[Insight-users] Access to Jacobian of LM Optimizer

[Mark Hastenteufel]

Hi Louis,

You're right. One can compute the singular values
from JtJ. But the version of vnl you use currently
does not implement the get_JtJ() method of the 
levenberg-marquardt optimizer. In the 
current cvs snapshot of vxl/vnl this method is implemented.
Are you plan to integrate the current vxl/vnl version
into ITK in next time?


Thanks,

Mark




Luis Ibanez schrieb:
> 
> Hi Mark,
> 
> The itkLevenbergMarquardtOptimizer allows to get
> access to the internal vnl optimizer through the
> GetOptimizer() method. This returns a pointer to
> a vnl_levenberg_marquardt object.
> 
> The vnl_levenberg_marquardt class returns the
> product of the Jacobian with its transpose at
> the end of the optimization. This is returned
> by the method:   get_JtJ();
> 
> This can be seen in
> 
> Insight/Code/Numerics/vxl/vnl/algo/vnl_levenberg_marquardt.h
> 
> Couldn't you compute the singular values of J
> form JtJ ?
> 
> The problem with changing vnl code is that we
> may want to make sure that the same modification
> goes in the Vxl project. ITK has a snapshot
> of the vnl library. Introducing unilateral changes
> will make very hard to update from the vnl
> trunk in the future.
> 
> You may want to post this same question to the
> vxl_list.
> 
>      http://vxl.sourceforge.net/#mail
> 
> Please let us know if you can see a better way
> to get around this problem.
> 
>    Thanks
> 
>      Luis
> 
> =======================================
> 
> Mark Hastenteufel wrote:
> 
> > Hallo!
> >
> > I use the Levenberg-Marquardt Algorithm for
> > optimization. For the computation of the condition
> > number I need access to the Jacobian matrix (condition
> > number is computed via the singular values
> > of the Jacobian).
> >
> > Are there any ideas how to solve this problem? It
> > seems that one have to change code in the vnl
> > library!?
> >
> > Thanks in advance,
> >
> > Mark
> >
> >
> >

-- 
Mark Hastenteufel
Deutsches Krebsforschungszentrum         (German Cancer Research Center)
Div. Medical and Biological Informatics H0100    Tel: (+49) 6221-42 2353
Im Neuenheimer Feld 280                          Fax: (+49) 6221-42 2345
D-69120 Heidelberg              e-mail            M.Hastenteufel@DKFZ.de
Germany                         http://www.dkfz.de/mbi/people/markh.html