Hi,

I would like to concatenate two sparse matrices or sparse+dense matrices like as follows as in matlab.

A=sparse(A);
B=sparse(B);
C = [A; B]

A=sparse(A);
B=full(B);
C= [A; B]

is there any way ?
thanks in advance.

Best Regards,
Yeonchool Park

nobody knows ?
please somebody help me.

Probably something like this should work:

Code: Select all

matrix c(w1+w2,h);
c.submatrix(0,0,w1,h)=a;
c.submatrix(w1,0,w2,h)=b;

But the exact syntax is obviously wrong (parameter order etc). I am not an expert for Eigen.

Thanks dzenanz,

However, unfortunately, Eigen sparse seems not to support yet block assignment. that's the problem.

You can do it manually:

C.resize(A.rows, A.cols()+B.cols());
C.middleCols(0,A.cols()) = A;
C.middleCols(A.cols(),B.cols()) = B;

If B is dense, then call B.sparseView() to remove the explicit zeros.

And you know, writing a block copy function should be quite simple. Just two canonical loops to iterate and copy the block.

Thanks ggael,

I will try your approach. it looks nice.
I didn't know that Eigen has this kind of function. where can I find this kind of information ? from Classes pages ?
I am not used to reading Classes pages.

Yeonchool

ggael wrote:You can do it manually:
C.resize(A.rows, A.cols()+B.cols());
C.middleCols(0,A.cols()) = A;
C.middleCols(A.cols(),B.cols()) = B;
If B is dense, then call B.sparseView() to remove the explicit zeros.

Thanks zoharl,

however I doubted the performance of my implementation because I need quite fast performance.
so, if Eigen had any function, I believed it could be more faster than my implementation.
from now on, I am going to check Eigen source code too to see if there is possibility of improving performance.

have a nice weekend

zoharl wrote:And you know, writing a block copy function should be quite simple. Just two canonical loops to iterate and copy the block.

Hi ggael,

I tested your approach. and, it works. however it has two weird things.

First, when I added reserve function, no values were assigned in the matrix(nonzeros() = 0) while it worked well without the function.
it seems to be a bug.

Code: Select all

   SpMat B(4, MGr+WUr);   //SpMat B=[Z;WU]   Z: null matrix (MGr x 4)
   B.reserve(Eigen::VectorXi::Constant(B.outerSize(), WU.nonZeros()/B.outerSize()));             //  <-- this line
   B.middleCols(MGr, WUr) = ((WU.transpose()).eval()).sparseView();     // I couldn't use middleRows() because B is a column-major matrix
   B = (B.transpose()).eval();


Second, speed.
computation time was slower than copying each element like as follows (2ms: the below code vs. 8ms: the above code).
I think it is caused by transpose() and sparseView() function.

Code: Select all

   for (int kA=0; kA<WU.outerSize(); ++kA)
   {
      for (Eigen::MatrixXf::InnerIterator itA(WU,kA); itA; ++itA)
      {
         const unsigned int   iA = itA.row(),    jA = itA.col();
         B.insert(iA+WUr,jA) = itA.value();
      }
   }


[Question] is it possible to copy one rows or one cols of Eigen matrix by using memcpy function ?

Best Regards,
Yeonchool

ggael wrote:You can do it manually:
C.resize(A.rows, A.cols()+B.cols());
C.middleCols(0,A.cols()) = A;
C.middleCols(A.cols(),B.cols()) = B;
If B is dense, then call B.sparseView() to remove the explicit zeros.

oh, I did not see that you wanted a vertical arrangement. Then you indeed need to do it with a manual for loop. Sparse transposition is expensive.

Also note that reserve only reserve memory. It does not create nonzeros.

However, I have some doubts about what you are trying to achieve. What's the point of inserting a dense matrix into a sparse representation? If you can, I would rather try to keep them separated even if that complexify a little the rest of the code. Dense representations are an order of magnitude faster than sparse one (unless the matrix is big and very sparse of course).

Interesting. I want to solve a similar problem, but
cannot even find the middleCols() as well as the other block matrix operation
functions under SparseMatrix class.

I am using the most recent eigen package.

You won't find them in the SparseMatrix.h file, but if you are using Eigen3.2, then they are accessible and documented: http://eigen.tuxfamily.org/dox/classEig ... atrix.html.

Make sure you are including Eigen 3.2 headers and not another one.

Thank you for the reply, but it is still not clear to me how to get access to those block functions.

The following is a simple test code I just wrote. I also use -I to include /users/dwang/eigen/.
I thought I should be able to use sm.block(..) after sm is declared, but I cannot,. sm does not
have have that function. Anything else should I do? Thanks!

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~`
#include <iostream>
#include <Eigen/Sparse>
#include <Eigen/Eigen>
#include <Eigen/Core>

using namespace Eigen;
using namespace std;
typedef Eigen::SparseMatrix<double> SpMat;

int main(int argc, char ** argv)
{
SpMat sm (100,100);
//I hope to call sm.block(...) here, but sm does not have block(...)
}

Try this:

Code: Select all

#include <Eigen/Sparse>

#if EIGEN_MAJOR_VERSION<2
#error wrong eigen version
#endif

typedef Eigen::SparseMatrix<double> SpMat;

int main(int argc, char ** argv)
{
  SpMat A(100,100), B;
  B = A.block(10,10,10,10);
}


Ah, I found that even though that function is not visible under eclipse,
the program still runs without any issue.

problem solved!