First commit
authorBenjamin Auder <benjamin.auder@somewhere>
Mon, 6 Dec 2021 23:29:59 +0000 (00:29 +0100)
committerBenjamin Auder <benjamin.auder@somewhere>
Mon, 6 Dec 2021 23:29:59 +0000 (00:29 +0100)
.gitignore [new file with mode: 0644]
DESCRIPTION [new file with mode: 0644]
LICENSE [new file with mode: 0644]
R/NAMESPACE.R [new file with mode: 0644]
R/ectd.R [new file with mode: 0644]
R/nng.R [new file with mode: 0644]
README.md [new file with mode: 0644]
src/knncpp.h [new file with mode: 0644]
src/nng.cpp [new file with mode: 0644]

diff --git a/.gitignore b/.gitignore
new file mode 100644 (file)
index 0000000..57d912a
--- /dev/null
@@ -0,0 +1,9 @@
+/NAMESPACE
+/man/
+!/man/nngd-package.Rd
+/R/RcppExports.R
+/src/RcppExports.cpp
+/src/*.o
+/src/*.so
+.RData
+.Rhistory
diff --git a/DESCRIPTION b/DESCRIPTION
new file mode 100644 (file)
index 0000000..e882735
--- /dev/null
@@ -0,0 +1,14 @@
+Package: nngd
+Type: Package
+Title: Compute distances based on the nearest-neighbors graph.
+Version: 1.0
+Date: 2021-12-06
+Author: Benjamin Auder <benjamin.auder@universite-paris-saclay.fr> [aut,cre]
+Maintainer: Benjamin Auder <benjamin.auder@universite-paris-saclay.fr>
+Description: Only two functions for now. nng() builds the nearest-neighbors graph,
+             and ectd() computes the Euclidian Commute-Time Distances.
+             Only undirected graph case implemented so far.
+License: MIT + file LICENSE
+Imports: Rcpp, RcppEigen, igraph, pracma
+LinkingTo: Rcpp, RcppEigen
+RoxygenNote: 7.1.2
diff --git a/LICENSE b/LICENSE
new file mode 100644 (file)
index 0000000..9f4ff56
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,2 @@
+YEAR: 2021-2021
+COPYRIGHT HOLDER: Benjamin Auder
diff --git a/R/NAMESPACE.R b/R/NAMESPACE.R
new file mode 100644 (file)
index 0000000..f213ec4
--- /dev/null
@@ -0,0 +1,7 @@
+#' @useDynLib nngd, .registration = TRUE
+#'
+#' @importFrom igraph is_directed vcount as_edgelist components graph_from_edgelist
+#' @importFrom pracma pinv
+#' @import Rcpp
+#'
+NULL
diff --git a/R/ectd.R b/R/ectd.R
new file mode 100644 (file)
index 0000000..727bbd4
--- /dev/null
+++ b/R/ectd.R
@@ -0,0 +1,43 @@
+#' Compute the Euclidian Commute-Time Distances (ECTD) in an undirected graph.
+#'
+#' Assuming similarity function doesn't depend on x, and undirected graph.
+#'
+#' @param o Output of \code{nng}.
+#' @param similarity function distance --> similarity.
+#' @param inf.replace Replace Inf values by large finite number ?
+#'
+#' @return A distances matrix (n x n)
+#'
+#' @export
+ectd <- function(o, similarity=function(x) 1, inf.replace=TRUE)
+{
+  if (!igraph::is_directed(o$graph))
+    stop("Directed graph case unimplemented yet")
+  n <- igraph::vcount(o$graph)
+  E <- igraph::as_edgelist(o$graph, names=FALSE)
+  W <- matrix(0, nrow=n, ncol=n)
+  for (i in seq_len(nrow(E))) {
+    edge <- c(E[i,1], E[i,2])
+    W[edge[1], edge[2]] <- similarity(o$distances[i])
+    W[edge[2], edge[1]] <- W[edge[1], edge[2]]
+  }
+  L <- diag(rowSums(W)) - W
+  cc <- igraph::components(o$graph)
+  distances <- matrix(Inf, nrow=n, ncol=n)
+  for (i in seq_len(cc$no)) {
+    indices <- which(cc$membership == i)
+    L_loc <- L[indices, indices]
+    n_loc <- cc$csize[i]
+    if (n_loc >= 2) L_inv = pracma::pinv(L_loc)
+    else L_inv = matrix(ifelse(L_loc != 0, 1/L_loc[1], 0))
+    Lii = matrix(rep(diag(L_inv), each=n_loc), nrow=n_loc, byrow=TRUE)
+    Ljj = matrix(rep(diag(L_inv), each=n_loc), nrow=n_loc)
+    # https://math.stackexchange.com/questions/1321305/commute-time-distance-in-a-graph
+    distances[indices, indices] <- Lii + Ljj - 2 * L_inv
+  }
+  if (inf.replace) {
+    finiteMax <- max(distances[is.finite(distances)])
+    distances[is.infinite(distances)] <- finiteMax + 1
+  }
+  distances
+}
diff --git a/R/nng.R b/R/nng.R
new file mode 100644 (file)
index 0000000..6f3a34d
--- /dev/null
+++ b/R/nng.R
@@ -0,0 +1,19 @@
+#' Compute the nearest-neighbors graph + pairwise distances.
+#'
+#' This is just a wrapper around C++/Rcpp function.
+#'
+#' @param data data.frame or matrix.
+#' @param k Number of neighbors at each point.
+#' @param mutual Whether or not to build mutual kNN graph.
+#'
+#' @return A list with $graph and $distances.
+#'
+#' @export
+nng <- function(data, k=ceil(sqrt(nrow(data))), mutual=TRUE)
+{
+  mdata <- data
+  if (is.data.frame(data)) mdata <- as.matrix(data)
+  res <- findNeighbors(mdata, k, mutual)
+  list(graph = igraph::graph_from_edgelist(t(res$edges), directed = !mutual),
+       distances = sqrt(res$euc_dists))
+}
diff --git a/README.md b/README.md
new file mode 100644 (file)
index 0000000..8dba5e4
--- /dev/null
+++ b/README.md
@@ -0,0 +1,16 @@
+# nngd
+
+Build nearest-neighbors graph in a ~efficient way.
+
+Compute graph distances + ECTD distances.
+
+Use file knncpp.h from https://github.com/Rookfighter/knn-cpp :
+wget https://raw.githubusercontent.com/Rookfighter/knn-cpp/master/include/knncpp.h
+
+## Usage
+
+roxygenize(".")
+R CMD INSTALL .
+library(nngd)
+?nng
+?ectd
diff --git a/src/knncpp.h b/src/knncpp.h
new file mode 100644 (file)
index 0000000..05d98ec
--- /dev/null
@@ -0,0 +1,1726 @@
+/* knncpp.h
+ *
+ * Author:     Fabian Meyer
+ * Created On: 22 Aug 2021
+ * License:    MIT
+ */
+
+#ifndef KNNCPP_H_
+#define KNNCPP_H_
+
+#include <Eigen/Geometry>
+#include <vector>
+#include <map>
+#include <set>
+
+#ifdef KNNCPP_FLANN
+
+#include <flann/flann.hpp>
+
+#endif
+
+namespace knncpp
+{
+    /********************************************************
+     * Matrix Definitions
+     *******************************************************/
+
+    typedef typename Eigen::MatrixXd::Index Index;
+
+    typedef Eigen::Matrix<Index, Eigen::Dynamic, 1> Vectori;
+    typedef Eigen::Matrix<Index, 2, 1> Vector2i;
+    typedef Eigen::Matrix<Index, 3, 1> Vector3i;
+    typedef Eigen::Matrix<Index, 4, 1> Vector4i;
+    typedef Eigen::Matrix<Index, 5, 1> Vector5i;
+
+    typedef Eigen::Matrix<Index, Eigen::Dynamic, Eigen::Dynamic> Matrixi;
+    typedef Eigen::Matrix<Index, 2, 2> Matrix2i;
+    typedef Eigen::Matrix<Index, 3, 3> Matrix3i;
+    typedef Eigen::Matrix<Index, 4, 4> Matrix4i;
+    typedef Eigen::Matrix<Index, 5, 5> Matrix5i;
+
+    typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic> Matrixf;
+    typedef Eigen::Matrix<double, Eigen::Dynamic, Eigen::Dynamic> Matrixd;
+
+    /********************************************************
+     * Distance Functors
+     *******************************************************/
+
+    /** Manhatten distance functor.
+      * This the same as the L1 minkowski distance but more efficient.
+      * @see EuclideanDistance, ChebyshevDistance, MinkowskiDistance */
+    template <typename Scalar>
+    struct ManhattenDistance
+    {
+        /** Compute the unrooted distance between two vectors.
+          * @param lhs vector on left hand side
+          * @param rhs vector on right hand side */
+        template<typename DerivedA, typename DerivedB>
+        Scalar operator()(const Eigen::MatrixBase<DerivedA> &lhs,
+            const Eigen::MatrixBase<DerivedB> &rhs) const
+        {
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedA>::Scalar,Scalar>::value,
+                "distance scalar and input matrix A must have same type");
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedB>::Scalar, Scalar>::value,
+                "distance scalar and input matrix B must have same type");
+
+            return (lhs - rhs).cwiseAbs().sum();
+        }
+
+        /** Compute the unrooted distance between two scalars.
+          * @param lhs scalar on left hand side
+          * @param rhs scalar on right hand side */
+        Scalar operator()(const Scalar lhs,
+            const Scalar rhs) const
+        {
+            return std::abs(lhs - rhs);
+        }
+
+        /** Compute the root of a unrooted distance value.
+          * @param value unrooted distance value */
+        Scalar operator()(const Scalar val) const
+        {
+            return val;
+        }
+    };
+
+    /** Euclidean distance functor.
+      * This the same as the L2 minkowski distance but more efficient.
+      * @see ManhattenDistance, ChebyshevDistance, MinkowskiDistance */
+    template <typename Scalar>
+    struct EuclideanDistance
+    {
+        /** Compute the unrooted distance between two vectors.
+          * @param lhs vector on left hand side
+          * @param rhs vector on right hand side */
+        template<typename DerivedA, typename DerivedB>
+        Scalar operator()(const Eigen::MatrixBase<DerivedA> &lhs,
+            const Eigen::MatrixBase<DerivedB> &rhs) const
+        {
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedA>::Scalar,Scalar>::value,
+                "distance scalar and input matrix A must have same type");
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedB>::Scalar, Scalar>::value,
+                "distance scalar and input matrix B must have same type");
+
+            return (lhs - rhs).cwiseAbs2().sum();
+        }
+
+        /** Compute the unrooted distance between two scalars.
+          * @param lhs scalar on left hand side
+          * @param rhs scalar on right hand side */
+        Scalar operator()(const Scalar lhs,
+            const Scalar rhs) const
+        {
+            Scalar diff = lhs - rhs;
+            return diff * diff;
+        }
+
+        /** Compute the root of a unrooted distance value.
+          * @param value unrooted distance value */
+        Scalar operator()(const Scalar val) const
+        {
+            return std::sqrt(val);
+        }
+    };
+
+    /** General minkowski distance functor.
+      * The infinite version is only available through the chebyshev distance.
+      * @see ManhattenDistance, EuclideanDistance, ChebyshevDistance  */
+    template <typename Scalar, int P>
+    struct MinkowskiDistance
+    {
+        struct Pow
+        {
+            Scalar operator()(const Scalar val) const
+            {
+                Scalar result = 1;
+                for(int i = 0; i < P; ++i)
+                    result *= val;
+                return result;
+            }
+        };
+
+        /** Compute the unrooted distance between two vectors.
+          * @param lhs vector on left hand side
+          * @param rhs vector on right hand side */
+        template<typename DerivedA, typename DerivedB>
+        Scalar operator()(const Eigen::MatrixBase<DerivedA> &lhs,
+            const Eigen::MatrixBase<DerivedB> &rhs) const
+        {
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedA>::Scalar,Scalar>::value,
+                "distance scalar and input matrix A must have same type");
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedB>::Scalar, Scalar>::value,
+                "distance scalar and input matrix B must have same type");
+
+            return (lhs - rhs).cwiseAbs().unaryExpr(MinkowskiDistance::Pow()).sum();
+        }
+
+        /** Compute the unrooted distance between two scalars.
+          * @param lhs scalar on left hand side
+          * @param rhs scalar on right hand side */
+        Scalar operator()(const Scalar lhs,
+            const Scalar rhs) const
+        {
+            return std::pow(std::abs(lhs - rhs), P);;
+        }
+
+        /** Compute the root of a unrooted distance value.
+          * @param value unrooted distance value */
+        Scalar operator()(const Scalar val) const
+        {
+            return std::pow(val, 1 / static_cast<Scalar>(P));
+        }
+    };
+
+    /** Chebyshev distance functor.
+      * This distance is the same as infinity minkowski distance.
+      * @see ManhattenDistance, EuclideanDistance, MinkowskiDistance */
+    template<typename Scalar>
+    struct ChebyshevDistance
+    {
+        /** Compute the unrooted distance between two vectors.
+          * @param lhs vector on left hand side
+          * @param rhs vector on right hand side */
+        template<typename DerivedA, typename DerivedB>
+        Scalar operator()(const Eigen::MatrixBase<DerivedA> &lhs,
+            const Eigen::MatrixBase<DerivedB> &rhs) const
+        {
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedA>::Scalar,Scalar>::value,
+                "distance scalar and input matrix A must have same type");
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedB>::Scalar, Scalar>::value,
+                "distance scalar and input matrix B must have same type");
+
+            return (lhs - rhs).cwiseAbs().maxCoeff();
+        }
+
+        /** Compute the unrooted distance between two scalars.
+          * @param lhs scalar on left hand side
+          * @param rhs scalar on right hand side */
+        Scalar operator()(const Scalar lhs,
+            const Scalar rhs) const
+        {
+            return std::abs(lhs - rhs);
+        }
+
+        /** Compute the root of a unrooted distance value.
+          * @param value unrooted distance value */
+        Scalar operator()(const Scalar val) const
+        {
+            return val;
+        }
+    };
+
+    /** Hamming distance functor.
+      * The distance vectors have to be of integral type and should hold the
+      * information vectors as bitmasks.
+      * Performs a XOR operation on the vectors and counts the number of set
+      * ones. */
+    template<typename Scalar>
+    struct HammingDistance
+    {
+        static_assert(std::is_integral<Scalar>::value,
+            "HammingDistance requires integral Scalar type");
+
+        struct XOR
+        {
+            Scalar operator()(const Scalar lhs, const Scalar rhs) const
+            {
+                return lhs ^ rhs;
+            }
+        };
+
+        struct BitCount
+        {
+            Scalar operator()(const Scalar lhs) const
+            {
+                Scalar copy = lhs;
+                Scalar result = 0;
+                while(copy != static_cast<Scalar>(0))
+                {
+                    ++result;
+                    copy &= (copy - 1);
+                }
+
+                return result;
+            }
+        };
+
+        /** Compute the unrooted distance between two vectors.
+          * @param lhs vector on left hand side
+          * @param rhs vector on right hand side */
+        template<typename DerivedA, typename DerivedB>
+        Scalar operator()(const Eigen::MatrixBase<DerivedA> &lhs,
+            const Eigen::MatrixBase<DerivedB> &rhs) const
+        {
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedA>::Scalar,Scalar>::value,
+                "distance scalar and input matrix A must have same type");
+            static_assert(
+                std::is_same<typename Eigen::MatrixBase<DerivedB>::Scalar, Scalar>::value,
+                "distance scalar and input matrix B must have same type");
+
+            return lhs.
+                binaryExpr(rhs, XOR()).
+                unaryExpr(BitCount()).
+                sum();
+        }
+
+        /** Compute the unrooted distance between two scalars.
+          * @param lhs scalar on left hand side
+          * @param rhs scalar on right hand side */
+        Scalar operator()(const Scalar lhs,
+            const Scalar rhs) const
+        {
+            BitCount cnt;
+            XOR xOr;
+            return cnt(xOr(lhs, rhs));
+        }
+
+        /** Compute the root of a unrooted distance value.
+          * @param value unrooted distance value */
+        Scalar operator()(const Scalar value) const
+        {
+            return value;
+        }
+    };
+
+    /** Efficient heap structure to query nearest neighbours. */
+    template<typename Scalar>
+    class QueryHeap
+    {
+    private:
+        Index *indices_ = nullptr;
+        Scalar *distances_ = nullptr;
+        size_t maxSize_ = 0;
+        size_t size_ = 0;
+    public:
+        /** Creates a query heap with the given index and distance memory regions. */
+        QueryHeap(Index *indices, Scalar *distances, const size_t maxSize)
+            : indices_(indices), distances_(distances), maxSize_(maxSize)
+        { }
+
+        /** Pushes a new query data set into the heap with the given
+          * index and distance.
+          * The index identifies the point for which the given distance
+          * was computed.
+          * @param idx index / ID of the query point
+          * @param dist distance that was computed for the query point*/
+        void push(const Index idx, const Scalar dist)
+        {
+            assert(!full());
+
+            // add new value at the end
+            indices_[size_] = idx;
+            distances_[size_] = dist;
+            ++size_;
+
+            // upheap
+            size_t k = size_ - 1;
+            size_t tmp = (k - 1) / 2;
+            while(k > 0 && distances_[tmp] < dist)
+            {
+                distances_[k] = distances_[tmp];
+                indices_[k] = indices_[tmp];
+                k = tmp;
+                tmp = (k - 1) / 2;
+            }
+            distances_[k] = dist;
+            indices_[k] = idx;
+        }
+
+        /** Removes the element at the front of the heap and restores
+          * the heap order. */
+        void pop()
+        {
+            assert(!empty());
+
+            // replace first element with last
+            --size_;
+            distances_[0] = distances_[size_];
+            indices_[0] = indices_[size_];
+
+            // downheap
+            size_t k = 0;
+            size_t j;
+            Scalar dist = distances_[0];
+            Index idx = indices_[0];
+            while(2 * k + 1 < size_)
+            {
+                j = 2 * k + 1;
+                if(j + 1 < size_ && distances_[j+1] > distances_[j])
+                    ++j;
+                // j references now greatest child
+                if(dist >= distances_[j])
+                    break;
+                distances_[k] = distances_[j];
+                indices_[k] = indices_[j];
+                k = j;
+            }
+            distances_[k] = dist;
+            indices_[k] = idx;
+        }
+
+        /** Returns the distance of the element in front of the heap. */
+        Scalar front() const
+        {
+            assert(!empty());
+            return distances_[0];
+        }
+
+        /** Determines if this query heap is full.
+          * The heap is considered full if its number of elements
+          * has reached its max size.
+          * @return true if the heap is full, else false */
+        bool full() const
+        {
+            return size_ >= maxSize_;
+        }
+
+        /** Determines if this query heap is empty.
+          * @return true if the heap contains no elements, else false */
+        bool empty() const
+        {
+            return size_ == 0;
+        }
+
+        /** Returns the number of elements within the query heap.
+          * @return number of elements in the heap */
+        size_t size() const
+        {
+            return size_;
+        }
+
+        /** Clears the query heap. */
+        void clear()
+        {
+            size_ = 0;
+        }
+
+        /** Sorts the elements within the heap according to
+          * their distance. */
+        void sort()
+        {
+            size_t cnt = size_;
+            for(size_t i = 0; i < cnt; ++i)
+            {
+                Index idx = indices_[0];
+                Scalar dist = distances_[0];
+                pop();
+                indices_[cnt - i - 1] = idx;
+                distances_[cnt - i - 1] = dist;
+            }
+        }
+    };
+
+    /** Class for performing brute force knn search. */
+    template<typename Scalar,
+        typename Distance=EuclideanDistance<Scalar>>
+    class BruteForce
+    {
+    public:
+        typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic> Matrix;
+        typedef Eigen::Matrix<Scalar, Eigen::Dynamic, 1> Vector;
+        typedef knncpp::Matrixi Matrixi;
+    private:
+        Distance distance_ = Distance();
+        Matrix dataCopy_ = Matrix();
+        const Matrix *data_ = nullptr;
+
+        bool sorted_ = true;
+        bool takeRoot_ = true;
+        Index threads_ = 1;
+        Scalar maxDist_ = 0;
+
+    public:
+
+        BruteForce() = default;
+
+        /** Constructs a brute force instance with the given data.
+          * @param data NxM matrix, M points of dimension N
+          * @param copy if true copies the data, otherwise assumes static data */
+        BruteForce(const Matrix &data, const bool copy = false)
+            : BruteForce()
+        {
+            setData(data, copy);
+        }
+
+        /** Set if the points returned by the queries should be sorted
+          * according to their distance to the query points.
+          * @param sorted sort query results */
+        void setSorted(const bool sorted)
+        {
+            sorted_ = sorted;
+        }
+
+        /** Set if the distances after the query should be rooted or not.
+          * Taking the root of the distances increases query time, but the
+          * function will return true distances instead of their powered
+          * versions.
+          * @param takeRoot set true if root should be taken else false */
+        void setTakeRoot(const bool takeRoot)
+        {
+            takeRoot_ = takeRoot;
+        }
+
+        /** Set the amount of threads that should be used for querying.
+          * OpenMP has to be enabled for this to work.
+          * @param threads amount of threads, 0 for optimal choice */
+        void setThreads(const unsigned int threads)
+        {
+            threads_ = threads;
+        }
+
+        /** Set the maximum distance for querying the tree.
+          * The search will be pruned if the maximum distance is set to any
+          * positive number.
+          * @param maxDist maximum distance, <= 0 for no limit */
+        void setMaxDistance(const Scalar maxDist)
+        {
+            maxDist_ = maxDist;
+        }
+
+        /** Set the data points used for this tree.
+          * This does not build the tree.
+          * @param data NxM matrix, M points of dimension N
+          * @param copy if true data is copied, assumes static data otherwise */
+        void setData(const Matrix &data, const bool copy = false)
+        {
+            if(copy)
+            {
+                dataCopy_ = data;
+                data_ = &dataCopy_;
+            }
+            else
+            {
+                data_ = &data;
+            }
+        }
+
+        void setDistance(const Distance &distance)
+        {
+            distance_ = distance;
+        }
+
+        void build()
+        { }
+
+        template<typename Derived>
+        void query(const Eigen::MatrixBase<Derived> &queryPoints,
+            const size_t knn,
+            Matrixi &indices,
+            Matrix &distances) const
+        {
+            if(data_ == nullptr)
+                throw std::runtime_error("cannot query BruteForce: data not set");
+            if(data_->size() == 0)
+                throw std::runtime_error("cannot query BruteForce: data is empty");
+            if(queryPoints.rows() != dimension())
+                throw std::runtime_error("cannot query BruteForce: data and query descriptors do not have same dimension");
+
+            const Matrix &dataPoints = *data_;
+
+            indices.setConstant(knn, queryPoints.cols(), -1);
+            distances.setConstant(knn, queryPoints.cols(), -1);
+
+            #pragma omp parallel for num_threads(threads_)
+            for(Index i = 0; i < queryPoints.cols(); ++i)
+            {
+                Index *idxPoint = &indices.data()[i * knn];
+                Scalar *distPoint = &distances.data()[i * knn];
+
+                QueryHeap<Scalar> heap(idxPoint, distPoint, knn);
+
+                for(Index j = 0; j < dataPoints.cols(); ++j)
+                {
+                    Scalar dist = distance_(queryPoints.col(i), dataPoints.col(j));
+
+                    // check if point is in range if max distance was set
+                    bool isInRange = maxDist_ <= 0 || dist <= maxDist_;
+                    // check if this node was an improvement if heap is already full
+                    bool isImprovement = !heap.full() ||
+                        dist < heap.front();
+                    if(isInRange && isImprovement)
+                    {
+                        if(heap.full())
+                            heap.pop();
+                        heap.push(j, dist);
+                    }
+                }
+
+                if(sorted_)
+                    heap.sort();
+
+                if(takeRoot_)
+                {
+                    for(size_t j = 0; j < knn; ++j)
+                    {
+                        if(idxPoint[j] < 0)
+                            break;
+                        distPoint[j] = distance_(distPoint[j]);
+                    }
+                }
+            }
+        }
+
+        /** Returns the amount of data points stored in the search index.
+          * @return number of data points */
+        Index size() const
+        {
+            return data_ == nullptr ? 0 : data_->cols();
+        }
+
+        /** Returns the dimension of the data points in the search index.
+          * @return dimension of data points */
+        Index dimension() const
+        {
+            return data_ == nullptr ? 0 : data_->rows();
+        }
+    };
+
+    // template<typename Scalar>
+    // struct MeanMidpointRule
+    // {
+    //     typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic> Matrix;
+    //     typedef knncpp::Matrixi Matrixi;
+
+    //     void operator(const Matrix &data, const Matrixi &indices, Index split)
+    // };
+
+    /** Class for performing k nearest neighbour searches with minkowski distances.
+      * This kdtree only works reliably with the minkowski distance and its
+      * special cases like manhatten or euclidean distance.
+      * @see ManhattenDistance, EuclideanDistance, ChebyshevDistance, MinkowskiDistance*/
+    template<typename _Scalar, int _Dimension, typename _Distance>
+    class KDTreeMinkowski
+    {
+    public:
+        typedef _Scalar Scalar;
+        typedef _Distance Distance;
+        typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic> Matrix;
+        typedef Eigen::Matrix<Scalar, _Dimension, Eigen::Dynamic> DataMatrix;
+        typedef Eigen::Matrix<Scalar, _Dimension, 1> DataVector;
+        typedef knncpp::Matrixi Matrixi;
+    private:
+        typedef Eigen::Matrix<Scalar, 2, 1> Bounds;
+        typedef Eigen::Matrix<Scalar, 2, _Dimension> BoundingBox;
+
+        /** Struct representing a node in the KDTree.
+          * It can be either a inner node or a leaf node. */
+        struct Node
+        {
+            /** Indices of data points in this leaf node. */
+            Index startIdx = 0;
+            Index length = 0;
+
+            /** Left child of this inner node. */
+            Index left = -1;
+            /** Right child of this inner node. */
+            Index right = -1;
+            /** Axis of the axis aligned splitting hyper plane. */
+            Index splitaxis = -1;
+            /** Translation of the axis aligned splitting hyper plane. */
+            Scalar splitpoint = 0;
+            /** Lower end of the splitpoint range */
+            Scalar splitlower = 0;
+            /** Upper end of the splitpoint range */
+            Scalar splitupper = 0;
+
+
+            Node() = default;
+
+            /** Constructor for leaf nodes */
+            Node(const Index startIdx, const Index length)
+                : startIdx(startIdx), length(length)
+            { }
+
+            /** Constructor for inner nodes */
+            Node(const Index splitaxis, const Scalar splitpoint,
+                const Index left, const Index right)
+                : left(left), right(right),
+                splitaxis(splitaxis), splitpoint(splitpoint)
+            { }
+
+            bool isLeaf() const
+            {
+                return !hasLeft() && !hasRight();
+            }
+
+            bool isInner() const
+            {
+                return hasLeft() && hasRight();
+            }
+
+            bool hasLeft() const
+            {
+                return left >= 0;
+            }
+
+            bool hasRight() const
+            {
+                return right >= 0;
+            }
+        };
+
+        DataMatrix dataCopy_ = DataMatrix();
+        const DataMatrix *data_ = nullptr;
+        std::vector<Index> indices_ = std::vector<Index>();
+        std::vector<Node> nodes_ = std::vector<Node>();
+
+        Index bucketSize_ = 16;
+        bool sorted_ = true;
+        bool compact_ = false;
+        bool balanced_ = false;
+        bool takeRoot_ = true;
+        Index threads_ = 0;
+        Scalar maxDist_ = 0;
+
+        Distance distance_ = Distance();
+
+        BoundingBox bbox_ = BoundingBox();
+
+        Index buildLeafNode(const Index startIdx,
+            const Index length,
+            BoundingBox &bbox)
+        {
+            nodes_.push_back(Node(startIdx, length));
+            calculateBoundingBox(startIdx, length, bbox);
+            return static_cast<Index>(nodes_.size() - 1);
+        }
+
+        /** Finds the minimum and maximum values of each dimension (row) in the
+          * data matrix. Only respects the columns specified by the index
+          * vector.
+          * @param startIdx starting index within indices data structure to search for bounding box
+          * @param length length of the block of indices*/
+        void calculateBoundingBox(const Index startIdx,
+            const Index length,
+            BoundingBox &bbox) const
+        {
+            assert(length > 0);
+            assert(startIdx >= 0);
+            assert(static_cast<size_t>(startIdx + length) <= indices_.size());
+            assert(data_->rows() == bbox.cols());
+
+            const DataMatrix &data = *data_;
+
+            // initialize bounds of the bounding box
+            Index first = indices_[startIdx];
+            for(Index i = 0; i < bbox.cols(); ++i)
+            {
+                bbox(0, i) = data(i, first);
+                bbox(1, i) = data(i, first);
+            }
+
+            // search for min / max values in data
+            for(Index i = 1; i < length; ++i)
+            {
+                // retrieve data index
+                Index col = indices_[startIdx + i];
+                assert(col >= 0 && col < data.cols());
+
+                // check min and max for each dimension individually
+                for(Index j = 0; j < data.rows(); ++j)
+                {
+                    bbox(0, j) = std::min(bbox(0, j), data(j, col));
+                    bbox(1, j) = std::max(bbox(1, j), data(j, col));
+                }
+            }
+        }
+
+        /** Calculates the bounds (min / max values) for the given dimension and block of data. */
+        void calculateBounds(const Index startIdx,
+            const Index length,
+            const Index dim,
+            Bounds &bounds) const
+        {
+            assert(length > 0);
+            assert(startIdx >= 0);
+            assert(static_cast<size_t>(startIdx + length) <= indices_.size());
+
+            const DataMatrix &data = *data_;
+
+            bounds(0) = data(dim, indices_[startIdx]);
+            bounds(1) = data(dim, indices_[startIdx]);
+
+            for(Index i = 1; i < length; ++i)
+            {
+                Index col = indices_[startIdx + i];
+                assert(col >= 0 && col < data.cols());
+
+                bounds(0) = std::min(bounds(0), data(dim, col));
+                bounds(1) = std::max(bounds(1), data(dim, col));
+            }
+        }
+
+        void calculateSplittingMidpoint(const Index startIdx,
+            const Index length,
+            const BoundingBox &bbox,
+            Index &splitaxis,
+            Scalar &splitpoint,
+            Index &splitoffset)
+        {
+            const DataMatrix &data = *data_;
+
+            // search for axis with longest distance
+            splitaxis = 0;
+            Scalar splitsize = static_cast<Scalar>(0);
+            for(Index i = 0; i < data.rows(); ++i)
+            {
+                Scalar diff = bbox(1, i) - bbox(0, i);
+                if(diff > splitsize)
+                {
+                    splitaxis = i;
+                    splitsize = diff;
+                }
+            }
+
+            // calculate the bounds in this axis and update our data
+            // accordingly
+            Bounds bounds;
+            calculateBounds(startIdx, length, splitaxis, bounds);
+            splitsize = bounds(1) - bounds(0);
+
+            const Index origSplitaxis = splitaxis;
+            for(Index i = 0; i < data.rows(); ++i)
+            {
+                // skip the dimension of the previously found splitaxis
+                if(i == origSplitaxis)
+                    continue;
+                Scalar diff = bbox(1, i) - bbox(0, i);
+                // check if the split for this dimension would be potentially larger
+                if(diff > splitsize)
+                {
+                    Bounds newBounds;
+                    // update the bounds to their actual current value
+                    calculateBounds(startIdx, length, splitaxis, newBounds);
+                    diff = newBounds(1) - newBounds(0);
+                    if(diff > splitsize)
+                    {
+                        splitaxis = i;
+                        splitsize = diff;
+                        bounds = newBounds;
+                    }
+                }
+            }
+
+            // use the sliding midpoint rule
+            splitpoint = (bounds(0) + bounds(1)) / static_cast<Scalar>(2);
+
+            Index leftIdx = startIdx;
+            Index rightIdx = startIdx + length - 1;
+
+            // first loop checks left < splitpoint and right >= splitpoint
+            while(leftIdx <= rightIdx)
+            {
+                // increment left as long as left has not reached right and
+                // the value of the left element is less than the splitpoint
+                while(leftIdx <= rightIdx && data(splitaxis, indices_[leftIdx]) < splitpoint)
+                    ++leftIdx;
+
+                // decrement right as long as left has not reached right and
+                // the value of the right element is greater than the splitpoint
+                while(leftIdx <= rightIdx && data(splitaxis, indices_[rightIdx]) >= splitpoint)
+                    --rightIdx;
+
+                if(leftIdx <= rightIdx)
+                {
+                    std::swap(indices_[leftIdx], indices_[rightIdx]);
+                    ++leftIdx;
+                    --rightIdx;
+                }
+            }
+
+            // remember this offset from starting index
+            const Index offset1 = leftIdx - startIdx;
+
+            rightIdx = startIdx + length - 1;
+            // second loop checks left <= splitpoint and right > splitpoint
+            while(leftIdx <= rightIdx)
+            {
+                // increment left as long as left has not reached right and
+                // the value of the left element is less than the splitpoint
+                while(leftIdx <= rightIdx && data(splitaxis, indices_[leftIdx]) <= splitpoint)
+                    ++leftIdx;
+
+                // decrement right as long as left has not reached right and
+                // the value of the right element is greater than the splitpoint
+                while(leftIdx <= rightIdx && data(splitaxis, indices_[rightIdx]) > splitpoint)
+                    --rightIdx;
+
+                if(leftIdx <= rightIdx)
+                {
+                    std::swap(indices_[leftIdx], indices_[rightIdx]);
+                    ++leftIdx;
+                    --rightIdx;
+                }
+            }
+
+            // remember this offset from starting index
+            const Index offset2 = leftIdx - startIdx;
+
+            const Index halfLength = length / static_cast<Index>(2);
+
+            // find a separation of points such that is best balanced
+            // offset1 denotes separation where equal points are all on the right
+            // offset2 denots separation where equal points are all on the left
+            if (offset1 > halfLength)
+                splitoffset = offset1;
+            else if (offset2 < halfLength)
+                splitoffset = offset2;
+            // when we get here offset1 < halflength and offset2 > halflength
+            // so simply split the equal elements in the middle
+            else
+                splitoffset = halfLength;
+        }
+
+        Index buildInnerNode(const Index startIdx,
+            const Index length,
+            BoundingBox &bbox)
+        {
+            assert(length > 0);
+            assert(startIdx >= 0);
+            assert(static_cast<size_t>(startIdx  + length) <= indices_.size());
+            assert(data_->rows() == bbox.cols());
+
+            // create node
+            const Index nodeIdx = nodes_.size();
+            nodes_.push_back(Node());
+
+            Index splitaxis;
+            Index splitoffset;
+            Scalar splitpoint;
+            calculateSplittingMidpoint(startIdx, length, bbox, splitaxis, splitpoint, splitoffset);
+
+            nodes_[nodeIdx].splitaxis = splitaxis;
+            nodes_[nodeIdx].splitpoint = splitpoint;
+
+            const Index leftStart = startIdx;
+            const Index leftLength = splitoffset;
+            const Index rightStart = startIdx + splitoffset;
+            const Index rightLength = length - splitoffset;
+
+            BoundingBox bboxLeft = bbox;
+            BoundingBox bboxRight = bbox;
+
+            // do left build
+            bboxLeft(1, splitaxis) = splitpoint;
+            Index left = buildR(leftStart, leftLength, bboxLeft);
+            nodes_[nodeIdx].left =  left;
+
+            // do right build
+            bboxRight(0, splitaxis) = splitpoint;
+            Index right = buildR(rightStart, rightLength, bboxRight);
+            nodes_[nodeIdx].right = right;
+
+            // extract the range of the splitpoint
+            nodes_[nodeIdx].splitlower = bboxLeft(1, splitaxis);
+            nodes_[nodeIdx].splitupper = bboxRight(0, splitaxis);
+
+            // update the bounding box to the values of the new bounding boxes
+            for(Index i = 0; i < bbox.cols(); ++i)
+            {
+                bbox(0, i) = std::min(bboxLeft(0, i), bboxRight(0, i));
+                bbox(1, i) = std::max(bboxLeft(1, i), bboxRight(1, i));
+            }
+
+            return nodeIdx;
+        }
+
+        Index buildR(const Index startIdx,
+            const Index length,
+            BoundingBox &bbox)
+        {
+            // check for base case
+            if(length <= bucketSize_)
+                return buildLeafNode(startIdx, length, bbox);
+            else
+                return buildInnerNode(startIdx, length, bbox);
+        }
+
+        bool isDistanceInRange(const Scalar dist) const
+        {
+            return maxDist_ <= 0 || dist <= maxDist_;
+        }
+
+        bool isDistanceImprovement(const Scalar dist, const QueryHeap<Scalar> &dataHeap) const
+        {
+            return !dataHeap.full() || dist < dataHeap.front();
+        }
+
+        template<typename Derived>
+        void queryLeafNode(const Node &node,
+            const Eigen::MatrixBase<Derived> &queryPoint,
+            QueryHeap<Scalar> &dataHeap) const
+        {
+            assert(node.isLeaf());
+
+            const DataMatrix &data = *data_;
+
+            // go through all points in this leaf node and do brute force search
+            for(Index i = 0; i < node.length; ++i)
+            {
+                const Index idx = node.startIdx + i;
+                assert(idx >= 0 && idx < static_cast<Index>(indices_.size()));
+
+                // retrieve index of the current data point
+                const Index dataIdx = indices_[idx];
+                const Scalar dist = distance_(queryPoint, data.col(dataIdx));
+
+                // check if point is within max distance and if the value would be
+                // an improvement
+                if(isDistanceInRange(dist) && isDistanceImprovement(dist, dataHeap))
+                {
+                    if(dataHeap.full())
+                        dataHeap.pop();
+                    dataHeap.push(dataIdx, dist);
+                }
+            }
+        }
+
+        template<typename Derived>
+        void queryInnerNode(const Node &node,
+            const Eigen::MatrixBase<Derived> &queryPoint,
+            QueryHeap<Scalar> &dataHeap,
+            DataVector &splitdists,
+            const Scalar mindist) const
+        {
+            assert(node.isInner());
+
+            const Index splitaxis = node.splitaxis;
+            const Scalar splitval = queryPoint(splitaxis, 0);
+            Scalar splitdist;
+            Index firstNode;
+            Index secondNode;
+            // check if right or left child should be visited
+            const bool visitLeft = (splitval - node.splitlower + splitval - node.splitupper) < 0;
+            if(visitLeft)
+            {
+                firstNode = node.left;
+                secondNode = node.right;
+                splitdist = distance_(splitval, node.splitupper);
+            }
+            else
+            {
+                firstNode = node.right;
+                secondNode = node.left;
+                splitdist = distance_(splitval, node.splitlower);
+            }
+
+            queryR(nodes_[firstNode], queryPoint, dataHeap, splitdists, mindist);
+
+            const Scalar mindistNew = mindist + splitdist - splitdists(splitaxis);
+
+            // check if node is in range if max distance was set
+            // check if this node was an improvement if heap is already full
+            if(isDistanceInRange(mindistNew) && isDistanceImprovement(mindistNew, dataHeap))
+            {
+                const Scalar splitdistOld = splitdists(splitaxis);
+                splitdists(splitaxis) = splitdist;
+                queryR(nodes_[secondNode], queryPoint, dataHeap, splitdists, mindistNew);
+                splitdists(splitaxis) = splitdistOld;
+            }
+        }
+
+        template<typename Derived>
+        void queryR(const Node &node,
+            const Eigen::MatrixBase<Derived> &queryPoint,
+            QueryHeap<Scalar> &dataHeap,
+            DataVector &splitdists,
+            const Scalar mindist) const
+        {
+            if(node.isLeaf())
+                queryLeafNode(node, queryPoint, dataHeap);
+            else
+                queryInnerNode(node, queryPoint, dataHeap, splitdists, mindist);
+        }
+
+        /** Recursively computes the depth for the given node. */
+        Index depthR(const Node &node) const
+        {
+            if(node.isLeaf())
+                return 1;
+            else
+            {
+                Index left = depthR(nodes_[node.left]);
+                Index right = depthR(nodes_[node.right]);
+                return std::max(left, right) + 1;
+            }
+        }
+
+    public:
+
+        /** Constructs an empty KDTree. */
+        KDTreeMinkowski()
+        { }
+
+        /** Constructs KDTree with the given data. This does not build the
+          * the index of the tree.
+          * @param data NxM matrix, M points of dimension N
+          * @param copy if true copies the data, otherwise assumes static data */
+        KDTreeMinkowski(const DataMatrix &data, const bool copy=false)
+        {
+            setData(data, copy);
+        }
+
+        /** Set the maximum amount of data points per leaf in the tree (aka
+          * bucket size).
+          * @param bucketSize amount of points per leaf. */
+        void setBucketSize(const Index bucketSize)
+        {
+            bucketSize_ = bucketSize;
+        }
+
+        /** Set if the points returned by the queries should be sorted
+          * according to their distance to the query points.
+          * @param sorted sort query results */
+        void setSorted(const bool sorted)
+        {
+            sorted_ = sorted;
+        }
+
+        /** Set if the tree should be built as balanced as possible.
+          * This increases build time, but decreases search time.
+          * @param balanced set true to build a balanced tree */
+        void setBalanced(const bool balanced)
+        {
+            balanced_ = balanced;
+        }
+
+        /** Set if the distances after the query should be rooted or not.
+          * Taking the root of the distances increases query time, but the
+          * function will return true distances instead of their powered
+          * versions.
+          * @param takeRoot set true if root should be taken else false */
+        void setTakeRoot(const bool takeRoot)
+        {
+            takeRoot_ = takeRoot;
+        }
+
+        /** Set if the tree should be built with compact leaf nodes.
+          * This increases build time, but makes leaf nodes denser (more)
+          * points. Thus less visits are necessary.
+          * @param compact set true ti build a tree with compact leafs */
+        void setCompact(const bool compact)
+        {
+            compact_ = compact;
+        }
+
+        /** Set the amount of threads that should be used for building and
+          * querying the tree.
+          * OpenMP has to be enabled for this to work.
+          * @param threads amount of threads, 0 for optimal choice */
+        void setThreads(const unsigned int threads)
+        {
+            threads_ = threads;
+        }
+
+        /** Set the maximum distance for querying the tree.
+          * The search will be pruned if the maximum distance is set to any
+          * positive number.
+          * @param maxDist maximum distance, <= 0 for no limit */
+        void setMaxDistance(const Scalar maxDist)
+        {
+            maxDist_ = maxDist;
+        }
+
+        /** Set the data points used for this tree.
+          * This does not build the tree.
+          * @param data NxM matrix, M points of dimension N
+          * @param copy if true data is copied, assumes static data otherwise */
+        void setData(const DataMatrix &data, const bool copy = false)
+        {
+            clear();
+            if(copy)
+            {
+                dataCopy_ = data;
+                data_ = &dataCopy_;
+            }
+            else
+            {
+                data_ = &data;
+            }
+        }
+
+        void setDistance(const Distance &distance)
+        {
+            distance_ = distance;
+        }
+
+        /** Builds the search index of the tree.
+          * Data has to be set and must be non-empty. */
+        void build()
+        {
+            if(data_ == nullptr)
+                throw std::runtime_error("cannot build KDTree; data not set");
+
+            if(data_->size() == 0)
+                throw std::runtime_error("cannot build KDTree; data is empty");
+
+            clear();
+            nodes_.reserve((data_->cols() / bucketSize_) + 1);
+
+            // initialize indices in simple sequence
+            indices_.resize(data_->cols());
+            for(size_t i = 0; i < indices_.size(); ++i)
+                indices_[i] = i;
+
+            bbox_.resize(2, data_->rows());
+            Index startIdx = 0;
+            Index length = data_->cols();
+
+            calculateBoundingBox(startIdx, length, bbox_);
+
+            buildR(startIdx, length, bbox_);
+        }
+
+        /** Queries the tree for the nearest neighbours of the given query
+          * points.
+          *
+          * The tree has to be built before it can be queried.
+          *
+          * The query points have to have the same dimension as the data points
+          * of the tree.
+          *
+          * The result matrices will be resized appropriatley.
+          * Indices and distances will be set to -1 if less than knn neighbours
+          * were found.
+          *
+          * @param queryPoints NxM matrix, M points of dimension N
+          * @param knn amount of neighbours to be found
+          * @param indices KNNxM matrix, indices of neighbours in the data set
+          * @param distances KNNxM matrix, distance between query points and
+          *        neighbours */
+        template<typename Derived>
+        void query(const Eigen::MatrixBase<Derived> &queryPoints,
+            const size_t knn,
+            Matrixi &indices,
+            Matrix &distances) const
+        {
+            if(nodes_.size() == 0)
+                throw std::runtime_error("cannot query KDTree; not built yet");
+
+            if(queryPoints.rows() != dimension())
+                throw std::runtime_error("cannot query KDTree; data and query points do not have same dimension");
+
+            distances.setConstant(knn, queryPoints.cols(), -1);
+            indices.setConstant(knn, queryPoints.cols(), -1);
+
+            Index *indicesRaw = indices.data();
+            Scalar *distsRaw = distances.data();
+
+            #pragma omp parallel for num_threads(threads_)
+            for(Index i = 0; i < queryPoints.cols(); ++i)
+            {
+
+                Scalar *distPoint = &distsRaw[i * knn];
+                Index *idxPoint = &indicesRaw[i * knn];
+
+                // create heap to find nearest neighbours
+                QueryHeap<Scalar> dataHeap(idxPoint, distPoint, knn);
+
+                Scalar mindist = static_cast<Scalar>(0);
+                DataVector splitdists(queryPoints.rows());
+
+                for(Index j = 0; j < splitdists.rows(); ++j)
+                {
+                    const Scalar value = queryPoints(j, i);
+                    const Scalar lower = bbox_(0, j);
+                    const Scalar upper = bbox_(1, j);
+                    if(value < lower)
+                    {
+                        splitdists(j) = distance_(value, lower);
+                    }
+                    else if(value > upper)
+                    {
+                        splitdists(j) = distance_(value, upper);
+                    }
+                    else
+                    {
+                        splitdists(j) = static_cast<Scalar>(0);
+                    }
+
+                    mindist += splitdists(j);
+                }
+
+                queryR(nodes_[0], queryPoints.col(i), dataHeap, splitdists, mindist);
+
+                if(sorted_)
+                    dataHeap.sort();
+
+                if(takeRoot_)
+                {
+                    for(size_t j = 0; j < knn; ++j)
+                    {
+                        if(distPoint[j] < 0)
+                            break;
+                        distPoint[j] = distance_(distPoint[j]);
+                    }
+                }
+            }
+        }
+
+        /** Clears the tree. */
+        void clear()
+        {
+            nodes_.clear();
+        }
+
+        /** Returns the amount of data points stored in the search index.
+          * @return number of data points */
+        Index size() const
+        {
+            return data_ == nullptr ? 0 : data_->cols();
+        }
+
+        /** Returns the dimension of the data points in the search index.
+          * @return dimension of data points */
+        Index dimension() const
+        {
+            return data_ == nullptr ? 0 : data_->rows();
+        }
+
+        /** Returns the maxximum depth of the tree.
+          * @return maximum depth of the tree */
+        Index depth() const
+        {
+            return nodes_.size() == 0 ? 0 : depthR(nodes_.front());
+        }
+    };
+
+    template<typename _Scalar, typename _Distance = EuclideanDistance<_Scalar>> using KDTreeMinkowski2 = KDTreeMinkowski<_Scalar, 2, _Distance>;
+    template<typename _Scalar, typename _Distance = EuclideanDistance<_Scalar>> using KDTreeMinkowski3 = KDTreeMinkowski<_Scalar, 3, _Distance>;
+    template<typename _Scalar, typename _Distance = EuclideanDistance<_Scalar>> using KDTreeMinkowski4 = KDTreeMinkowski<_Scalar, 4, _Distance>;
+    template<typename _Scalar, typename _Distance = EuclideanDistance<_Scalar>> using KDTreeMinkowski5 = KDTreeMinkowski<_Scalar, 5, _Distance>;
+    template<typename _Scalar, typename _Distance = EuclideanDistance<_Scalar>> using KDTreeMinkowskiX = KDTreeMinkowski<_Scalar, Eigen::Dynamic, _Distance>;
+
+    /** Class for performing KNN search in hamming space by multi-index hashing. */
+    template<typename Scalar>
+    class MultiIndexHashing
+    {
+    public:
+        static_assert(std::is_integral<Scalar>::value, "MultiIndexHashing Scalar has to be integral");
+
+        typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic> Matrix;
+        typedef Eigen::Matrix<Scalar, Eigen::Dynamic, 1> Vector;
+        typedef knncpp::Matrixi Matrixi;
+
+    private:
+        HammingDistance<Scalar> distance_;
+
+        Matrix dataCopy_;
+        const Matrix *data_;
+
+        bool sorted_;
+        Scalar maxDist_;
+        Index substrLen_;
+        Index threads_;
+        std::vector<std::map<Scalar, std::vector<Index>>> buckets_;
+
+        template<typename Derived>
+        Scalar extractCode(const Eigen::MatrixBase<Derived> &data,
+            const Index idx,
+            const Index offset) const
+        {
+            Index leftShift = std::max<Index>(0, static_cast<Index>(sizeof(Scalar)) - offset - substrLen_);
+            Index rightShift = leftShift + offset;
+
+            Scalar code = (data(idx, 0) << (leftShift * 8)) >> (rightShift * 8);
+
+            if(static_cast<Index>(sizeof(Scalar)) - offset < substrLen_ && idx + 1 < data.rows())
+            {
+                Index shift = 2 * static_cast<Index>(sizeof(Scalar)) - substrLen_ - offset;
+                code |= data(idx+1, 0) << (shift * 8);
+            }
+
+            return code;
+        }
+    public:
+        MultiIndexHashing()
+            : distance_(), dataCopy_(), data_(nullptr), sorted_(true),
+            maxDist_(0), substrLen_(1), threads_(1)
+        { }
+
+        /** Constructs an index with the given data.
+          * This does not build the the index.
+          * @param data NxM matrix, M points of dimension N
+          * @param copy if true copies the data, otherwise assumes static data */
+        MultiIndexHashing(const Matrix &data, const bool copy=false)
+            : MultiIndexHashing()
+        {
+            setData(data, copy);
+        }
+
+        /** Set the maximum distance for querying the index.
+          * Note that if no maximum distance is used, this algorithm performs
+          * basically a brute force search.
+          * @param maxDist maximum distance, <= 0 for no limit */
+        void setMaxDistance(const Scalar maxDist)
+        {
+            maxDist_ = maxDist;
+        }
+
+        /** Set if the points returned by the queries should be sorted
+          * according to their distance to the query points.
+          * @param sorted sort query results */
+        void setSorted(const bool sorted)
+        {
+            sorted_ = sorted;
+        }
+
+        /** Set the amount of threads that should be used for building and
+          * querying the tree.
+          * OpenMP has to be enabled for this to work.
+          * @param threads amount of threads, 0 for optimal choice */
+        void setThreads(const unsigned int threads)
+        {
+            threads_ = threads;
+        }
+
+        /** Set the length of substrings (in bytes) used for multi index hashing.
+          * @param len lentth of bucket substrings in bytes*/
+        void setSubstringLength(const Index len)
+        {
+            substrLen_ = len;
+        }
+
+        /** Set the data points used for the KNN search.
+          * @param data NxM matrix, M points of dimension N
+          * @param copy if true data is copied, assumes static data otherwise */
+        void setData(const Matrix &data, const bool copy = false)
+        {
+            clear();
+            if(copy)
+            {
+                dataCopy_ = data;
+                data_ = &dataCopy_;
+            }
+            else
+            {
+                data_ = &data;
+            }
+        }
+
+        void build()
+        {
+            if(data_ == nullptr)
+                throw std::runtime_error("cannot build MultiIndexHashing; data not set");
+            if(data_->size() == 0)
+                throw std::runtime_error("cannot build MultiIndexHashing; data is empty");
+
+            const Matrix &data = *data_;
+            const Index bytesPerVec = data.rows() * static_cast<Index>(sizeof(Scalar));
+            if(bytesPerVec % substrLen_ != 0)
+                throw std::runtime_error("cannot build MultiIndexHashing; cannot divide byte count per vector by substring length without remainings");
+
+            buckets_.clear();
+            buckets_.resize(bytesPerVec / substrLen_);
+
+            for(size_t i = 0; i < buckets_.size(); ++i)
+            {
+                Index start = static_cast<Index>(i) * substrLen_;
+                Index idx = start / static_cast<Index>(sizeof(Scalar));
+                Index offset = start % static_cast<Index>(sizeof(Scalar));
+                std::map<Scalar, std::vector<Index>> &map = buckets_[i];
+
+                for(Index c = 0; c < data.cols(); ++c)
+                {
+                    Scalar code = extractCode(data.col(c), idx, offset);
+                    if(map.find(code) == map.end())
+                        map[code] = std::vector<Index>();
+                    map[code].push_back(c);
+                }
+            }
+        }
+
+        template<typename Derived>
+        void query(const Eigen::MatrixBase<Derived> &queryPoints,
+            const size_t knn,
+            Matrixi &indices,
+            Matrix &distances) const
+        {
+            if(buckets_.size() == 0)
+                throw std::runtime_error("cannot query MultiIndexHashing; not built yet");
+            if(queryPoints.rows() != dimension())
+                throw std::runtime_error("cannot query MultiIndexHashing; data and query points do not have same dimension");
+
+            const Matrix &data = *data_;
+
+            indices.setConstant(knn, queryPoints.cols(), -1);
+            distances.setConstant(knn, queryPoints.cols(), -1);
+
+            Index *indicesRaw = indices.data();
+            Scalar *distsRaw = distances.data();
+
+            Scalar maxDistPart = maxDist_ / buckets_.size();
+
+            #pragma omp parallel for num_threads(threads_)
+            for(Index c = 0; c < queryPoints.cols(); ++c)
+            {
+                std::set<Index> candidates;
+                for(size_t i = 0; i < buckets_.size(); ++i)
+                {
+                    Index start = static_cast<Index>(i) * substrLen_;
+                    Index idx = start / static_cast<Index>(sizeof(Scalar));
+                    Index offset = start % static_cast<Index>(sizeof(Scalar));
+                    const std::map<Scalar, std::vector<Index>> &map = buckets_[i];
+
+                    Scalar code = extractCode(queryPoints.col(c), idx, offset);
+                    for(const auto &x: map)
+                    {
+                        Scalar dist = distance_(x.first, code);
+                        if(maxDistPart <= 0 || dist <= maxDistPart)
+                        {
+                            for(size_t j = 0; j < x.second.size(); ++j)
+                                candidates.insert(x.second[j]);
+                        }
+                    }
+                }
+
+                Scalar *distPoint = &distsRaw[c * knn];
+                Index *idxPoint = &indicesRaw[c * knn];
+                // create heap to find nearest neighbours
+                QueryHeap<Scalar> dataHeap(idxPoint, distPoint, knn);
+
+                for(Index idx: candidates)
+                {
+                    Scalar dist = distance_(data.col(idx), queryPoints.col(c));
+
+                    bool isInRange = maxDist_ <= 0 || dist <= maxDist_;
+                    bool isImprovement = !dataHeap.full() ||
+                        dist < dataHeap.front();
+                    if(isInRange && isImprovement)
+                    {
+                        if(dataHeap.full())
+                            dataHeap.pop();
+                        dataHeap.push(idx, dist);
+                    }
+                }
+
+                if(sorted_)
+                    dataHeap.sort();
+            }
+        }
+
+        /** Returns the amount of data points stored in the search index.
+          * @return number of data points */
+        Index size() const
+        {
+            return data_ == nullptr ? 0 : data_->cols();
+        }
+
+        /** Returns the dimension of the data points in the search index.
+          * @return dimension of data points */
+        Index dimension() const
+        {
+            return data_ == nullptr ? 0 : data_->rows();
+        }
+
+        void clear()
+        {
+            data_ = nullptr;
+            dataCopy_.resize(0, 0);
+            buckets_.clear();
+        }
+
+    };
+
+    #ifdef KNNCPP_FLANN
+
+    /** Wrapper class of FLANN kdtrees for the use with Eigen3. */
+    template<typename Scalar,
+        typename Distance=flann::L2_Simple<Scalar>>
+    class KDTreeFlann
+    {
+    public:
+        typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic> Matrix;
+        typedef Eigen::Matrix<Scalar, Eigen::Dynamic, 1> Vector;
+        typedef Eigen::Matrix<int, Eigen::Dynamic, Eigen::Dynamic> Matrixi;
+
+    private:
+        typedef flann::Index<Distance> FlannIndex;
+
+        Matrix dataCopy_;
+        Matrix *dataPoints_;
+
+        FlannIndex *index_;
+        flann::SearchParams searchParams_;
+        flann::IndexParams indexParams_;
+        Scalar maxDist_;
+
+    public:
+        KDTreeFlann()
+            : dataCopy_(), dataPoints_(nullptr), index_(nullptr),
+            searchParams_(32, 0, false),
+            indexParams_(flann::KDTreeSingleIndexParams(15)),
+            maxDist_(0)
+        {
+        }
+
+        KDTreeFlann(Matrix &data, const bool copy = false)
+            : KDTreeFlann()
+        {
+            setData(data, copy);
+        }
+
+        ~KDTreeFlann()
+        {
+            clear();
+        }
+
+        void setIndexParams(const flann::IndexParams &params)
+        {
+            indexParams_ = params;
+        }
+
+        void setChecks(const int checks)
+        {
+            searchParams_.checks = checks;
+        }
+
+        void setSorted(const bool sorted)
+        {
+            searchParams_.sorted = sorted;
+        }
+
+        void setThreads(const int threads)
+        {
+            searchParams_.cores = threads;
+        }
+
+        void setEpsilon(const float eps)
+        {
+            searchParams_.eps = eps;
+        }
+
+        void setMaxDistance(const Scalar dist)
+        {
+            maxDist_ = dist;
+        }
+
+        void setData(Matrix &data, const bool copy = false)
+        {
+            if(copy)
+            {
+                dataCopy_ = data;
+                dataPoints_ = &dataCopy_;
+            }
+            else
+            {
+                dataPoints_ = &data;
+            }
+
+            clear();
+        }
+
+        void build()
+        {
+            if(dataPoints_ == nullptr)
+                throw std::runtime_error("cannot build KDTree; data not set");
+            if(dataPoints_->size() == 0)
+                throw std::runtime_error("cannot build KDTree; data is empty");
+
+            if(index_ != nullptr)
+                delete index_;
+
+            flann::Matrix<Scalar> dataPts(
+                dataPoints_->data(),
+                dataPoints_->cols(),
+                dataPoints_->rows());
+
+            index_ = new FlannIndex(dataPts, indexParams_);
+            index_->buildIndex();
+        }
+
+        void query(Matrix &queryPoints,
+            const size_t knn,
+            Matrixi &indices,
+            Matrix &distances) const
+        {
+            if(index_ == nullptr)
+                throw std::runtime_error("cannot query KDTree; not built yet");
+            if(dataPoints_->rows() != queryPoints.rows())
+                throw std::runtime_error("cannot query KDTree; KDTree has different dimension than query data");
+
+            // resize result matrices
+            distances.resize(knn, queryPoints.cols());
+            indices.resize(knn, queryPoints.cols());
+
+            // wrap matrices into flann matrices
+            flann::Matrix<Scalar> queryPts(
+                queryPoints.data(),
+                queryPoints.cols(),
+                queryPoints.rows());
+            flann::Matrix<int> indicesF(
+                indices.data(),
+                indices.cols(),
+                indices.rows());
+            flann::Matrix<Scalar> distancesF(
+                distances.data(),
+                distances.cols(),
+                distances.rows());
+
+            // if maximum distance was set then use radius search
+            if(maxDist_ > 0)
+                index_->radiusSearch(queryPts, indicesF, distancesF, maxDist_, searchParams_);
+            else
+                index_->knnSearch(queryPts, indicesF, distancesF, knn, searchParams_);
+
+            // make result matrices compatible to API
+            #pragma omp parallel for num_threads(searchParams_.cores)
+            for(Index i = 0; i < indices.cols(); ++i)
+            {
+                bool found = false;
+                for(Index j = 0; j < indices.rows(); ++j)
+                {
+                    if(indices(j, i) == -1)
+                        found = true;
+
+                    if(found)
+                    {
+                        indices(j, i) = -1;
+                        distances(j, i) = -1;
+                    }
+                }
+            }
+        }
+
+        Index size() const
+        {
+            return dataPoints_ == nullptr ? 0 : dataPoints_->cols();
+        }
+
+        Index dimension() const
+        {
+            return dataPoints_ == nullptr ? 0 : dataPoints_->rows();
+        }
+
+        void clear()
+        {
+            if(index_ != nullptr)
+            {
+                delete index_;
+                index_ = nullptr;
+            }
+        }
+
+        FlannIndex &flannIndex()
+        {
+            return index_;
+        }
+    };
+
+    typedef KDTreeFlann<double> KDTreeFlannd;
+    typedef KDTreeFlann<float> KDTreeFlannf;
+
+    #endif
+}
+
+#endif
\ No newline at end of file
diff --git a/src/nng.cpp b/src/nng.cpp
new file mode 100644 (file)
index 0000000..a326414
--- /dev/null
@@ -0,0 +1,65 @@
+#include <iostream>
+#include <Rcpp.h>
+#include <RcppEigen.h>
+#include "knncpp.h"
+
+using namespace Rcpp;
+
+// [[Rcpp::depends(RcppEigen)]]
+// [[Rcpp::export]]
+List findNeighbors(NumericMatrix data, int k, bool mutual) {
+  int n = data.nrow(),
+      d = data.ncol();
+
+  Eigen::MatrixXd dataPoints(d, n);
+  for (int row = 0; row < n; ++row) {
+    for (int col = 0; col < d; ++col) {
+      dataPoints(col,row) = data(row,col); //dataPoints: by columns
+    }
+  }
+
+  knncpp::KDTreeMinkowskiX<double, knncpp::EuclideanDistance<double>> kdtree(dataPoints);
+  kdtree.setBucketSize(16);
+  kdtree.setSorted(false);
+  kdtree.setTakeRoot(false);
+  kdtree.setMaxDistance(0);
+  kdtree.setThreads(0);
+  kdtree.build();
+
+  knncpp::Matrixi indices;
+  Eigen::MatrixXd distances;
+  // k+1 because i is always a neighbor of i (to discard)
+  kdtree.query(dataPoints, k+1, indices, distances);
+
+  NumericVector res_edges(0);
+  NumericVector res_dists(0);
+  for (int i = 0; i <= k; ++i) {
+    for (int j = 0; j < n; ++j) {
+      if (indices(i,j) == j)
+        continue;
+      bool addRow = false;
+      if (!mutual)
+        addRow = true;
+      else if (mutual && j < indices(i,j)) {
+        int l = 0;
+        for (; l <= k; ++l) {
+          if (indices(l,indices(i,j)) == j)
+            break;
+        }
+        if (l <= k)
+          addRow = true;
+      }
+      if (addRow) {
+        // R indices from 1 to n:
+        res_edges.push_back(j+1);
+        res_edges.push_back(indices(i,j)+1);
+        res_dists.push_back(distances(i,j));
+      }
+    }
+  }
+
+  res_edges.attr("dim") = Dimension(2, res_edges.length() / 2);
+  List L = List::create(Named("edges") = res_edges,
+                        Named("euc_dists") = res_dists);
+  return L;
+}