search

neurophysik / jitcdde

Just-in-time compilation for delay differential equations
Other
56 stars 14 forks source link

How do I use non-symbolic functions? #31

Closed saurabhmalani1994 closed 2 years ago

saurabhmalani1994 commented 4 years ago

Is there any way to use JITCDDE for DDEs that cannot be expressed symbolically? I am training a neural network using Tensorflow+Keras to learn a DDE, and I want to use a DDE solver to numerically integrate this learned DDE. However, Tensorflow+Keras can not take in symbolic inputs.

Wrzlprmft commented 4 years ago

Yes, you can use the callback functionality. Search the documentation for callback to read about it. Briefly, it allows you to call a Python function within the derivative. You may need to write a wrapper to make it fit the format.

Be aware that this may be horribly inefficient (as well as it may be not) and it may be worth trying to find a way to export your learned solution in symbolic form. You also may want to look at this issue.