SP controlpanel()?

[codykarcher]

So I'm doing the example from #590, and I can't do a controlpanel() on the SP. Is there a way to do this?

[bqpd]

Only Model implements .controlpanel() right now.

[bqpd]

I'll make it so you can make that a Model

[bqpd]

Hmm, but I think conrtolpanels might still only work with .solve ...

[codykarcher]

Yea, that was the error I was getting. Sorry, should have posted:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-7-dcd8be9650e5> in <module>()
----> 1 m.controlpanel()

/Users/codykarcher/Dropbox (MIT)/Research/gpkit/gpkit/constraints/costed.pyc in controlpanel(self, *args, **kwargs)
     83         """
     84         from ..interactive.widgets import modelcontrolpanel
---> 85         return modelcontrolpanel(self, *args, **kwargs)

/Users/codykarcher/Dropbox (MIT)/Research/gpkit/gpkit/interactive/widgets.pyc in modelcontrolpanel(model, *args, **kwargs)
     94     """
     95 
---> 96     sliders = model.interact(*args, **kwargs)
     97     sliderboxes = []
     98     for sl in sliders.children:

/Users/codykarcher/Dropbox (MIT)/Research/gpkit/gpkit/constraints/costed.pyc in interact(self, ranges, fn_of_sol, **solvekwargs)
     74         """
     75         from ..interactive.widgets import modelinteract
---> 76         return modelinteract(self, ranges, fn_of_sol, **solvekwargs)
     77 
     78     def controlpanel(self, *args, **kwargs):

/Users/codykarcher/Dropbox (MIT)/Research/gpkit/gpkit/interactive/widgets.pyc in modelinteract(model, ranges, fn_of_sol, **solvekwargs)
     82             #     print(out)
     83 
---> 84     resolve()
     85 
     86     return widgets.interactive(resolve, **ranges)

/Users/codykarcher/Dropbox (MIT)/Research/gpkit/gpkit/interactive/widgets.pyc in resolve(**subs)
     65         try:
     66             try:
---> 67                 model.solve(**solvekwargs)
     68             except ValueError:
     69                 model.localsolve(**solvekwargs)

AttributeError: 'SignomialProgram' object has no attribute 'solve'

[bqpd]

Okay, so now the following code works:

m = Model(objective, constraints)
sp = m.sp(require_signomials=False)
sp.localsolve('cvxopt', verbosity=1)
sp.controlpanel()

The control panel does what you'd expect, so it's more interesting off the bat if you set Re to 1e6.

[bqpd]

@cjk12, @whoburg, these "signomial programs that do not require signomials" could probably use a catchier name: LocalPrograms? Which would imply you could create one with m.lp()...

[codykarcher]

Did you push this?

[bqpd]

oops, committed but did not push. Pushed now!

[codykarcher]

Ok, this is pretty cool... Whatever nomenclature is fine with me.

[whoburg]

can someone post the full working example?

[codykarcher]

from gpkit import VectorVariable, Variable, units, Model
from gpkit.constraints.set import ConstraintSet

Re = Variable("Re", 1e6, "-", "Reynolds Number")
CL = Variable("C_L", "-", "Lift Coefficient")
CD = Variable("C_D", "-", "Drag Coefficient")

def xfoil(x0):
    return [CD >= 0.01]

class AeroModel(ConstraintSet):

    def as_gpconstr(self, x0):
        gpconstr = super(AeroModel, self).as_gpconstr(x0)
        if x0 and x0["Re"] >= 1e6:
            constraints = xfoil(x0)
        else:
            constraints = [CD >= 1. + CL**2]
        return ConstraintSet([gpconstr, constraints])

constraints = AeroModel([CL >= 1.0, Re >= 1e4*CL])

objective = CD

m = Model(objective, constraints)
sp = m.sp(require_signomials=False)
sp.localsolve('cvxopt', verbosity=1)
sp.controlpanel()

[whoburg]

Thanks. @bqpd, yeah - I tend to think that ideally users would not need the lines

sp = m.sp(require_signomials=False)
sp.localsolve('cvxopt', verbosity=1)

and that Model.controlpanel() would just work. Might make sense to update the error logic that spits out No Signomials remained after substitution.

A more precise name for LocalProgramsmight be SequentialGP. Maybe we should rename CostedConstraintSet to SequentialGP? It's a model that can be solved as a sequence of GP's, and this includes the AeroModel above as well as SignomialProgram.

[bqpd]

SequentialGP seems strictly less precise to me; we have a program whose constraints are a sequence of local approximations of GPs.

re Model.controlpanel working, all that's missing is localsolve. Should localsolve work for non-SP locally approximable programs?

[bqpd]

Closing this issue, will link in the PR thread.

[whoburg]

yes, I think it's fine for localsolve to work for models that contain no signomials but that contain locally approximable constraints (e.g. for external solvers)

[bqpd]

Hmm, okay! Great.

[whoburg]

that sounded suspicious -- am I missing something? I guess the elephant in the room is whether localsolve should work for GPs?

[bqpd]

Yup. See the PR for the added

    def as_posyslt1(self):
        raise ValueError("must be checked during solve")

which lets Cody's example Just Work.