python - Scipy: Do sparse matrices support advanced indexing?

Question

No problem:

>>> t = np.array([[1,1,1,1,1],[2,2,2,2,2],[3,3,3,3,3],[4,4,4,4,4],[5,5,5,5,5]])
>>> x = np.arange(5).reshape((-1,1)); y = np.arange(5)
>>> print (t[[x]],t[[y]])

Big problem:

>>> s = scipy.sparse.csr_matrix(t)
>>> print (s[[x]].toarray(),s[[y]].toarray())
Traceback (most recent call last):
  File "<pyshell#22>", line 1, in <module>
:               :
:               :
ValueError: data, indices, and indptr should be rank 1

s.toarray()[[x]] works great, but defeats the whole purpose of me using sparse matrices as my arrays are too big. I've checked the Attributes and Methods associated with some of the sparse matrices for anything referencing Advanced Indexing, but no dice. Any ideas?

Answer 1

sparse matrices have a very limited indexing support, and what is available depends on the format of the matrix.

For example:

>>> a = scipy.sparse.rand(100,100,format='coo')
>>> a[2:5, 6:8]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: 'coo_matrix' object has no attribute '__getitem__'

but

>>> a = scipy.sparse.rand(100,100,format='csc')
>>> a[2:5, 6:8]
<3x2 sparse matrix of type '<type 'numpy.float64'>'
    with 0 stored elements in Compressed Sparse Column format>

although

>>> a[2:5:2, 6:8:3]
Traceback (most recent call last):
...
ValueError: slicing with step != 1 not supported

There is also

>>> a = scipy.sparse.rand(100,100,format='dok')
>>> a[2:5:2, 6:8:3]
Traceback (most recent call last):
...
NotImplementedError: fancy indexing supported over one axis only
>>> a[2:5:2,1]
<3x1 sparse matrix of type '<type 'numpy.float64'>'
    with 0 stored elements in Dictionary Of Keys format>

And even

>>> a = scipy.sparse.rand(100,100,format='lil')
>>> a[2:5:2,1]
<2x1 sparse matrix of type '<type 'numpy.int32'>'
    with 0 stored elements in LInked List format>
C:Python27libsite-packagesscipysparselil.py:230: SparseEfficiencyWarning: Indexing into a lil_matrix with multiple indices is slow. Pre-converting to CSC or CSR beforehand is more efficient.
  SparseEfficiencyWarning)
>>> a[2:5:2, 6:8:3]
<2x1 sparse matrix of type '<type 'numpy.int32'>'
    with 0 stored elements in LInked List format>