Add sampler to iterative refine an initial course posterior #4341
Conversation
|
This adds a simple sampler that builds an internal kde from samples and does iterative evolution. There is no mcmc done, just weighted mc draws. This is useful for finalizing a posterior that has already identified the key modes (or if the mode structure can be known ahead of time), but has not fully converged. It has trivial parallelization. The goal is to also be able to use it to slightly change prior / model choices in a way that will converge very quickly on the right answer. It includes some checks on the internal kde representation convergence by looking at the JS divergence between iterations and the value of the evidence. Some failure modes (well, you'll get few actually independent samples back) include if the shape of the posterior is badly smoothed by the default kde used. I am exploring some alternate kdes that will do better there. I think awkde used by @tdent may work well. |
There was a problem hiding this comment.
No comment on the code. This is a new model and doesn't touch anything else, so it's hard to say anything meaningful without having tried it out ... It definitely won't break anything else!
It would be good to have some documentation of this. Collin has set a high standard for inference docs, so I think he would ask for the things I suggest here. I'm pressing approve and leaving this to your conscience, however!
| models._global_instance.loglikelihood) | ||
|
|
||
|
|
||
| def resample_equal(samples, logwt, seed=0): |
There was a problem hiding this comment.
Please add a docstring for this function. I think it's important enough that we should know what it does. (sorry these comments come late. I wrote them while github was misbehaving, and it looks like they got swallowed!)
|
|
||
| class RefineSampler(DummySampler): | ||
| """Sampler for kde drawn refinement of existing posterior estimate | ||
|
|
There was a problem hiding this comment.
Can we provide an example of how this might work in the code/documentation? I assume we generate a normal output, and then run this. Is it valid for both nested sampling and MCMC?
|
Definitely some of the codeclimate comments should be addressed. I do think calling things like |
…4341) * add kde refinement sampler * cc * more logging * cc * cc
…4341) * add kde refinement sampler * cc * more logging * cc * cc
…4341) * add kde refinement sampler * cc * more logging * cc * cc
No description provided.