.
Code in datasets/
, adapted from Free-form Jacobian of Reversible Dynamics (FFJORD), will create an interface for the MINIBOONE
dataset once it's downloaded.
It is called from the training script.
Code in lib
is modified from google/jax under the license.
@article{kelly2020easynode,
title={Learning Differential Equations that are Easy to Solve},
author={Kelly, Jacob and Bettencourt, Jesse and Johnson, Matthew James and Duvenaud, David},
journal={arXiv preprint arXiv:2007.04504},
year={2020}
}