Hi everyone,
The ONNX community has recently launched a Probabilistic Programming Working Group aimed at supporting probabilistic models and Bayesian inference directly within the ONNX ecosystem.
The goal is to define a standardized probabilistic operator domain and runtime semantics that allow models to be exported and executed across frameworks and hardware.
Areas we’re exploring include:
-
Probability distributions and log-probability operators
-
Bijectors and parameter constraints
-
Stateless RNG semantics compatible with parallel execution
-
Special mathematical functions used in probabilistic models
-
Inference algorithms such as Laplace, Pathfinder, INLA, HMC, NUTS, and SMC
-
Export pathways for frameworks including Pyro, NumPyro, PyMC, Stan, TensorFlow Probability, JAX-based systems, BayesFlow, and Julia/Turing
We’d particularly appreciate input from the Pyro and NumPyro communities given their work around tracing, vectorized inference, and probabilistic programming in PyTorch and JAX.
If you’re interested in participating or providing feedback, feel free to reach out to:
-
Andreas Fehlner Andreas Fehlner - TRUMPF | LinkedIn
-
Adam Pocock Adam Pocock - Oracle | LinkedIn
-
Brian Parbhu Brian Parbhu - M&T Bank | LinkedIn
You’re also welcome to join the working group meetings:
Fridays @ 12 PM EST, every two weeks
Working group repository:
https://github.com/onnx/working-groups/tree/main/probabilistic-programming