Sparse matrix with fast access

Question

While working with large SciPy CSR sparse matrices I noticed that slicing the matrix to get a single row from the matrix was very slow as it seems to make a copy.

Is there any way to make a sparse matrix that takes a reference of the existing row instead of copying it, perhaps there is a more fitting implementation than CSR matrix?

What I need for my implementation is fast lookup for elements and rows and fast lookup of all non zero indices of a vector. I never need to change the matrix in any way or perform other operations on the matrix.

Answer 1

You can take advantage of the CSR representation to slice the underlying arrays directly and share the data with a new CSR matrix:

mat = # some CSR matrix
i = # the index of whatever row you want
start, stop = mat.indptr[i], mat.indptr[i+1]
noncopy_row_i = scipy.sparse.csr_matrix((mat.data[start:stop],
                                         mat.indices[start:stop],
                                         numpy.array([0, stop-start])),
                                        shape=(1, mat.shape[1]))

Answer 2

Numpy supports different types os sparces matrices: https://docs.scipy.org/doc/scipy/reference/sparse.html#usage-information

May be coo_matrix will give faster elements lookup, but also you can loose on some other operations. I think the best way is to benchmark on your data and algorythms.