Cannot create NeighborSearch with arma:fmat MatType

[victor1234]

I cannot declare type:

using KnnModel = mlpack::neighbor::NeighborSearch<mlpack::neighbor::NearestNeighborSort,
                                                  mlpack::metric::ManhattanDistance,
                                                  arma::fmat,
                                                  mlpack::tree::KDTree>;

error:

/usr/include/mlpack/core/tree/hrectbound_impl.hpp:372:30: 
error: no matching function for call to 
‘arma::Col<double>::Col(const arma::Op<arma::subview<float>, arma::op_min>)’
   arma::vec mins(min(data, 1));

Is threre no implementation for float or I'm doing something wrong?

[rcurtin]

Hi there,

It turns out that I never wrote a test for this use case, and fixing this will take some more work than I had anticipated. I think that when I allowed arbitrary MatType, I was thinking of arma::mat and arma::sp_mat, but definitely the other matrix types are useful too! (See also #290) I've started the work in 3561ad7, but there are lots of other pieces that need to come together. I'd like to hope I can have this done in a few weeks (or sooner).

[victor1234]

I think you should note in library overview about incomplete float support. It was critical for me and I spent some time for library familiarization before know about it (API looks like very generic and I didn't expect it).

[versusvoid]

Similar problem in KMeans:

float mat[] = { 1, 2, 3, 4, 5, 6 };
const arma::fmat data(mat, 2, 3, false, true);

size_t number_of_clusters = 2;  

arma::Row<size_t> assignments(3);

arma::mat centroids(2, 3);

km::KMeans<mlpack::metric::EuclideanDistance, km::RandomPartition, 
           km::MaxVarianceNewCluster, km::NaiveKMeans, arma::fmat> k;

k.Cluster(data, number_of_clusters, assignments, centroids);

/usr/include/mlpack/methods/kmeans/kmeans_impl.hpp: In instantiation of ‘void mlpack::kmeans::KMeans<MetricType, InitialPartitionPolicy, EmptyClusterPolicy, LloydStepType, MatType>::Cluster(const MatType&, size_t, arma::Row<long unsigned int>&, arma::mat&, bool, bool) [with MetricType = mlpack::metric::LMetric<2, true>; InitialPartitionPolicy = mlpack::kmeans::RandomPartition; EmptyClusterPolicy = mlpack::kmeans::MaxVarianceNewCluster; LloydStepType = mlpack::kmeans::NaiveKMeans; MatType = arma::Mat<float>; size_t = long unsigned int; arma::mat = arma::Mat<double>]’:
test.cxx:44:65:   required from here
/usr/include/mlpack/methods/kmeans/kmeans_impl.hpp:239:37: error: no matching function for call to ‘arma::Col<double>::Col(const arma::subview_col<float>)’
       centroids.col(assignments[i]) += arma::vec(data.col(i));

[kartikdutt18]

This issue still exists. @ojhalakshya and I are taking this up.

[ojhalakshya]

Hey @rcurtin
I had been working on this issue for a while now, even after making a dozen of template changes I am not sure whether it is the right direction to move forward in.
I do have a branch for it currently, @kartikdutt18 has also made some changes there. If you see fit I can make a pr, or maybe you can first take a look at it.
Thanks.

[ojhalakshya]

@rcurtin how must we proceed in this issue.
Thanks.

[rcurtin]

@ojhalakshya this is not an easy issue to resolve. :) It will require templating everything involved with NeighborSearch to accept different types of matrices. I'd suggest starting from the simple example posted in the original issue, and making that work. I think many of the things in the tree infrastructure are already templatized, so, there may not be much work left to do. Consider taking the test case DualTreeVsNaive1 from knn_test.cpp and adapting that to use arma::fmat instead of arma::mat; once that works correctly, we can probably call the issue solved. :)

[ojhalakshya]

okay @rcurtin will see into these files, and try to make some things work out.
Will ping you if I need more info :)
Thanks.