How to discard the sampling results of some variables when running mcmc？

numpyro
1

By default, numpyro will save all the random variables, however, I want to run a very long mcmc and too many random variables will take up too much memory. Is there a way to save only a subset of variables instead of all? Thanks in advance!

2

I think you can subclass HMC/NUTS and remove unnecessary terms in postprocess_fn.

1 Like
3

Actually, currently we perform postprocess_fn after collecting all states. So the above solution won’t save memory. I think we can add a flag to apply postprocess_fn during collecting all states at this place. Please make a FR/PR for this. :slight_smile:

1 Like
4

note there there is also a thinning argument here. this won’t let you subset variables but it’ll let you keep only e.g. every 10th sample which also leads to large memory savings

1 Like