Dynesty warnings added

[arrjon]

The wrapper for the dynesty sampler differs significantly to the other samplers in pyPESTO. Usually the samplers expect the objective to be the posterior and x_priors to be defined, from which they compute the likelihood if needed. I added the argument objective_type to the sampler to switch between posterior and likelihood with default set to posterior. I also added checks in the initialize function to very if x_priors is defined in the problem or not.

Furthermore, I added a warning that by default, a uniform prior with bounds specified in the problem is assumed. This warning was missing, but already mentioned in the prior_transform function.

[dilpath]

Thanks!

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 21.42857% with 11 lines in your changes are missing coverage. Please review.

Project coverage is 53.47%. Comparing base (b1298ea) to head (75bc446).

Files	Patch %	Lines
pypesto/sample/dynesty.py	8.33%	11 Missing ⚠️

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@             Coverage Diff              @@
##           develop    #1324       +/-   ##
============================================
- Coverage    81.32%   53.47%   -27.86%     
============================================
  Files          153      153               
  Lines        12515    12529       +14     
============================================
- Hits         10178     6700     -3478     
- Misses        2337     5829     +3492     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[vwiela]

Do we want to add this flag only in the dynesty sampler? I think it would be better to add it on the objective level.

[dilpath]

Feel free to merge with just the warning and leave the objective_type things to a new PR where objective attributes are added.

[vwiela]

Fine then as a temporary fix for the dynesty sampler and to have some warnings.

[Doresic]

Looks good. Good temporary solution for dynesty.

I have a general question, possibly stupid.
As much as I understand, the OBJECTIVE_NEGLOGLIKE and NEGLOGPOST constants are only passed to the dynesty sampler upon construction such that it knows whether it needs to remove the priors from the objective function when calculating the likelihood.

My question is whether the constants are necessary? Why not define the loglikelihood of dynesty to always remove the prior in cases a prior exists? The only case when it should not remove the prior from the objective is when it does not exist, right? I might be missing something.

[arrjon]

My question is whether the constants are necessary? Why not define the loglikelihood of dynesty to always remove the prior in cases a prior exists? The only case when it should not remove the prior from the objective is when it does not exist, right? I might be missing something.

If you load a petab problem, that should work fine. But if you define your objective yourself (which could be the nllh) and you define x_priors as well, you would compute something wrong without giving any warnings. That is not only a problem in dynesty, but also with the other samplers, see #1325.