[mlpack-git] [mlpack] mlpack KMeans class much slower than armadillo kmeans() (#514)

[jciminogh]

I seem to be getting much faster clustering from armadillo's [kmeans()](http://armasourceforgenet/docshtml#kmeans) function in comparison to mlpack's kmeans::KMeans<> class, which is about 2x to 6x slower Using the latest mlpack code from the git repo Am I doing something wrong?

Using the code below I get the following timings on my machine (Intel i5, 64 bit, g++ version 53)

Compiled without openmp:
`g++ kmeans_testcpp -o kmeans_test -O3 -std=c++11 -larmadillo -lmlpack`
mlpack_kmeans time: 173024
arma::kmeans time: 916399

Compiled with openmp:
`g++ kmeans_testcpp -o kmeans_test -O3 -std=c++11 -larmadillo -lmlpack -fopenmp`
mlpack_kmeans time: 177575
arma::kmeans time: 287675

```
#include <fstream>
#include <mlpack/methods/kmeans/kmeanshpp>
#include <armadillo>
    
int main() {
    arma::uword dims = 20; // number of dimensions
    arma::uword samples = 5000000;
    
    arma::uword max_iterations = 10;
    arma::uword k = 10;  // number of clusters
    
    arma::arma_rng::set_seed_random(); // random start
    
    std::cout << "Generating some synthetic data " << std::endl;
    
    arma::mat data(dims, samples, arma::fill::zeros);
    
    // generate data with unique centroids, added with a small amount of noise
    for (arma::uword i=0; i<samples; i++) {
        arma::uword c = as_scalar( arma::randi<arma::uvec>(1, arma::distr_param(0,k-1)) );
        datacol(i) = arma::linspace<arma::vec>(c, c+dims-1, dims) + 025*arma::randn<arma::vec>(dims);
      }

    arma::wall_clock timer;
    
    std::cout << "mlpack_kmeans start " << std::endl;
    
    arma::Row<size_t> mlpack_assignments;
    arma::mat mlpack_centroids;
    
    mlpack::kmeans::KMeans<> mlpack_kmeans(max_iterations);
    
    timertic();
    mlpack_kmeansCluster(data, k, mlpack_assignments, mlpack_centroids);
    
    std::cout << "mlpack_kmeans time: " << timertoc() << std::endl;
    
    std::cout << "---" << std::endl;
    std::cout << "arma::kmeans start " << std::endl;
    
    arma::mat arma_centroids;
    
    timertic();
    arma::kmeans(arma_centroids, data, k, arma::random_subset, max_iterations, false);
    
    std::cout << "arma::kmeans time: " << timertoc() << std::endl;
    
    std::cout << "---" << std::endl;
    
    mlpack_centroidsprint("mlpack_centroids:");
    arma_centroidsprint("arma_centroids:");
    
    return 0;
}
```


---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/issues/514
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20160201/01f0a34a/attachment.html>