[Insight-users] Re: SPSA Optimizer
Mathieu De Craene
decraene at tele.ucl.ac.be
Thu Mar 24 05:47:35 EST 2005
Hi Stefan,
Thanks a lot for such a nice work. I just would like to make sure I
understand the way to deal with scales in optimization.
If we refer to the classical gradient descent optimizer,
http://www.itk.org/cgi-bin/viewcvs.cgi/Code/Numerics/itkGradientDescentOptimizer.cxx?root=Insight&view=markup
it seems there is to phases. In the first one, the gradient is computed
without taking into account the scales
m_CostFunction->GetValueAndDerivative(
this->GetCurrentPosition(), m_Value, m_Gradient );
In the second phase, the scales are applied to the gradient before
updating the parameters
transformedGradient[j] = m_Gradient[j] / scales[j];
In your implementation, you take into account the scales in the first
phase (gradient computation) when applying the perturbation
m_Delta[j] /= sqrt(scales[j]);
I understand than applying too big perturbations can bring the optimizer
in a region where the metric is not nicely behaving... But aren't we
biasing the gradient computation using this trick ?
Dank voor de moeite,
Math.
Le jeudi 24 mars 2005 à 11:08 +0100, Stefan Klein a écrit :
> Hi Mathieu, Daniel and others,
>
> I put together a new SPSA optimizer, based on the three version that
> we had. I also added a Test.
>
> You may download the zip from:
> http://www.isi.uu.nl/People/Stefan/
>
>
> > - what should be the default values for a, A, C?
> > - I used a=A=c=0.1 (Don't trust my implementation...)
> > - Mathieu uses a=01.6, A = 60(!), C=0.01
> > - Stefan uses a=A=c=1.0
> > - Are we OK with Alpha = 0.602, Gamma = 0.101?
>
> alpha = 0.602 and gamma = 0.101, as J. Spall suggests.
>
> c depends on the amount of noise in your problem and the range of your
> parameters, so could by anything. So, 1.0 seemed the most appropriate
> to me as a default.
>
> a depends on your specific problem as well, so is set to 1.0 by
> default.
>
> A is set to 10.0, which is 10% of the default
> MaximumNumberOfIterations.
>
> The Scales thing gave me quite a headache but I think I did it correct
> now. In the .cxx I tried to explain why I did it like i did...
>
> Please let me know what you think about it!
>
> Regards,
> Stefan.
>
>
> At 10:04 18/03/05, Mathieu De Craene wrote:
> > Le jeudi 17 mars 2005 à 09:10 -0500, Blezek, Daniel J (Research) a
> > écrit :
> > > Stefan, Marius, Mathieu,
> > >
> > > Excellent. I liked the try / catch blocks, and the stopping
> > > criteria in Mathieu's implementation. Essentially, the code is
> > the
> > > same across all 3 implementations... As you would expect from the
> > > nicely written paper.
> > >
> > > - what should be the default values for a, A, C?
> > > - I used a=A=c=0.1 (Don't trust my implementation...)
> > > - Mathieu uses a=01.6, A = 60(!), C=0.01
> >
> >
> > In fact, this value should be changed to 0.5 or 1. I used a Set()
> > method
> > in the main code to change the default value. For rigid registration
> > for
> > instance, I doubt you will measure anything by changing the
> > translations
> > coefficients by 0.01 mm ;-)
> >
> > If the A coefficient is big, it means that the optimizer will make
> > big
> > steps at the initial iterations.
> > An idea of the successive trials made by the spsa optimizer in rigid
> > registration can be seen on this animation :
> > http://www.tele.ucl.ac.be/~decraene/anims/RIGID.mpeg
> >
> > I think it is good to keep this parameter big enough if we want to
> > keep
> > the robust behavior of this optimizer.
> >
> > > - Stefan uses a=A=c=1.0
> > > - Are we OK with Alpha = 0.602, Gamma = 0.101?
> > >
> >
> > I have never experimented other set of values but these work fine
> > for
> > me !
> >
> > >
> > > -dan
> > > -----Original Message-----
> > > From: Stefan Klein [mailto:stefan at isi.uu.nl]
> > > Sent: Thursday, March 17, 2005 8:55 AM
> > > To: Blezek, Daniel J (Research); Mathieu De Craene
> > > Cc: Insight-users at itk.org; insight-developers at itk.org
> > > Subject: RE: [Insight-users] some contributions
> > >
> > >
> > > Eh, but then we'll have 2 SPSA-optimizers....
> > > Wouldn't it be the best to have one that combines the good
> > of
> > > all 3 versions?
> > >
> > > plan:
> > > - we include the scales thing and the test and some
> > elements
> > > from Mathieu's implementation
> > > - we'll mail it to the list.
> > > - you check it (if we combined all features correctly) and
> > > correct it
> > > - Daniel commits it to the itk.
> > >
> > > Marius&Stefan.
> > >
> > >
> > >
> > > At 08:39 17/03/05, Blezek, Daniel J (Research) wrote:
> > > > Oddly enough, I also have an implementation of the SPSA
> > > > algorithm! I offered to put Stefan's implementation in,
> > > > provided he account for m_Scale and write a test. I'll
> > make
> > > > the same offer for you, Mathieu. I've put my code on
> > the
> > > > itk Wiki @ http://itk.org/Wiki/SPSAOptimizer however, I
> > > > haven't tested my code very well, as it's new.
> > > >
> > > > -dan
> > > >
> > > > -----Original Message-----
> > > > From: insight-users-bounces at itk.org
> > > > [mailto:insight-users-bounces at itk.org]On Behalf Of
> > Mathieu
> > > > De Craene
> > > > Sent: Thursday, March 17, 2005 8:20 AM
> > > > To: Stefan Klein
> > > > Cc: Insight-users at itk.org; insight-developers at itk.org
> > > > Subject: Re: [Insight-users] some contributions
> > > >
> > > >
> > > > Le mercredi 16 mars 2005 à 21:34 +0100, Stefan Klein a
> > > > écrit :
> > > > > Hi all,
> > > > >
> > > > > We have created some classes that are useful for us.
> > We
> > > > hope also for other
> > > > > people, so we would like to contribute them. The code
> > can
> > > > be found for
> > > > > download on:
> > > > > http://www.isi.uu.nl/People/Stefan/
> > > > >
> > > > > - itkEulerTransform.h: This class is a wrap around the
> > > > Euler2DTransform and
> > > > > the Euler3DTransform. In this way it is templated over
> > the
> > > > dimension, like
> > > > > other Transform-classes.
> > > > >
> > > > > - SimultaneousPerturbation.zip: This class implements
> > the
> > > > so-called
> > > > > Simultaneous Perturbation method as an ITK optimizer.
> > For
> > > > more info on this
> > > > > optimization method, go to:
> > > > > http://www.jhuapl.edu/SPSA
> > > > >
> > > >
> > > > I have also have some code for SPSA optimization. I
> > found it
> > > > quite
> > > > useful for rigid registration (it seems to be pretty
> > > > robust). The source
> > > > code for a rigid registration application based on ITK
> > is
> > > > available on
> > > >
> > > > http://euterpe.tele.ucl.ac.be/Waleo2/vesale/
> > > >
> > > > I would like to compare our implementations of this
> > > > optimizer.
> > > >
> > > > Regards,
> > > >
> > > > Mathieu.
> > > >
> > > > > - FullSearch.zip: This class implements a (semi-)Full
> > > > Search optimization
> > > > > routine. This can be useful for research, to scan the
> > > > optimization surface,
> > > > > or to evaluate the result of other optimizers. It
> > allows
> > > > the user to
> > > > > specify which parameters should be evaluated, in which
> > > > range.
> > > > >
> > > > > Since we have contributed 3 things now, we also have a
> > > > feature request; as
> > > > > a reward :-)
> > > > >
> > > > > Would it be possible to add a public function to the
> > > > > MattesMutualInformationImageToImageMetric that forces
> > the
> > > > metric to select
> > > > > new spatial samples? This would be useful if you would
> > > > want new spatial
> > > > > samples every iteration for example.
> > > > >
> > > > > We hope that the contributions are useful. Any
> > comments
> > > > are welcome!!
> > > > >
> > > > > Marius and Stefan.
> > > > >
> > > > >
> > > > > Marius Staring, Stefan Klein
> > > > > Image Sciences Institute
> > > > > University Medical Centre Utrecht
> > > > > Heidelberglaan 100, 3584 CX Utrecht, The Netherlands
> > > > > phone: +31 (0)30 250 3186, fax: +31 (0)30 251 3399
> > > > > {marius,stefan}@isi.uu.nl,
> > > > > http://www.isi.uu.nl/People/Marius
> > > > > http://www.isi.uu.nl/People/?stefan
> > > > >
> > > >
>
More information about the Insight-users
mailing list