Element-wise power of scipy.sparse matrix

Question

How do I raise a scipy.sparse matrix to a power, element-wise? numpy.power should, according to its manual, do this, but it fails on sparse matrices:

>>> X
<1353x32100 sparse matrix of type '<type 'numpy.float64'>'
        with 144875 stored elements in Compressed Sparse Row format>

>>> np.power(X, 2)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File ".../scipy/sparse/base.py", line 347, in __pow__
    raise TypeError('matrix is not square')
TypeError: matrix is not square

Same problem with X**2. Converting to a dense array works, but wastes precious seconds.

I've had the same problem with np.multiply, which I solved using the sparse matrix's multiply method, but there seems to be no pow method.

Heraclid

I'm not familiar enough with numpy to tell you the answer, but your code doesn't contradict the documentation. The second argument to numpy.power should not be a number, but another matrix. – Ichnography

It looks like it's calling the object's __pow__ method, which is attempting to square the entire matrix, rather than doing it elementwise. That fails because, as it says, the matrix isn't square. – Starla

@RoundTower: actually, the second element should be an array, which it not the same as a matrix in Numpy, but a scalar is the same as a (1,) or (1,1) array for the purposes of numpy.power on a dense array. – Heraclid

Answer 1

This is a little low-level, but for element-wise operations you can work with the underlying data array directly:

>>> import scipy.sparse
>>> X = scipy.sparse.rand(1000,1000, density=0.003)
>>> X = scipy.sparse.csr_matrix(X)
>>> Y = X.copy()
>>> Y.data **= 3
>>> 
>>> abs((X.toarray()**3-Y.toarray())).max()
0.0

Brokenhearted

Answer 2

I just ran into the same question and find that sparse matrix now supports element-wise power. For the case above, it should be:

 X.power(2)

Floatstone

Is this for python 2.7? – Splanchnology

this doesn't seem to be true for python 3.6 – Bromism

it might depend on the format of a sparse matrix, for csr it seems to work – Macadam