pypesto.profile.parameter_profile leaves Problem in modified state if an error occurs during profiling

[dweindl]

What happens:

fixed_parameters_before = problem.x_fixed_indices

parameter_profile(problem, profile_index=[0], ...) # <- fails for unrelated reasons

assert problem.x_fixed_indices == fixed_parameters_before # <- fails, because parameters were fixed during profiling but not unfixed

# retry computing profiles
parameter_profile(problem, profile_index=[0], ...) # <- nothing happens, because that parameter is still set to fixed

parameter_profile should either create a copy of problem, or make sure that it is in the original state when parameter_profile exists, independent of any potential errors.

[PaulJonasJost]

Would go with copy for now. Any strong arguments against that?

[dweindl]

Would go with copy for now. Any strong arguments against that?

No. Considering the cost of profiling, the copying-overhead should be negligible.
That Problem/Objective is copy-able is already expected elsewhere, so it shouldn't be a problem here either.

[PaulJonasJost]

yes, that's what I thought as well. 👍🏼

[dweindl]

Somewhat independent issue, but related to the pseudo-code above: Would be good to issue some warning when attempting to compute profiles for fixed parameters instead of just not doing anything.

[PaulJonasJost]

You are right, just went through the profile code and noticed that as well. As it Is a somewhat independent issue, I will create a new issue and fix it as well 👍🏼