Regularized Neural ODEs (RNODE)

This repository contains code for reproducing the results in "How to train your Neural ODE: the world of Jacobian and Kinetic regularization".

Requirements

• PyTorch 1.0+
• Install torchdiffeq, which provides Python CUDA ODE solvers, from https://github.com/rtqichen/torchdiffeq

Examples

The paper applies regularized neural ODEs to density estimation and generative modeling using the FFJORD framework. Example training scripts for MNIST, CIFAR10, ImageNet64 and 5bit CelebAHQ-256 are found in example-scripts/

Data preprocessing

Follow instructions in preprocessing/

Citation

Please cite as

@article{finlay2020how,
  author    = {Chris Finlay and
               J{\"{o}}rn{-}Henrik Jacobsen and
               Levon Nurbekyan and
               Adam M. Oberman},
  title     = {How to train your neural {ODE}: the world of {Jacobian} and {Kinetic} regularization},
  journal   = {CoRR},
  volume    = {abs/2002.02798},
  year      = {2020},
  url       = {https://arxiv.org/abs/2002.02798},
  archivePrefix = {arXiv},
  eprint    = {2002.02798},
}

Many thanks

FFJORD was gratefully forked from https://github.com/rtqichen/ffjord.