Fix HierarchicalAmiciCalculator.__call__ not setting 'hess' in result

[dweindl]

HierarchicalAmiciCalculator.__call__ needs to set 'hess' in the returned dict, also in case of simulation failures.

Closes #1160

[codecov-commenter]

Codecov Report

Merging #1161 (81613c1) into develop (160c2a8) will decrease coverage by 3.77%.
Report is 407 commits behind head on develop.
The diff coverage is 87.87%.

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

@@             Coverage Diff             @@
##           develop    #1161      +/-   ##
===========================================
- Coverage    88.16%   84.39%   -3.77%     
===========================================
  Files           79      148      +69     
  Lines         5257    11673    +6416     
===========================================
+ Hits          4635     9852    +5217     
- Misses         622     1821    +1199     

Files	Coverage Δ
pypesto/C.py	100.00% <100.00%> (ø)
pypesto/__init__.py	100.00% <100.00%> (ø)
pypesto/engine/__init__.py	100.00% <100.00%> (ø)
pypesto/engine/multi_process.py	100.00% <100.00%> (+7.69%)	⬆️
pypesto/engine/multi_thread.py	100.00% <100.00%> (ø)
pypesto/engine/single_core.py	100.00% <100.00%> (ø)
pypesto/engine/task.py	100.00% <100.00%> (ø)
pypesto/ensemble/__init__.py	100.00% <100.00%> (ø)
pypesto/ensemble/task.py	100.00% <100.00%> (ø)
pypesto/hierarchical/__init__.py	100.00% <100.00%> (ø)
... and 138 more

[Doresic]

Yes, I think for simulation failures it should return a nan hessian. I think I encountered a situation where that caused an issue if fides (with BFGS hessian, not analytical) is used with hierarchical optimization. So the changes as they are are good.

However, just to mention: I believe that the true hessian is not the same as the one given by AMICI, because of the hierarchical optimization of inner parameters. The gradient from AMICI is correct as the hierarchical parameters are optimal, so the derivatives cancel out. But I'm not sure that happens for the Hessian. It could even be correct, thinking about it now briefly, since the inner problem optimization is convex so for the optimum both first and second-order derivatives should be 0. But I'm not sure, would have to look at it in detail.

[dweindl]

The gradient from AMICI is correct as the hierarchical parameters are optimal, so the derivatives cancel out. But I'm not sure that happens for the Hessian. It could even be correct, thinking about it now briefly, since the inner problem optimization is convex so for the optimum both first and second-order derivatives should be 0. But I'm not sure, would have to look at it in detail.

Would be great if you could double-check that. If not, we need to let Objective know that we can't provide the Hessian instead of happily providing a potentially incorrect one.

That's independent of this PR - will merge this already.