Expose minuit instance in fit

[alexander-held]

Description

A fit via pyhf.infer.mle.fit() with minuit backend currently does not offer access to the minuit instance to for further operations, e.g. access to the correlation matrix. It would be convenient to optionally return it.

Is your feature request related to a problem? Please describe.

The provided pyhf.infer.mle.fit() can only be used for studies with limited scope, as important information such as parameter correlations are inaccessible.

Describe the solution you'd like

A keyword argument to optionally return the minuit instance, similar to return_uncertainties. Alternatively, the correlation matrix might be returned directly, though the full minuit instance might be more convenient (e.g. for Minos errors).

Describe alternatives you've considered

A user wanting more detailed control over the fit could implement it separately and not go through infer to bypass this limitation. This might be a good solution to prevent supporting every last use case (and will offer more detailed control over minuit, which might be desired).

Relevant Issues and Pull Requests

#881

Additional context

none

[kratsg]

You can't get the optimizer used via _, opt = pyhf.get_backend() ?

[alexander-held]

That returns a pyhf.optimize.opt_minuit.minuit_optimizer, and the minuit instance is built within _make_minuit() as mm, but not saved as a member variable after minimize(). I think I cannot get the right object this way.

[lukasheinrich]

I thin i'd rather provide APIs for the common information (eg return_correlation or return_hessian) instead of directly exposing the underlying implementation / minuit instance.

the ctor of the minuit optimizer should be able to take on fine-grained global options

pyhf.optimize.opt_minuit.minuit_optimizer(option1 = , option2 = )

what do you think @alexander-held ?

[alexander-held]

An API to get the available information, like return_correlation, would be very convenient. This prevents users from having to worry about writing backend-agnostic implementations (to the extent that this is possible, since not all backends provide all the information).

the ctor of the minuit optimizer should be able to take on fine-grained global options

Is this already possible now, or a goal? I have not managed so far to pass information to the Minuit instance through pyhf. I am in particular thinking of customizing things like strategy / tolerance.

[kratsg]

As of v0.5.0, the minimize() API provides a return_result_obj which contains a minuit attribute for the underlying minuit instance.

[alexander-held]

The updates in #951 and #988 cover most of what I can think of. The last related thing that would be convenient is support for propagating strategy / tol through to Minuit, see scikit-hep/iminuit#466. Thanks a lot for these updates!

[alexander-held]

Given the outcome of scikit-hep/iminuit#466, I am not sure whether support for strategy / tol through pyhf.infer is possible. I ended up re-implementing the relevant functionality to get access to these settings.