What should I use as hyperparameters for the DefaultNS ?

[renecotyfanboy]

Hi everyone!
First, many thanks to Joshua for bringing us this awesome jax package for nested sampling, it is much appreciated.
I am currently learning to use jaxns and nested sampling in general, but I struggle to find a good parametrization for the sampler. Comparing my runs with NUTS shows that I can't catch the desired correlations and values in my model, as shown below. For information, I am using the contributed code to numpyro to directly use their interface, as my models are already defined with this library.

I tried to increase the number of livepoints to 1000, and enabled "difficult_model" and "parameter_estimation", but this does not converge as intended. What would be your take on this ? Diagnostics below

Also if you have any kind of practical resources explaining the choice and meaning of each hyperparameter in addition to the existing papers, I would be glad to read them!

[Joshuaalbert]

Can you share a gist of the numpyro model? The posterior looks quite innocent such that default parameters should solve this no problem. Also, what version of jaxns are you using? pip freeze | grep jaxns

[renecotyfanboy]

Hi, thank you very much for your help. For you to know, I use a package I am developing to define the numpyro model so you might need to install it with pip install jaxspec. If you'd like, I might rewrite it in pure numpyro, but it will take me some time to do so, in the meantime, here is a gist using jaxspec. I am using jaxns 2.5.1 in my environment.

Maybe this is due to the wrapping with numpyro ? I have no idea tbh

[renecotyfanboy]

Some news about this, I tested this same problem using a package implementing ultranest sampler, and got similar results to what is obtained from NUTS using the default parameters.

I also tested the Gaussian Shell example from numpyro+ jaxns to check if the contributed code is in default, but it seems to work properly too.

So I guess this is an issue with my model or so. I'll investigate this further.

[renecotyfanboy]

Rerunning this in a fresh environment lead to expected results, I probably have an incompatible cocktail package in my dev environment 🥲

[renecotyfanboy]

After a full day of tracking the why of this issue, it simply comes from the prior distributions I chose, which were too restrictive for jaxns to work properly. I moved some to LogUniform instead of Uniform and expanded them a bit and it now works like a charm. So, not a package issue or similar.

[Joshuaalbert]

Glad you figured it out!