[Insight-users] gradientdescentoptimizer scales

[Elena Pavlovskaia]

Hi Felix,

I am not a prof in ITK, yet I have experience with
itkRegularStepGradientDescentOptimizer
in BSpline registration. Here is what I think.

You need to set one scale for each of your parameters.

Your number of parameters 338 = 2 * 13 * 13 makes me to think that you are
using
2D images and the spline total grid size = 13 in every dimension. Splines
always
have 3 border points in every dimension. This means that your grid size on
image
is 10 * 10, because 13 = 3 + 10.

If your spline grid sizes are indeed as I calculated, then you need to
specify
the same number (338) of scales. Else you need to fix your spline settings.

Elena

--------------------------------------
[Insight-users] gradientdescentoptimizer scales
Felix Bollenbeck bollen at ipk-gatersleben.de
Thu Mar 29 08:05:04 EST 2007

----------------------------------------------------------------------------
----

Hi all,
i want to use a gradient descent optimizer for bspline registration
(LBFGS optimizer fails during linesearch)

vnl_lbfgs: ** EEEK **
 IFLAG= -1  LINE SEARCH FAILED.

i want to use simple gradientdescent for a start to see whats happening.

i assume that itkRegularStepGradientDescentOptimizer has scales for each
dimension and these are max 128 large and identity.

Description: itk::ERROR: RegularStepGradientDescentOptimizer(0x8566940):
The size of Scales is 128, but the NumberOfParameters for the
CostFunction is 338.

which means I have to pass a larger scales vector to the optimizer?

regards,

felix.