[Insight-users] Regular Step Gradient Descent optimizer

[Luis Ibanez]

Hi Tom,


That sounds like a reasonable feature request.

It shouldn't be too hard to implement for the case of the
GradientDescent optimizers.

The OnePlusOne optimizer is already designed in this way,
that is, it retains the best possible value.

Could you please log this as a feature request in the
bugTracker ?


     http://public.kitware.com/Bug/index.php


If you don't have an account yet, you can creat it by
just providing an email.


Please let us know if you find any problem with accessing
the bug tracker.


   Thanks


      Luis



=====================
Tom Vercauteren wrote:
> Hi all,
> 
> Thanks for all the nice optimizers in ITK! Here's a small feature
> request that IMHO could make many of them easier to use.
> 
> As mentioned in this thread, it is possible for several optimizers in
> ITK not to optimize the metric value at each step. Hence the last
> position might not be the best position visited by the optimizer.
> 
> The "disturbing" thing is that the optimizers (and registration
> methods) only allow to directly query for the last position (using
> optimizer->GetCurrentPosition() or
> registration->GetLastTransformParameters() ).
> 
> I know it is possible to keep the best position by using observers,
> but it would be nice if the optimizers could keep this best position
> and if we could simply call something like
> optimizer->GetBestPosition() or
> registration->->GetBestTransformParameters().
> 
> Of course some optimizers (e.g. Powell, LBFGS, FRPR etc.) do not need
> to store this position as they can only optimize the metric value. In
> such a case GetBestPosition() could simply call GetCurrentPosition().
> 
> Regards,
> Tom
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
> 
>