search

pyro-ppl / numpyro

Probabilistic programming with NumPy powered by JAX for autograd and JIT compilation to GPU/TPU/CPU.
https://num.pyro.ai
Apache License 2.0
2.32k stars 246 forks source link

Tutorial or example on embarrassingly parallel/consensus MCMC #417

Open fehiepsi opened 5 years ago

fehiepsi commented 5 years ago

Despite that NumPyro is very fast (comparing to other frameworks), running MCMC for large datasets is still slow. Even that GPU helps us increase 10x the speed, it is still slow. This is when subposterior methods show their advantages. It is also a good chance to illustrate some available utilities consensus and parametric, which merge subposteriors together, to users.

EdwardRaff commented 11 months ago

I have some upcoming work that I think this might solve, but the documentation is very sparse. Are the subposteriors the states from previous NUTS runs? Can it work with the DiscreteHMCGibbs sampler?

fehiepsi commented 11 months ago

Yes, subposteriors come from NUTS runs with a shard of data. The method does not work with discrete ones I think.