numpy-discussion

[Numpy-discussion] convolve2d

[Michel S. <sa...@sc...>]

Hello,

I started using numarray for doing 2D convolutions on images.  I noticed 
that

import numarray.examples.convolve.high_level as convolve
convolve.Convolve2d(kernel, in, out)

only works on square images.  For images that are not square I get lots 
of noise in the background.

Also I was wondering is using the high_level API is most efficient?  
Currently my image is a Numeric
array (grabbed from the OpenGL frame buffer) which I convert to a 
numarray to do the convolution
and back to a Numeric array.  In the future I hope to completely replace 
Numeric by numarray.

Thanks for any input

-Michel

-- 

-----------------------------------------------------------------------
   o
  /    Michel F. Sanner Ph.D.            The Scripps Research Institute
o      Associate Professor               Department of Molecular Biology
  \                                      10550 North Torrey Pines Road
   o   Tel. (858) 784-2341               La Jolla, CA 92037
  /    Fax. (858) 784-2860
o     sa...@sc...                 http://www.scripps.edu/~sanner
-----------------------------------------------------------------------

Re: [Numpy-discussion] convolve2d

[Todd M. <jm...@st...>]

On Tue, 2004-06-22 at 11:51, Michel Sanner wrote:
> Hello,
> 
> I started using numarray for doing 2D convolutions on images.  I noticed 
> that
> 
> import numarray.examples.convolve.high_level as convolve
> convolve.Convolve2d(kernel, in, out)
> 
> only works on square images.  For images that are not square I get lots 
> of noise in the background.

I looked at the code and couldn't find any row/column typos.  In theory,
Convolve2d works for non-square arrays.

What do kernel.info() and data.info() say?

> Also I was wondering is using the high_level API is most efficient?  

For CPU, my guess is yes.  The point access macros and functions are
definitely slower than the high level API in any scenario where you're
not running out of memory.  The 1D API improves space usage but is less
efficient in time.  The Numeric compatible API is layered over the high
level API.

> Currently my image is a Numeric
> array (grabbed from the OpenGL frame buffer) which I convert to a 
> numarray to do the convolution and back to a Numeric array.  

It's also possible something is happening during this conversion.  It'd
be good to round-trip the Numeric array and make sure the end product
looks like the original.

> In the future I hope to completely replace Numeric by numarray.

Great!

Regards,
Todd

Re: [Numpy-discussion] convolve2d

[Todd M. <jm...@st...>]

On Tue, 2004-06-22 at 11:51, Michel Sanner wrote:
> Hello,
> 
> I started using numarray for doing 2D convolutions on images.  I noticed 
> that
> 
> import numarray.examples.convolve.high_level as convolve
> convolve.Convolve2d(kernel, in, out)
> 

Here's another (quick?) test which might help localize the problem.  Do:

convolve.Convolve2d(kernel, in, out, fft=1)

The underlying Convolve2d implementations are very different depending
on the use of FFT.

Regards,
Todd