Implementation of neural ordinary differential equations built for deeplearning4j.
[Arxiv]
[Pytorch repo by paper authors]
NOTE: This is very much a work in progress and given that I haven't touched a differential equation since school chances are that there are conceptual misunderstandings.
As of now, there is no ODE solver implementation which uses ND4J and this makes the ODE net very slow and therefore impractical to evaluate. As such, the implementation is basically unverified except for indications that test score steadily decreases in the MNIST example.
GIT clone and run with maven or in IDE.
git clone https://github.com/DrChainsaw/neuralODE4j
cd neuralODE4j
mvn install
Currently only the MNIST toy experiment from the paper is implemented [link]
Maven and GIT. Project uses ND4Js CUDA 10 backend as default which requires CUDA 10. To use CPU backend instead, set the maven property backend-CPU (e.g. through the -P flag when running from command line).
All contributions are welcome. Head over to the issues page and either add a new issue or pick up and existing one.
TBD
- Christian Skärby - Initial work - DrChainsaw
This project is licensed under the MIT License - see the LICENSE.md file for details
- Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, David Duvenaud for a very cool and inspiring paper
- Deeplearning4j for neural nets