FYI: mcx, numpyro: new NUTS implementations in Python using jax

[ariddell]

Seems like at least two groups are implementing NUTS using Jax. (Jax seems to be the new favorite among Python automatic differentiation libraries.)

• mcx, https://github.com/rlouf/mcx
• numpyro, https://github.com/pyro-ppl/numpyro (a rewrite of pyro to use jax, I think)

These should be added to the list—in case anyone is keeping track—of Python implementations. pymc3 is already on the list. There are probably a few more too.

[ariddell]

I missed oryx. (It’s buried inside of tensorflow probability.)

No NUTS implementation as far as I can see, but it does have HMC using jax.

• “Homepage”: https://github.com/tensorflow/probability/tree/master/spinoffs/oryx
• Examples: https://github.com/tensorflow/probability/spinoffs/oryx/examples/notebooks/a_tour_of_oryx.ipynb

[rlouf]

MCX author here. The PyMC devs, numpyro devs and I are currently discussing getting the HMC/NUTS implementations out of our respective libraries and move them to a separate, modeling language agnostic , repository.

[maedoc]

that syntax tho

sigma <~ dist.Exponential(lmbda)

reminds of the spaceship operator in C++.