Warnings for scipy and laplace prior.

[PaulJonasJost]

Closes #1189.

Error should only appear for priors? Otherwise could also throw a general error, but not sure how helpful that one would be.

[codecov-commenter]

Codecov Report

Attention: 1768 lines in your changes are missing coverage. Please review.

Comparison is base (160c2a8) 88.16% compared to head (223ed7f) 71.92%.
Report is 454 commits behind head on develop.

Files	Patch %	Lines
pypesto/hierarchical/optimal_scaling/solver.py	9.95%	362 Missing ⚠️
...ypesto/hierarchical/spline_approximation/solver.py	12.15%	253 Missing ⚠️
pypesto/hierarchical/optimal_scaling/problem.py	16.74%	179 Missing ⚠️
pypesto/hierarchical/petab.py	0.00%	149 Missing ⚠️
pypesto/hierarchical/inner_calculator_collector.py	17.12%	121 Missing ⚠️
...pesto/hierarchical/spline_approximation/problem.py	20.94%	117 Missing ⚠️
pypesto/hierarchical/problem.py	23.07%	100 Missing ⚠️
pypesto/ensemble/ensemble.py	69.36%	68 Missing ⚠️
pypesto/hierarchical/solver.py	21.68%	65 Missing ⚠️
pypesto/hierarchical/util.py	17.10%	63 Missing ⚠️
... and 19 more

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

Additional details and impacted files

@@             Coverage Diff              @@
##           develop    #1228       +/-   ##
============================================
- Coverage    88.16%   71.92%   -16.24%     
============================================
  Files           79      148       +69     
  Lines         5257    11863     +6606     
============================================
+ Hits          4635     8533     +3898     
- Misses         622     3330     +2708     

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

[dilpath]

How does this affect optimization? i.e. does the gradient then become nan when the parameter is at the mean of the Laplace, even if there are finite contributions from the likelihood? If so, then it might be better to define the gradient contribution from the Laplace at the mean as 0, so that the optimizer can still use the gradient information from the likelihood.

[dweindl]

i.e. does the gradient then become nan when the parameter is at the mean of the Laplace, even if there are finite contributions from the likelihood?

Yes. I am not sure whether this is really a problem in practice. I'd guess it's rather unlikely that the optimizer steps exactly on the mean.

If so, then it might be better to define the gradient contribution from the Laplace at the mean as 0, so that the optimizer can still use the gradient information from the likelihood.

Also fine with me.