mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-06-12 20:42:53 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

flann/kmeans: Fix non-determinism of KMeans index

Browse Source 
When running with >1 OpenCV thread, KMeans index generation was
non-deterministic because of a RWW race. Issue is resolved by removing
the offending logic from the parallel section.
This commit is contained in:
Matthew K. Gumbel 2019-03-19 15:59:35 -07:00 committed by Matt Gumbel
parent 6c862fae13
commit 74e3403620
1 changed files with 27 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
modules/flann/include/opencv2/flann/kmeans_index.h
View File

@ -276,17 +276,15 @@ public:
public: public:
KMeansDistanceComputer(Distance _distance, const Matrix<ElementType>& _dataset, KMeansDistanceComputer(Distance _distance, const Matrix<ElementType>& _dataset,
const int _branching, const int* _indices, const Matrix<double>& _dcenters, const size_t _veclen, const int _branching, const int* _indices, const Matrix<double>& _dcenters, const size_t _veclen,
int* _count, int* _belongs_to, std::vector<DistanceType>& _radiuses, bool& _converged) std::vector<int> &_new_centroids, std::vector<DistanceType> &_sq_dists)
: distance(_distance) : distance(_distance)
, dataset(_dataset) , dataset(_dataset)
, branching(_branching) , branching(_branching)
, indices(_indices) , indices(_indices)
, dcenters(_dcenters) , dcenters(_dcenters)
, veclen(_veclen) , veclen(_veclen)
, count(_count) , new_centroids(_new_centroids)
, belongs_to(_belongs_to) , sq_dists(_sq_dists)
, radiuses(_radiuses)
, converged(_converged)
{ {
} }
@ -297,8 +295,8 @@ public:
for( int i = begin; i<end; ++i) for( int i = begin; i<end; ++i)
{ {
DistanceType sq_dist = distance(dataset[indices[i]], dcenters[0], veclen); DistanceType sq_dist(distance(dataset[indices[i]], dcenters[0], veclen));
int new_centroid = 0; int new_centroid(0);
for (int j=1; j<branching; ++j) { for (int j=1; j<branching; ++j) {
DistanceType new_sq_dist = distance(dataset[indices[i]], dcenters[j], veclen); DistanceType new_sq_dist = distance(dataset[indices[i]], dcenters[j], veclen);
if (sq_dist>new_sq_dist) { if (sq_dist>new_sq_dist) {
@ -306,15 +304,8 @@ public:
sq_dist = new_sq_dist; sq_dist = new_sq_dist;
} }
} }
if (sq_dist > radiuses[new_centroid]) { sq_dists[i] = sq_dist;
radiuses[new_centroid] = sq_dist; new_centroids[i] = new_centroid;
}
if (new_centroid != belongs_to[i]) {
CV_XADD(&count[belongs_to[i]], -1);
CV_XADD(&count[new_centroid], 1);
belongs_to[i] = new_centroid;
converged = false;
}
} }
} }
@ -325,10 +316,8 @@ public:
const int* indices; const int* indices;
const Matrix<double>& dcenters; const Matrix<double>& dcenters;
const size_t veclen; const size_t veclen;
int* count; std::vector<int> &new_centroids;
int* belongs_to; std::vector<DistanceType> &sq_dists;
std::vector<DistanceType>& radiuses;
bool& converged;
KMeansDistanceComputer& operator=( const KMeansDistanceComputer & ) { return *this; } KMeansDistanceComputer& operator=( const KMeansDistanceComputer & ) { return *this; }
}; };
@ -796,10 +785,27 @@ private:
} }
} }
std::vector<int> new_centroids(indices_length);
std::vector<DistanceType> sq_dists(indices_length);
// reassign points to clusters // reassign points to clusters
KMeansDistanceComputer invoker(distance_, dataset_, branching, indices, dcenters, veclen_, count, belongs_to, radiuses, converged); KMeansDistanceComputer invoker(distance_, dataset_, branching, indices, dcenters, veclen_, new_centroids, sq_dists);
parallel_for_(cv::Range(0, (int)indices_length), invoker); parallel_for_(cv::Range(0, (int)indices_length), invoker);
for (int i=0; i < (int)indices_length; ++i) {
DistanceType sq_dist(sq_dists[i]);
int new_centroid(new_centroids[i]);
if (sq_dist > radiuses[new_centroid]) {
radiuses[new_centroid] = sq_dist;
}
if (new_centroid != belongs_to[i]) {
count[belongs_to[i]]--;
count[new_centroid]++;
belongs_to[i] = new_centroid;
converged = false;
}
}
for (int i=0; i<branching; ++i) { for (int i=0; i<branching; ++i) {
// if one cluster converges to an empty cluster, // if one cluster converges to an empty cluster,
// move an element into that cluster // move an element into that cluster