TMinuit in PyRoot

ROOT
1

Hi,
I want to fit several histograms with a shared parameter, so I’m doing it by using TMinuit directly.
To speed up the fitting I’m using a compiled function as the minimization function while I still want to do the setup in PyRoot. I get an exception.
I’ve reduced the problem to this:
in fit.cxx:

void My_FCN(Int_t&, Double_t*, Double_t&f, Double_t*par, Int_t){
	f=par[0]*par[0];
}

in fit.py:

import ROOT
ROOT.gROOT.ProcessLine(".L fit.cxx+")
fitter=ROOT.TMinuit(1)
fitter.SetFCN(ROOT.My_FCN)
fitter.DefineParameter(0,"p0",1,1,0,0)
fitter.Migrad()

After the setup phase I get the following output:

...
 **********
 **    1 **MIGRAD
 **********
 FIRST CALL TO USER FUNCTION AT NEW START POINT, WITH IFLAG=4.
TypeError: void ::My_FCN(Int_t&, Double_t*, Double_t& f, Double_t* par, Int_t) =>
    could not convert argument 1
Traceback (most recent call last):
  File "fit.py", line 6, in <module>
    fitter.Migrad()
Exception: TMinuit python fit function call failed (C++ exception)

The strange thing is that the following c++ code works both compiled and interpreted:

#include "fit.cxx" #include <TMinuit.h> void FitROOT(){ TMinuit* fit= new TMinuit(1); fit->SetFCN(My_FCN); fit->DefineParameter(0,"p0",1,1,0,0); fit->Migrad(); }
I would have regarded the c++ code and the Python code to be equivalent.
My ROOT version is
ROOT 5.29/01 (trunk@38210, Feb 25 2011, 09:57:26 on macosx64)
my python version is Python 2.7.1.
Any thoughts on why this fails?

2

Hi,

there are a couple of issues, but what it comes down to is that the parameters created by the stub function to had to python are expected to work in a real python function.

Now fixed in trunk, but do note that this setting does not peel away the python layer (it’s not trying to locate the C++ function and passing that). Consequently, you’re unlikely to get a speedup this way.

Cheers,
Wim

3

Hi, thanks a lot Wim,
I have noticed that I might want to use Minuit2 anyways. I’ve set up a quick testcase following the example in math/minuit2/test/testMinimize.cxx, again compiling the function and doing the configuring in PyRoot. This seems to work perfectly well.
What are your thoughts on speed there?
Cheers
Simon

4

Simon,

same problem. If you put the function in C++ anyway, you’re better of calling the SetFCN on the C++ side as well.

At issue is that the calling of python functions works by setting up a custom stub for CINT and handing that to SetFCN. Since creating C++ functions on the fly isn’t supported by the language (it’s easy to do, but not easy to do without setting off virus scanners :slight_smile: ), I’m reusing one stub function wherein the python callable is looked up. On top of that, there’s the translation of variables from C++ -> python -> C++ in your case. All that is much more than the code in the fit function itself.

This is actually a great case to solve with PyPy. Now only if I had time to work on that. :frowning:

Cheers,
Wim