| BVLS (NNLS) without loops? [message #86953] |
Mon, 16 December 2013 22:32  |
JP
Messages: 55
Registered: April 2008
|
Member |
|
|
Hi IDLers,
I am currently using BVLS (http://www-astro.physics.ox.ac.uk/~mxc/idl/bvls.pro) for spectral unmixing. It works great, but, on one pixel at a time. I want to implement for many pixels without having to loop. Example below:
IDL> help, a
A FLOAT = Array[85, 3]
IDL> help, b
B FLOAT = Array[85]
IDL> help, bnd
BND FLOAT = Array[2, 3]
IDL> print, bnd
0.000000 1.00000
0.000000 1.00000
0.000000 1.00000
bvls, A, B, BND, X_BVLS
IDL> help, x_bvls
X_BVLS FLOAT = Array[3]
in my example A is a vector of "endmembers" (85 spectral bands and 3 fractions), B is a pixel (vector of 85 bands), BND are bounds (don't go negative nor >1) and the X_BVLS are the estimated fractions returned for that pixel.
Now as you could imagine I have many many pixels (n) (my B is really a 2D array of [85, n]) and i've written a function which loops through n, but it gets very slow for large n.
Any ideas?
Thanks
|
|
|
|
| Re: BVLS (NNLS) without loops? [message #86962 is a reply to message #86953] |
Tue, 17 December 2013 09:35  |
Mike[5]
Messages: 5
Registered: October 2013
|
Junior Member |
|
|
Hi Juan Pablo,
From what I understand, your BVLS (Bounded-Variables Least-Squares) problems in every pixel are completely independent. They don't share the coefficients of the A array nor the B vector. The only similarity between the different problems is the size of the arrays.
If this is the case I don't see how one could speed-up the problem by some form of vectorization.
The only real solution would be for IDL to include a compiled BVLS routine in the core language. I had hoped for this to happen for some time, given that so much spectral analysis work has to solve this kind of problem.
So I take the opportunity of your message to suggest to Exelis to consider the inclusion of a compiled version of BVLS (http://www.netlib.org/lawson-hanson/all) in the core language.
Cheers,
Michele
On Tuesday, December 17, 2013 6:32:00 AM UTC, JP wrote:
> Hi IDLers,
>
> I am currently using BVLS (http://www-astro.physics.ox.ac.uk/~mxc/idl/bvls.pro) for spectral unmixing. It works great, but, on one pixel at a time. I want to implement for many pixels without having to loop. Example below:
>
>
>
> IDL> help, a
>
> A FLOAT = Array[85, 3]
>
> IDL> help, b
>
> B FLOAT = Array[85]
>
> IDL> help, bnd
>
> BND FLOAT = Array[2, 3]
>
> IDL> print, bnd
>
> 0.000000 1.00000
>
> 0.000000 1.00000
>
> 0.000000 1.00000
>
>
>
>
>
> bvls, A, B, BND, X_BVLS
>
>
>
> IDL> help, x_bvls
>
> X_BVLS FLOAT = Array[3]
>
>
>
> in my example A is a vector of "endmembers" (85 spectral bands and 3 fractions), B is a pixel (vector of 85 bands), BND are bounds (don't go negative nor >1) and the X_BVLS are the estimated fractions returned for that pixel.
>
> Now as you could imagine I have many many pixels (n) (my B is really a 2D array of [85, n]) and i've written a function which loops through n, but it gets very slow for large n.
>
>
>
> Any ideas?
>
>
>
> Thanks
|
|
|
|
|
|
| Re: BVLS (NNLS) without loops? [message #87189 is a reply to message #87178] |
Tue, 14 January 2014 17:19 |
JP
Messages: 55
Registered: April 2008
|
Member |
|
|
Thanks Michele
JP
On Wednesday, 15 January 2014 02:42:43 UTC+11, Mike wrote:
> Hi Juan Pablo,
>
>
>
> I don't see any obvious way to speed up your problem.
>
>
>
> However I take the opportunity to suggest the inclusion of a compiled version of BVLS (or NNLS) into a future version of IDL. I have been using that routine for years in a numbers of problems. I keep missing a faster built-in IDL version.
>
>
>
> Cheers,
>
>
>
> Michele
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
