Thread Subject: Faster sparse from triplets

I have written my own version of repmat for use on sparse arrays,
because the default implementation creates big ones arrays for which
there is insufficient memory due to the dimensions involved.

My implementation basically involves repmatting the results of
find(...) so I end up with two index vectors and a values vector
representing the new repmatted array.

The problem comes when trying to convert those vectors into a sparse
array with sparse(x, y, values). It just takes far too long.

The dimensions of the vectors are in the region of 18e6 and the sparse
density is around 0.2. sprand can create a sparse of the same
dimensions and density in no time at all. How can I get sparse(x, y,
values) to speed up??

Have tried CXSparse's cs_sparse but no noticeable improvement.

Thanks.

On 7 Feb, 16:42, Pietro <pietro...@gmail.com> wrote:
> I have written my own version of repmat for use on sparse arrays,
> because the default implementation creates big ones arrays for which
> there is insufficient memory due to the dimensions involved.
>
> My implementation basically involves repmatting the results of
> find(...) so I end up with two index vectors and a values vector
> representing the new repmatted array.
>
> The problem comes when trying to convert those vectors into a sparse
> array with sparse(x, y, values). It just takes far too long.
>
> The dimensions of the vectors are in the region of 18e6 and the sparse
> density is around 0.2. sprand can create a sparse of the same
> dimensions and density in no time at all. How can I get sparse(x, y,
> values) to speed up??
>
> Have tried CXSparse's cs_sparse but no noticeable improvement.
>
> Thanks.

I had a similar problem and attempted a similar solution a while ago.
Its posted below if you want to compare. Unfortunately when I tried it
with the dimensions/density you mentioned it was way too slow to be
acceptable.

function B = sprepmat(A, M, N)
% SPREPMAT Space efficient repmat for sparse matrices
%
% A the matrix to be tiled
% M the number of times to tile vertically
% N the number of times to tile horizontally
%
[m, n] = size(A);
[x, y, values] = find(A);
if(m == 1)
x = x';
y = y';
values = values';
end
nz = length(values);

i = (0:nz * M * N - 1)';
B = sparse( x(mod(i, nz) + 1) + m * floor(i / (nz * N)), ...
y(mod(i, nz) + 1) + n * floor(mod(i, nz * N) / nz), ...
values(mod(i, nz) + 1) );

Pietro <pietromas@gmail.com> wrote in message
<b3058c25-ab52-422b-b0e2-9d1a87124808@e23g2000prf.googlegroups.com>...
> I have written my own version of repmat for use on sparse
arrays,
> because the default implementation creates big ones arrays
for which
> there is insufficient memory due to the dimensions involved.
>
> My implementation basically involves repmatting the results of
> find(...) so I end up with two index vectors and a values
vector
> representing the new repmatted array.
>
> The problem comes when trying to convert those vectors
into a sparse
> array with sparse(x, y, values). It just takes far too long.
>
> The dimensions of the vectors are in the region of 18e6
and the sparse
> density is around 0.2. sprand can create a sparse of the same
> dimensions and density in no time at all. How can I get
sparse(x, y,
> values) to speed up??
>
> Have tried CXSparse's cs_sparse but no noticeable improvement.
>
> Thanks.

Try CHOLMOD's "sparse2" function. It's much faster than
MATLAB's "sparse", particularly when the indices are unsorted.