[Numpy-discussion] ndarray of matrices (original) (raw)

Pau Gargallo pau.gargallo at gmail.com
Thu Jun 8 08:42:47 EDT 2006

On 6/8/06, Alexandre Guimond <alexandre.guimond at mirada-solutions.com> wrote:

Hi all. i work mainly with "volume" (3d) images, and numpy.ndarray answers most of my needs (addition of images, etc.). The problem I'm faced now with is that I have images of matrices and vectors and would like that when I do imageofmatrices * imageofvector is does the dot product of each of my matrices with all of my vectors, and when I do imageofmatrices.mean() it gives me the mean matrix. Basically, I want the same functionalities that are currently provided with scalars, but applied to matrices. It seems that a nice way of doing this is to have and ndarray of numpy.matrix, but this isn't supported it seems. Can anyone recommend a good way of implementing this? I'm new with the numpy thing and I'm not sure if subclassing ndarray is a good idea since I'll have to overload all the operators and i don't believe this will result in a very fast implementation, but I might be mistaken. Another possibility may be to create a new dtype for numpy.matrix, but I don't know if this is possible. Anyone have recommandations? Thx for any help.

We are several of us wondering which is the best way to do this kind of things. We were discussing this before (http://aspn.activestate.com/ASPN/Mail/Message/numpy-discussion/3130104), and some solutions were proposed, but we still don't have the definitive answer.

Building arrays of matrices objects will be too inefficient. For me the best thing would be to have n-dimensional universal functions, but this don't exist yet.

Meanwhile, I am using the following code (which is not the solution):

from numpy import *

nz,ny,nx = 1,1,1

im_of_mat = rand( nz, ny, nx, 3,3 ) im_of_vec = rand( nz, ny, nx, 3 )

im_of_products = ( im_of_mat * im_of_vec[...,newaxis,:] ).sum(axis=-1)

test that everything it's ok

for m,v,p in zip(im_of_mat.reshape(-1,3,3), im_of_vec.reshape(-1,3), im_of_products.reshape(-1,3)): assert allclose( dot(m,v), p )

pau

More information about the NumPy-Discussion mailing list