How to convert a scipy row matrix into a numpy array

Question

Consider the following example:

import numpy as np
import scipy.sparse
A = scipy.sparse.csr_matrix((2,2))
b = A.sum(axis=0)

The matrix b now has the form

matrix([[ 0., 0.]])

However, I'd like it to become an array like this:

array([ 0., 0.])

This can by done by b = np.asarray(b)[0], but that does not seem very elegant, especially compared to MATLAB's b(:). Is there a more elegant way to do this?

• Warren Weckesser: See stackoverflow.com/questions/26576524/…; in particular, note that the sparse matrix has the method .toarray(), the result of which you could reshape into a one-dimensional numpy array.
• hpaulj: MATLAB's b(:) - what's the size? Still 2d isnt' it? (:) is a bit like a np.ravel, except MATLAB is always 2d or higher.

Answer 1

b.A1 will do the job.

In [83]: A
Out[83]: 
<2x2 sparse matrix of type '<class 'numpy.float64'>'
 with 0 stored elements in Compressed Sparse Row format>
In [84]: A.A
Out[84]: 
array([[ 0., 0.],
 [ 0., 0.]])
In [85]: b=A.sum(axis=0)
In [86]: b
Out[86]: matrix([[ 0., 0.]])
In [87]: b.A1
Out[87]: array([ 0., 0.])
In [88]: A.A.sum(axis=0) # another way
Out[88]: array([ 0., 0.])

You can up vote this, or add to my top grossing answer here: Numpy matrix to array :)

A is a sparse matrix. Sparse sum is performed with a matrix product (an appropriate matrix of 1s). The result is a dense matrix.

Sparse matrix has a toarray() method, with a .A shortcut.

Dense matrix also has those, but it also has a .A1 (poorly documented - hence all my hits), which flattens as well.

The doc for A1:

Return `self` as a flattened `ndarray`.
Equivalent to ``np.asarray(x).ravel()``

In fact the code is

return self.__array__().ravel()

====================

Is MATLAB b(:) really the equivalent?

A(:) is all the elements of A, regarded as a single column.

If I read that correctly, the numpy equivalent is a transpose, or b.ravel().T. The shape would be (2,1). But in MATLAB a column matrix is the simplest form of matrix.

In [94]: b.T
Out[94]: 
matrix([[ 0.],
 [ 0.]])

(I'm an old MATLAB programmer, with Octave on my standby computer. And a copy of 3.5 on some old Windows disk. :) ).

• Kurt Peek: Indeed, accoding to docs.scipy.org/doc/numpy/reference/generated/…, .A1 is equivalent to np.asarray(x).ravel(), and seems like the most concise option.

Answer 2

There are different options here. For example, you could start by converting matrix b to a 2D array. Then you'll need to transform it into a 1D array. This can be easily accomplished through NumPy's squeeze or reshape:

In [208]: np.asarray(b).squeeze()
Out[208]: array([ 0., 0.])
In [209]: np.asarray(b).reshape((b.size,))
Out[209]: array([ 0., 0.])

Alternatively, you could convert A to an array as suggested in @Warren Weckesser's comment. This would make it unnecessary to further convert b:

In [210]: A.toarray().sum(axis=0)
Out[210]: array([ 0., 0.])