Hello folks,

I've been using Eigen 3.0 to implement EM using a huge sparse matrix.
At that time of being I was using DynamicSparseMatrix as it was the only (unofficial) sparse matrix available in Eigen.
Now I see it got deprecated in Eigen 3.1.
But if I replace it by SparseMatrix my old code runs like 50-500 times slower.
I boiled it down to the coeff() and coeffRef() calls I am using in this loop. I will post a snippet of my code here (tDist and count are sparse matrices, there is one more outer loop over i):

Code: Select all

 82     // iterate over bag of words of the two aligned sentences
 83     // - collect total counts
 84     double v;
 85     for (int d=0;d<dstSent[i].size();d++) {
 86       for (int s=0;s<srcSent[i].size();s++) {
 87         const int& dst=dstSent[i][d];
 88         const int& src=srcSent[i][s];
 89         v=firstIter?(1.0/(double)tDist.cols()):(tDist.coeff(src,dst));
 90         total_s[dst]+=v;
 91       }
 92     }
 93 
 94     // iterate over bag of words of the two aligned sentences
 95     // - normalize counts
 96     double q;
 97     for (int d=0;d<dstSent[i].size();d++) {
 98       for (int s=0;s<srcSent[i].size();s++) {
 99         const int& dst=dstSent[i][d];
100         const int& src=srcSent[i][s];
101         q=(firstIter?(1.0/(double)tDist.cols()):(tDist.coeff(src,dst)))/total_s[dst];
102         count.coeffRef(src,dst)+=q;
103         total[src]+=q;
104       }
105     }


It should be possible to rewrite the code in order to NOT use coeff/coeffRef but I am not sure what would be most promising (Triplets?)
By the way, I am reading and writing the sparse matrices to disk using this code:

Code: Select all

 38   for (int k=0;k<t.outerSize();k++)
 39   {
 40     for (Eigen::SparseMatrix<double,Eigen::RowMajor>::InnerIterator it(t,k); it; ++it)
 41     {
 42       out<<it.value()<<" "<<it.row()<<" "<<it.col()<<std::endl;
 43     }
 44   }


It is horribly slow, in fact it takes almost as long as one iteration of my EM training. Is there a better way to dump sparse matrices to disk/load them?

cheers & thank you
katakombi

Using .coeffRef() to assemble a SparseMatrix is fine as soon as you can predict the upper bound of the number of non-zeros per column, e.g. if it is constant:

mat.reserve( VectorXi(nnzPerCol, mat.outerSize()) );

If it is not constant at all, you can fill a VectorXi to send to reserve.

In doubt, the easiest way to be sure to get nearly optimal performance is to use a triplet list as recommended in the documentation.

One more hint: if the number of calls to coeffRef(i,j) += ... with the same i,j pair is quite important then using coeffRef is probably the best.

For IO, we have in unsupported/Eigen/SparseExtra saveMarket and loadMarket functions.

Thanks a lot!
I am going to try it as soon as possible...

Hi there,

Ive found that slow writing of matrices was basically not Eigen-related at all.
I wasnt aware of std::endl - quite contrarly to '\n' - flushes the output buffer. this results in like 6-10x slowdown when I write large matrices

greets
Stefan

Dear ggael,

thanks for your help!
After fiddling around for some time I have abandoned my plans to rewrite my code from DynamicSparseMatrix to SparseMatrix.
The thing is I construct a sparse matrix and really need random access (I can't precompute triplets since I accumulate counts in no predefined order). Quite sadly so, I also don't have a guess about how many entries per column will be needed as it may vary between zero and e.g. 5%
of all existing values.
However, my code runs considerably fast using the DynamicSparseMatrix. So, is there any chance that the officially supported SparseMatrix will support these much faster coeff/coeffRef accessors in future? Would it be technically feasable?

cheers
Stefan

PS: could you please explain what you meant here?

One more hint: if the number of calls to coeffRef(i,j) += ... with the same i,j pair is quite important then using coeffRef is probably the best.

I just realized that I can accumulate counts also using setFromTriplets! Great, that way I should be able to rewrite my code...