AmiciObjective: check that sensitivities wrt all relevant parameters …

[dweindl]

…can be computed

Adds a check whether the amici model in an AmiciObjective is able to compute sensitivities w.r.t. to the objective parameters.

Closes #1414

This raised the question of whether an objective can have implicitly fixed parameters. I.e. whether it should be considered a use case, that some objective parameters that may change the objective value, have zero entries in the gradient. This is what's happening if one uses PetabImporter.create_objective() without passing it through PetabImporter.create_problem(). So far, pypesto is rather unclear on that. Whether this affects optimization will depend on into which Problem this objective will be put.

My assumption here is that this is undesirable and that changes in the objective value should be reflected in the gradient.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 31.05%. Comparing base (0bf7f36) to head (f107c48).

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

Additional details and impacted files

@@             Coverage Diff              @@
##           develop    #1416       +/-   ##
============================================
- Coverage    83.88%   31.05%   -52.83%     
============================================
  Files          160      160               
  Lines        13096    13105        +9     
============================================
- Hits         10985     4070     -6915     
- Misses        2111     9035     +6924     

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

[PaulJonasJost]

Thanks, the check looks sound to me, but probably good if @FFroehlich has another look at it.

Conceptual question: I get it right, that if currently (and henceforth) i create the objective only with PetabImporter.create_objective(), i get an objective back, which will artificially compute gradients/sensitivities for all parameters independent of whether they are fixed or not? artificially in the sense that for fixed parameters it returns 0? So the only difference between passing it/not passing it to .create_problem would be the dimensionality?

[PaulJonasJost]

Should we have a test to check the expected behavior if one does not pass it to PetabImporter.create_problem?

[dweindl]

The problem is that we don't know the expected behaviour if one does not pass it to PetabImporter.create_problem

[PaulJonasJost]

This might be a minor thing, but i find this a really unintuitive way of creating the objective function correctly.

[dweindl]

The problem is, that petab fixed parameters are only handled in create_problem, but not in create_objective. Maybe this could/should be changed. Not completely sure what might rely on the current behaviour.

[FFroehlich]

well both the importer and the problem should be aware of fixes parameters, so I agree that one should naturally expect problem.create_objective, importer.create_objective and importer.create_problem().objective to be equivalent and that passing importer.create_objective(max_sensi_order=1) to importer.create_problem isn't necessary.

[FFroehlich]

Suggested change

	but aren't available missing.
	but aren't available.

[FFroehlich]

well both the importer and the problem should be aware of fixes parameters, so I agree that one should naturally expect problem.create_objective, importer.create_objective and importer.create_problem().objective to be equivalent and that passing importer.create_objective(max_sensi_order=1) to importer.create_problem isn't necessary.