How to access the loss for each proposal and how to tell the accepted and rejected samples?

Misc.
1

Hi Pyro community,

I’m new to the Pyro package, and I want to extend my thanks to the developers for this powerful tool. I’m currently working on PDE-constrained parameter optimization and have set up Pyro to run HMC sampling for parameters in a PDE to fit observed data. I have a few questions and would appreciate any guidance:

  1. How can I access the loss for each of the proposed parameter samples?
  2. I understand that I can get the samples using mcmc.get_samples(), but I believe these include both accepted and rejected samples. How can I obtain only the accepted samples?

Thank you in advance for your help.

2

By loss, do you mean the potential energy?

For “rejected” info, you can compare two contiguous samples, something like norm(samples[:-1] - samples[1:]) < 1e-6. Note that “reject” is just a part of Metropolis Hasting proposal - you should include those “repeated” samples when computing posterior diagnostics/predictions. We do not return “rejected” samples in the mcmc.get_samples() method.

3

Certainly! Here’s a refined version of your text:

Thank you for your response. I now understand question #2.

By “loss,” I am referring to the potential energy, which is calculated as log(-misfit). I realize that I can use the predictive function to make predictions. However, running PDE simulations in my case is time-consuming, and I prefer not to repeat them for the proposed samples. Therefore, I would like to know if there is a way to access the potential energy after running MCMC?

Thanks!

4

If you use NumPyro, you can collect it via extra_fields of MCMC.run. We currently don’t have such api in Pyro.

1 Like
5

I see. Thank you!