Implement pymc-backend

[flo-schu]

PYMC is a very mature PPL for Bayesian inference. The main challenge is to use ODEs in the computational graph of the probabilistic model

ODE support

Using Sundials for pymc had problems. A better solution would be to use JAX/diffrax for obtaining the gradients. This way not 2 backends need to be maintained.

There is already a blogpost explaining this: https://www.pymc-labs.com/blog-posts/jax-functions-in-pymc-3-quick-examples/

There is an active PR that tries to implement this in pymc: pymc-devs/pytensor#1120,

• Based on the implementation in https://icomo.readthedocs.io/en/stable/api/jax2pytensor.html
• which was mentioned in a forum: https://discourse.pymc.io/t/new-package-transforming-jax-to-pytensor-for-odes-and-other-applications/16191

Additional Resources

• https://calibr8.readthedocs.io/
• https://murefi.readthedocs.io/
• https://doi.org/10.1371/journal.pcbi.1009223
• https://github.com/michaelosthege/pytensor-federated
• https://betanalpha.github.io/assets/case_studies/hierarchical_modeling (Bearbeitet)