[SciPy-user] Sparse with fast element-wise multiply?

[David Warde-Farley]

On 16-Dec-07, at 2:03 PM, lorenzo bolla wrote:

> converting the matrix from lil_matrix to csc or csr format gives a  
> 140x speed improvement on multiplication.
>
> In [4]: A = scipy.sparse.lil_matrix((1000, 1000))
>
> In [5]: A[:,100] = scipy.rand(1000)
>
> In [6]: A[99, :] = scipy.rand(1000)
>
> In [7]: A.setdiag(scipy.rand(1000))

Hi Lorenzo,

Thanks for your reply. Actually, though, I'm not looking for matrix- 
multiply; I know CSR/CSC is faster for that.

I'm looking for elementwise multiply, i.e. take two matrices of the  
same size, multiply each element in matrix 1 with the corresponding  
element in matrix 2, and put the result in the same position in the  
result matrix (as with matlab's a .* b syntax). It seems lil_matrix is  
the only sparse type that implements .multiply(), and converting to  
lil_matrix is slow, so I've written my own function for multiplying  
two coo_matrix's together. It's a LITTLE slower than .multiply() but  
faster than converting both matrices to lil_matrix and then multiplying.

I'm not entirely sure how CSR and CSC work so it might be possible to  
implement faster elementwise multiplication on top of them.

David