Feat/jh/add callback to scipy interface

[jameshammondpx]

Description

Implements callbacks in scipy_interface wrapper of cyipopt.Problem, addressing the Issue raised in #290

Callbacks should be defined as a callable callback(xk), where xk is the current parameter vector, consistent with: https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.minimize.html.

Development Notes

• Refactors IpoptProblemWrapper to subclass cyipopt.Problem
• Enables use of self.get_current_iterate() in intermediate to pass the current candidate xk to the callback
• Docstrings updated accordingly

[moorepants]

Thanks for working on this. We'll need to keep the __init__ method backwards compatible, so you'll need to think of an alternative way of implementing this. We'll need some unit tests to check the functionality also.

[jameshammondpx]

Thanks, have you got any ideas for keeping the __init__ method backwards compatible? I will look at adding some unit tests.

One idea for backwards compatibility is to:

• Make n,m,lb,ub,cl,cu optional arguments to the constructor
• Only call super.__init__() if those arguments are passed
• Only allow callbacks if those arguments are passed

But it's not particularly elegant. The issue is self.get_current_iterate() is required to get information for the callback, but is only defined in cyipopt.Problem.

[moorepants]

Can Problem be an internal attribute/variable? (instead of subclassing)?

[moorepants]

I see that the scipy code creates a Problem without subclassing Problem, is it that the call backs require a subclass of Problem? Maybe making a second IpoptWrapper class based on the subclass would avoid the backwards incompatibilities.

[jameshammondpx]

I've created a new IpoptProblem class which subclasses Problem and: (1) instantiates a IpoptProblemWrapper, (2) overrides the default intermediate method. This way IpoptProblemWrapper remains untouched.

Interested to hear your thoughts.

[jameshammondpx]

I've created a new IpoptProblem class which subclasses Problem and: (1) instantiates a IpoptProblemWrapper, (2) overrides the default intermediate method. This way IpoptProblemWrapper remains untouched.

Interested to hear your thoughts.

@moorepants - what do you think?

[moorepants]

I think this is fine.

[moorepants]

Can you merge in master and lets see if the tests run?

[moorepants]

I merged master and pushed.

[moorepants]

This error occurs FAILED cyipopt/tests/unit/test_scipy_optional.py::test_gh115_eps_option - AttributeError: 'IpoptProblem' object has no attribute 'nfev'.

[moorepants]

get_current_iterate() is only available for Ipopt >=3.14, so you'll need to make this compatible with Ipopt 3.13 also.