HMCECS: variance of the log likelihood estimator

numpyro
1

Hello!

Thank you for realization the subsampling HMC!

I have two questions about the implementation

  1. As I know, the variance of the log likelihood is one of the key parameters of the subsampling MCMC. But I can not find a way how to control it or return it to user.

The variance calculates in the function perturbed_method and it is not stored in the memory.

Am I miss smth or in the current version I can not control/return the variance?

  1. Now, you are using bias-corrected estimator for the log-likehood. If I prefer another estimation, Is it enough to change estimator in perturbed_method?
    I asked it, because in the estimate_likelihood hardcode string about bias corrected estimator
2

Hi @ChernovAndrey, it would be complicated to expose the variance term. I think you can try to use host_callback with numpyro source code to print out those variance values, though I’m not positive about this. How about just calculating that term by hand?

If I prefer another estimation, Is it enough to change estimator in perturbed_method?

Yes, I think so.