Given the code snippet:
B = A @ M - T
where A is a CSR scipy sparse matrix, M and T are two numpy arrays.
Question: During the matrix operations, does numpy treat A as a dense matrix, or M and T as two sparse matrices?
I suspect that the latter case is true since the resulting matrix B is not in the sparse format.
I also notices that this operation is much slower if I change the format of A to dense, which sort of contradicts my guess.
Numpy doesn't do sparse matrices. Scipy does the matrix multiplication (this means no multithreading, unlike numpy).
A is kept sparse but A @ M fills a dense array if M is a dense array.
>>> import numpy as np
>>> from scipy import sparse
>>> A = sparse.random(100, 10, density=0.1, format='csr')
>>> B = np.random.rand(10, 10)
>>> type(A@B)
<class 'numpy.ndarray'>
>>> type([email protected])
<class 'numpy.ndarray'>
Note that some sparse operations still give matrixes, not arrays:
>>> N = sparse.random(100, 10, density=0.1, format='csr')
>>> type(A@B - N)
<class 'numpy.matrix'>
Athe sparse matrix controls the operation. It does one thing if the other object is sparse, another if dense (with different result type) – hpaulj@like*passes the task tosparse.__mul__, which in turn calls one of several methods depending on what theotheris. Follow the[source]link on this page if you want more details: docs.scipy.org/doc/scipy/reference/generated/… – hpaulj