umappp 0.1.2
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +7 -0
- data/LICENSE.txt +25 -0
- data/README.md +110 -0
- data/ext/umappp/extconf.rb +25 -0
- data/ext/umappp/numo.hpp +867 -0
- data/ext/umappp/umappp.cpp +225 -0
- data/lib/umappp/version.rb +5 -0
- data/lib/umappp.rb +41 -0
- data/vendor/Eigen/Cholesky +45 -0
- data/vendor/Eigen/CholmodSupport +48 -0
- data/vendor/Eigen/Core +384 -0
- data/vendor/Eigen/Dense +7 -0
- data/vendor/Eigen/Eigen +2 -0
- data/vendor/Eigen/Eigenvalues +60 -0
- data/vendor/Eigen/Geometry +59 -0
- data/vendor/Eigen/Householder +29 -0
- data/vendor/Eigen/IterativeLinearSolvers +48 -0
- data/vendor/Eigen/Jacobi +32 -0
- data/vendor/Eigen/KLUSupport +41 -0
- data/vendor/Eigen/LU +47 -0
- data/vendor/Eigen/MetisSupport +35 -0
- data/vendor/Eigen/OrderingMethods +70 -0
- data/vendor/Eigen/PaStiXSupport +49 -0
- data/vendor/Eigen/PardisoSupport +35 -0
- data/vendor/Eigen/QR +50 -0
- data/vendor/Eigen/QtAlignedMalloc +39 -0
- data/vendor/Eigen/SPQRSupport +34 -0
- data/vendor/Eigen/SVD +50 -0
- data/vendor/Eigen/Sparse +34 -0
- data/vendor/Eigen/SparseCholesky +37 -0
- data/vendor/Eigen/SparseCore +69 -0
- data/vendor/Eigen/SparseLU +50 -0
- data/vendor/Eigen/SparseQR +36 -0
- data/vendor/Eigen/StdDeque +27 -0
- data/vendor/Eigen/StdList +26 -0
- data/vendor/Eigen/StdVector +27 -0
- data/vendor/Eigen/SuperLUSupport +64 -0
- data/vendor/Eigen/UmfPackSupport +40 -0
- data/vendor/Eigen/src/Cholesky/LDLT.h +688 -0
- data/vendor/Eigen/src/Cholesky/LLT.h +558 -0
- data/vendor/Eigen/src/Cholesky/LLT_LAPACKE.h +99 -0
- data/vendor/Eigen/src/CholmodSupport/CholmodSupport.h +682 -0
- data/vendor/Eigen/src/Core/ArithmeticSequence.h +413 -0
- data/vendor/Eigen/src/Core/Array.h +417 -0
- data/vendor/Eigen/src/Core/ArrayBase.h +226 -0
- data/vendor/Eigen/src/Core/ArrayWrapper.h +209 -0
- data/vendor/Eigen/src/Core/Assign.h +90 -0
- data/vendor/Eigen/src/Core/AssignEvaluator.h +1010 -0
- data/vendor/Eigen/src/Core/Assign_MKL.h +178 -0
- data/vendor/Eigen/src/Core/BandMatrix.h +353 -0
- data/vendor/Eigen/src/Core/Block.h +448 -0
- data/vendor/Eigen/src/Core/BooleanRedux.h +162 -0
- data/vendor/Eigen/src/Core/CommaInitializer.h +164 -0
- data/vendor/Eigen/src/Core/ConditionEstimator.h +175 -0
- data/vendor/Eigen/src/Core/CoreEvaluators.h +1741 -0
- data/vendor/Eigen/src/Core/CoreIterators.h +132 -0
- data/vendor/Eigen/src/Core/CwiseBinaryOp.h +183 -0
- data/vendor/Eigen/src/Core/CwiseNullaryOp.h +1001 -0
- data/vendor/Eigen/src/Core/CwiseTernaryOp.h +197 -0
- data/vendor/Eigen/src/Core/CwiseUnaryOp.h +103 -0
- data/vendor/Eigen/src/Core/CwiseUnaryView.h +132 -0
- data/vendor/Eigen/src/Core/DenseBase.h +701 -0
- data/vendor/Eigen/src/Core/DenseCoeffsBase.h +685 -0
- data/vendor/Eigen/src/Core/DenseStorage.h +652 -0
- data/vendor/Eigen/src/Core/Diagonal.h +258 -0
- data/vendor/Eigen/src/Core/DiagonalMatrix.h +391 -0
- data/vendor/Eigen/src/Core/DiagonalProduct.h +28 -0
- data/vendor/Eigen/src/Core/Dot.h +318 -0
- data/vendor/Eigen/src/Core/EigenBase.h +160 -0
- data/vendor/Eigen/src/Core/ForceAlignedAccess.h +150 -0
- data/vendor/Eigen/src/Core/Fuzzy.h +155 -0
- data/vendor/Eigen/src/Core/GeneralProduct.h +465 -0
- data/vendor/Eigen/src/Core/GenericPacketMath.h +1040 -0
- data/vendor/Eigen/src/Core/GlobalFunctions.h +194 -0
- data/vendor/Eigen/src/Core/IO.h +258 -0
- data/vendor/Eigen/src/Core/IndexedView.h +237 -0
- data/vendor/Eigen/src/Core/Inverse.h +117 -0
- data/vendor/Eigen/src/Core/Map.h +171 -0
- data/vendor/Eigen/src/Core/MapBase.h +310 -0
- data/vendor/Eigen/src/Core/MathFunctions.h +2057 -0
- data/vendor/Eigen/src/Core/MathFunctionsImpl.h +200 -0
- data/vendor/Eigen/src/Core/Matrix.h +565 -0
- data/vendor/Eigen/src/Core/MatrixBase.h +547 -0
- data/vendor/Eigen/src/Core/NestByValue.h +85 -0
- data/vendor/Eigen/src/Core/NoAlias.h +109 -0
- data/vendor/Eigen/src/Core/NumTraits.h +335 -0
- data/vendor/Eigen/src/Core/PartialReduxEvaluator.h +232 -0
- data/vendor/Eigen/src/Core/PermutationMatrix.h +605 -0
- data/vendor/Eigen/src/Core/PlainObjectBase.h +1128 -0
- data/vendor/Eigen/src/Core/Product.h +191 -0
- data/vendor/Eigen/src/Core/ProductEvaluators.h +1179 -0
- data/vendor/Eigen/src/Core/Random.h +218 -0
- data/vendor/Eigen/src/Core/Redux.h +515 -0
- data/vendor/Eigen/src/Core/Ref.h +381 -0
- data/vendor/Eigen/src/Core/Replicate.h +142 -0
- data/vendor/Eigen/src/Core/Reshaped.h +454 -0
- data/vendor/Eigen/src/Core/ReturnByValue.h +119 -0
- data/vendor/Eigen/src/Core/Reverse.h +217 -0
- data/vendor/Eigen/src/Core/Select.h +164 -0
- data/vendor/Eigen/src/Core/SelfAdjointView.h +365 -0
- data/vendor/Eigen/src/Core/SelfCwiseBinaryOp.h +47 -0
- data/vendor/Eigen/src/Core/Solve.h +188 -0
- data/vendor/Eigen/src/Core/SolveTriangular.h +235 -0
- data/vendor/Eigen/src/Core/SolverBase.h +168 -0
- data/vendor/Eigen/src/Core/StableNorm.h +251 -0
- data/vendor/Eigen/src/Core/StlIterators.h +463 -0
- data/vendor/Eigen/src/Core/Stride.h +116 -0
- data/vendor/Eigen/src/Core/Swap.h +68 -0
- data/vendor/Eigen/src/Core/Transpose.h +464 -0
- data/vendor/Eigen/src/Core/Transpositions.h +386 -0
- data/vendor/Eigen/src/Core/TriangularMatrix.h +1001 -0
- data/vendor/Eigen/src/Core/VectorBlock.h +96 -0
- data/vendor/Eigen/src/Core/VectorwiseOp.h +784 -0
- data/vendor/Eigen/src/Core/Visitor.h +381 -0
- data/vendor/Eigen/src/Core/arch/AVX/Complex.h +372 -0
- data/vendor/Eigen/src/Core/arch/AVX/MathFunctions.h +228 -0
- data/vendor/Eigen/src/Core/arch/AVX/PacketMath.h +1574 -0
- data/vendor/Eigen/src/Core/arch/AVX/TypeCasting.h +115 -0
- data/vendor/Eigen/src/Core/arch/AVX512/Complex.h +422 -0
- data/vendor/Eigen/src/Core/arch/AVX512/MathFunctions.h +362 -0
- data/vendor/Eigen/src/Core/arch/AVX512/PacketMath.h +2303 -0
- data/vendor/Eigen/src/Core/arch/AVX512/TypeCasting.h +89 -0
- data/vendor/Eigen/src/Core/arch/AltiVec/Complex.h +417 -0
- data/vendor/Eigen/src/Core/arch/AltiVec/MathFunctions.h +90 -0
- data/vendor/Eigen/src/Core/arch/AltiVec/MatrixProduct.h +2937 -0
- data/vendor/Eigen/src/Core/arch/AltiVec/MatrixProductCommon.h +221 -0
- data/vendor/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h +629 -0
- data/vendor/Eigen/src/Core/arch/AltiVec/PacketMath.h +2711 -0
- data/vendor/Eigen/src/Core/arch/CUDA/Complex.h +258 -0
- data/vendor/Eigen/src/Core/arch/Default/BFloat16.h +700 -0
- data/vendor/Eigen/src/Core/arch/Default/ConjHelper.h +117 -0
- data/vendor/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +1649 -0
- data/vendor/Eigen/src/Core/arch/Default/GenericPacketMathFunctionsFwd.h +110 -0
- data/vendor/Eigen/src/Core/arch/Default/Half.h +942 -0
- data/vendor/Eigen/src/Core/arch/Default/Settings.h +49 -0
- data/vendor/Eigen/src/Core/arch/Default/TypeCasting.h +120 -0
- data/vendor/Eigen/src/Core/arch/GPU/MathFunctions.h +103 -0
- data/vendor/Eigen/src/Core/arch/GPU/PacketMath.h +1685 -0
- data/vendor/Eigen/src/Core/arch/GPU/TypeCasting.h +80 -0
- data/vendor/Eigen/src/Core/arch/HIP/hcc/math_constants.h +23 -0
- data/vendor/Eigen/src/Core/arch/MSA/Complex.h +648 -0
- data/vendor/Eigen/src/Core/arch/MSA/MathFunctions.h +387 -0
- data/vendor/Eigen/src/Core/arch/MSA/PacketMath.h +1233 -0
- data/vendor/Eigen/src/Core/arch/NEON/Complex.h +584 -0
- data/vendor/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h +183 -0
- data/vendor/Eigen/src/Core/arch/NEON/MathFunctions.h +75 -0
- data/vendor/Eigen/src/Core/arch/NEON/PacketMath.h +4587 -0
- data/vendor/Eigen/src/Core/arch/NEON/TypeCasting.h +1419 -0
- data/vendor/Eigen/src/Core/arch/SSE/Complex.h +351 -0
- data/vendor/Eigen/src/Core/arch/SSE/MathFunctions.h +199 -0
- data/vendor/Eigen/src/Core/arch/SSE/PacketMath.h +1505 -0
- data/vendor/Eigen/src/Core/arch/SSE/TypeCasting.h +142 -0
- data/vendor/Eigen/src/Core/arch/SVE/MathFunctions.h +44 -0
- data/vendor/Eigen/src/Core/arch/SVE/PacketMath.h +752 -0
- data/vendor/Eigen/src/Core/arch/SVE/TypeCasting.h +49 -0
- data/vendor/Eigen/src/Core/arch/SYCL/InteropHeaders.h +232 -0
- data/vendor/Eigen/src/Core/arch/SYCL/MathFunctions.h +301 -0
- data/vendor/Eigen/src/Core/arch/SYCL/PacketMath.h +670 -0
- data/vendor/Eigen/src/Core/arch/SYCL/SyclMemoryModel.h +694 -0
- data/vendor/Eigen/src/Core/arch/SYCL/TypeCasting.h +85 -0
- data/vendor/Eigen/src/Core/arch/ZVector/Complex.h +426 -0
- data/vendor/Eigen/src/Core/arch/ZVector/MathFunctions.h +233 -0
- data/vendor/Eigen/src/Core/arch/ZVector/PacketMath.h +1060 -0
- data/vendor/Eigen/src/Core/functors/AssignmentFunctors.h +177 -0
- data/vendor/Eigen/src/Core/functors/BinaryFunctors.h +541 -0
- data/vendor/Eigen/src/Core/functors/NullaryFunctors.h +189 -0
- data/vendor/Eigen/src/Core/functors/StlFunctors.h +166 -0
- data/vendor/Eigen/src/Core/functors/TernaryFunctors.h +25 -0
- data/vendor/Eigen/src/Core/functors/UnaryFunctors.h +1131 -0
- data/vendor/Eigen/src/Core/products/GeneralBlockPanelKernel.h +2645 -0
- data/vendor/Eigen/src/Core/products/GeneralMatrixMatrix.h +517 -0
- data/vendor/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +317 -0
- data/vendor/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h +145 -0
- data/vendor/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h +124 -0
- data/vendor/Eigen/src/Core/products/GeneralMatrixVector.h +518 -0
- data/vendor/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h +136 -0
- data/vendor/Eigen/src/Core/products/Parallelizer.h +180 -0
- data/vendor/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +544 -0
- data/vendor/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h +295 -0
- data/vendor/Eigen/src/Core/products/SelfadjointMatrixVector.h +262 -0
- data/vendor/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h +118 -0
- data/vendor/Eigen/src/Core/products/SelfadjointProduct.h +133 -0
- data/vendor/Eigen/src/Core/products/SelfadjointRank2Update.h +94 -0
- data/vendor/Eigen/src/Core/products/TriangularMatrixMatrix.h +472 -0
- data/vendor/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h +317 -0
- data/vendor/Eigen/src/Core/products/TriangularMatrixVector.h +350 -0
- data/vendor/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h +255 -0
- data/vendor/Eigen/src/Core/products/TriangularSolverMatrix.h +337 -0
- data/vendor/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h +167 -0
- data/vendor/Eigen/src/Core/products/TriangularSolverVector.h +148 -0
- data/vendor/Eigen/src/Core/util/BlasUtil.h +583 -0
- data/vendor/Eigen/src/Core/util/ConfigureVectorization.h +512 -0
- data/vendor/Eigen/src/Core/util/Constants.h +563 -0
- data/vendor/Eigen/src/Core/util/DisableStupidWarnings.h +106 -0
- data/vendor/Eigen/src/Core/util/ForwardDeclarations.h +322 -0
- data/vendor/Eigen/src/Core/util/IndexedViewHelper.h +186 -0
- data/vendor/Eigen/src/Core/util/IntegralConstant.h +272 -0
- data/vendor/Eigen/src/Core/util/MKL_support.h +137 -0
- data/vendor/Eigen/src/Core/util/Macros.h +1464 -0
- data/vendor/Eigen/src/Core/util/Memory.h +1163 -0
- data/vendor/Eigen/src/Core/util/Meta.h +812 -0
- data/vendor/Eigen/src/Core/util/NonMPL2.h +3 -0
- data/vendor/Eigen/src/Core/util/ReenableStupidWarnings.h +31 -0
- data/vendor/Eigen/src/Core/util/ReshapedHelper.h +51 -0
- data/vendor/Eigen/src/Core/util/StaticAssert.h +221 -0
- data/vendor/Eigen/src/Core/util/SymbolicIndex.h +293 -0
- data/vendor/Eigen/src/Core/util/XprHelper.h +856 -0
- data/vendor/Eigen/src/Eigenvalues/ComplexEigenSolver.h +346 -0
- data/vendor/Eigen/src/Eigenvalues/ComplexSchur.h +462 -0
- data/vendor/Eigen/src/Eigenvalues/ComplexSchur_LAPACKE.h +91 -0
- data/vendor/Eigen/src/Eigenvalues/EigenSolver.h +622 -0
- data/vendor/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h +418 -0
- data/vendor/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h +226 -0
- data/vendor/Eigen/src/Eigenvalues/HessenbergDecomposition.h +374 -0
- data/vendor/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h +158 -0
- data/vendor/Eigen/src/Eigenvalues/RealQZ.h +657 -0
- data/vendor/Eigen/src/Eigenvalues/RealSchur.h +558 -0
- data/vendor/Eigen/src/Eigenvalues/RealSchur_LAPACKE.h +77 -0
- data/vendor/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +904 -0
- data/vendor/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h +87 -0
- data/vendor/Eigen/src/Eigenvalues/Tridiagonalization.h +561 -0
- data/vendor/Eigen/src/Geometry/AlignedBox.h +486 -0
- data/vendor/Eigen/src/Geometry/AngleAxis.h +247 -0
- data/vendor/Eigen/src/Geometry/EulerAngles.h +114 -0
- data/vendor/Eigen/src/Geometry/Homogeneous.h +501 -0
- data/vendor/Eigen/src/Geometry/Hyperplane.h +282 -0
- data/vendor/Eigen/src/Geometry/OrthoMethods.h +235 -0
- data/vendor/Eigen/src/Geometry/ParametrizedLine.h +232 -0
- data/vendor/Eigen/src/Geometry/Quaternion.h +870 -0
- data/vendor/Eigen/src/Geometry/Rotation2D.h +199 -0
- data/vendor/Eigen/src/Geometry/RotationBase.h +206 -0
- data/vendor/Eigen/src/Geometry/Scaling.h +188 -0
- data/vendor/Eigen/src/Geometry/Transform.h +1563 -0
- data/vendor/Eigen/src/Geometry/Translation.h +202 -0
- data/vendor/Eigen/src/Geometry/Umeyama.h +166 -0
- data/vendor/Eigen/src/Geometry/arch/Geometry_SIMD.h +168 -0
- data/vendor/Eigen/src/Householder/BlockHouseholder.h +110 -0
- data/vendor/Eigen/src/Householder/Householder.h +176 -0
- data/vendor/Eigen/src/Householder/HouseholderSequence.h +545 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h +226 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +212 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +229 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/IncompleteCholesky.h +394 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h +453 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +444 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/LeastSquareConjugateGradient.h +198 -0
- data/vendor/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h +117 -0
- data/vendor/Eigen/src/Jacobi/Jacobi.h +483 -0
- data/vendor/Eigen/src/KLUSupport/KLUSupport.h +358 -0
- data/vendor/Eigen/src/LU/Determinant.h +117 -0
- data/vendor/Eigen/src/LU/FullPivLU.h +877 -0
- data/vendor/Eigen/src/LU/InverseImpl.h +432 -0
- data/vendor/Eigen/src/LU/PartialPivLU.h +624 -0
- data/vendor/Eigen/src/LU/PartialPivLU_LAPACKE.h +83 -0
- data/vendor/Eigen/src/LU/arch/InverseSize4.h +351 -0
- data/vendor/Eigen/src/MetisSupport/MetisSupport.h +137 -0
- data/vendor/Eigen/src/OrderingMethods/Amd.h +435 -0
- data/vendor/Eigen/src/OrderingMethods/Eigen_Colamd.h +1863 -0
- data/vendor/Eigen/src/OrderingMethods/Ordering.h +153 -0
- data/vendor/Eigen/src/PaStiXSupport/PaStiXSupport.h +678 -0
- data/vendor/Eigen/src/PardisoSupport/PardisoSupport.h +545 -0
- data/vendor/Eigen/src/QR/ColPivHouseholderQR.h +674 -0
- data/vendor/Eigen/src/QR/ColPivHouseholderQR_LAPACKE.h +97 -0
- data/vendor/Eigen/src/QR/CompleteOrthogonalDecomposition.h +635 -0
- data/vendor/Eigen/src/QR/FullPivHouseholderQR.h +713 -0
- data/vendor/Eigen/src/QR/HouseholderQR.h +434 -0
- data/vendor/Eigen/src/QR/HouseholderQR_LAPACKE.h +68 -0
- data/vendor/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +335 -0
- data/vendor/Eigen/src/SVD/BDCSVD.h +1366 -0
- data/vendor/Eigen/src/SVD/JacobiSVD.h +812 -0
- data/vendor/Eigen/src/SVD/JacobiSVD_LAPACKE.h +91 -0
- data/vendor/Eigen/src/SVD/SVDBase.h +376 -0
- data/vendor/Eigen/src/SVD/UpperBidiagonalization.h +414 -0
- data/vendor/Eigen/src/SparseCholesky/SimplicialCholesky.h +697 -0
- data/vendor/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h +174 -0
- data/vendor/Eigen/src/SparseCore/AmbiVector.h +378 -0
- data/vendor/Eigen/src/SparseCore/CompressedStorage.h +274 -0
- data/vendor/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +352 -0
- data/vendor/Eigen/src/SparseCore/MappedSparseMatrix.h +67 -0
- data/vendor/Eigen/src/SparseCore/SparseAssign.h +270 -0
- data/vendor/Eigen/src/SparseCore/SparseBlock.h +571 -0
- data/vendor/Eigen/src/SparseCore/SparseColEtree.h +206 -0
- data/vendor/Eigen/src/SparseCore/SparseCompressedBase.h +370 -0
- data/vendor/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +722 -0
- data/vendor/Eigen/src/SparseCore/SparseCwiseUnaryOp.h +150 -0
- data/vendor/Eigen/src/SparseCore/SparseDenseProduct.h +342 -0
- data/vendor/Eigen/src/SparseCore/SparseDiagonalProduct.h +138 -0
- data/vendor/Eigen/src/SparseCore/SparseDot.h +98 -0
- data/vendor/Eigen/src/SparseCore/SparseFuzzy.h +29 -0
- data/vendor/Eigen/src/SparseCore/SparseMap.h +305 -0
- data/vendor/Eigen/src/SparseCore/SparseMatrix.h +1518 -0
- data/vendor/Eigen/src/SparseCore/SparseMatrixBase.h +398 -0
- data/vendor/Eigen/src/SparseCore/SparsePermutation.h +178 -0
- data/vendor/Eigen/src/SparseCore/SparseProduct.h +181 -0
- data/vendor/Eigen/src/SparseCore/SparseRedux.h +49 -0
- data/vendor/Eigen/src/SparseCore/SparseRef.h +397 -0
- data/vendor/Eigen/src/SparseCore/SparseSelfAdjointView.h +659 -0
- data/vendor/Eigen/src/SparseCore/SparseSolverBase.h +124 -0
- data/vendor/Eigen/src/SparseCore/SparseSparseProductWithPruning.h +198 -0
- data/vendor/Eigen/src/SparseCore/SparseTranspose.h +92 -0
- data/vendor/Eigen/src/SparseCore/SparseTriangularView.h +189 -0
- data/vendor/Eigen/src/SparseCore/SparseUtil.h +186 -0
- data/vendor/Eigen/src/SparseCore/SparseVector.h +478 -0
- data/vendor/Eigen/src/SparseCore/SparseView.h +254 -0
- data/vendor/Eigen/src/SparseCore/TriangularSolver.h +315 -0
- data/vendor/Eigen/src/SparseLU/SparseLU.h +923 -0
- data/vendor/Eigen/src/SparseLU/SparseLUImpl.h +66 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_Memory.h +226 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_Structs.h +110 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h +375 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_Utils.h +80 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_column_bmod.h +181 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_column_dfs.h +179 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h +107 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_gemm_kernel.h +280 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h +126 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_kernel_bmod.h +130 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_panel_bmod.h +223 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_panel_dfs.h +258 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_pivotL.h +137 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_pruneL.h +136 -0
- data/vendor/Eigen/src/SparseLU/SparseLU_relax_snode.h +83 -0
- data/vendor/Eigen/src/SparseQR/SparseQR.h +758 -0
- data/vendor/Eigen/src/StlSupport/StdDeque.h +116 -0
- data/vendor/Eigen/src/StlSupport/StdList.h +106 -0
- data/vendor/Eigen/src/StlSupport/StdVector.h +131 -0
- data/vendor/Eigen/src/StlSupport/details.h +84 -0
- data/vendor/Eigen/src/SuperLUSupport/SuperLUSupport.h +1025 -0
- data/vendor/Eigen/src/UmfPackSupport/UmfPackSupport.h +642 -0
- data/vendor/Eigen/src/misc/Image.h +82 -0
- data/vendor/Eigen/src/misc/Kernel.h +79 -0
- data/vendor/Eigen/src/misc/RealSvd2x2.h +55 -0
- data/vendor/Eigen/src/misc/blas.h +440 -0
- data/vendor/Eigen/src/misc/lapack.h +152 -0
- data/vendor/Eigen/src/misc/lapacke.h +16292 -0
- data/vendor/Eigen/src/misc/lapacke_mangling.h +17 -0
- data/vendor/Eigen/src/plugins/ArrayCwiseBinaryOps.h +358 -0
- data/vendor/Eigen/src/plugins/ArrayCwiseUnaryOps.h +696 -0
- data/vendor/Eigen/src/plugins/BlockMethods.h +1442 -0
- data/vendor/Eigen/src/plugins/CommonCwiseBinaryOps.h +115 -0
- data/vendor/Eigen/src/plugins/CommonCwiseUnaryOps.h +177 -0
- data/vendor/Eigen/src/plugins/IndexedViewMethods.h +262 -0
- data/vendor/Eigen/src/plugins/MatrixCwiseBinaryOps.h +152 -0
- data/vendor/Eigen/src/plugins/MatrixCwiseUnaryOps.h +95 -0
- data/vendor/Eigen/src/plugins/ReshapedMethods.h +149 -0
- data/vendor/aarand/aarand.hpp +114 -0
- data/vendor/annoy/annoylib.h +1495 -0
- data/vendor/annoy/kissrandom.h +120 -0
- data/vendor/annoy/mman.h +242 -0
- data/vendor/hnswlib/bruteforce.h +152 -0
- data/vendor/hnswlib/hnswalg.h +1192 -0
- data/vendor/hnswlib/hnswlib.h +108 -0
- data/vendor/hnswlib/space_ip.h +282 -0
- data/vendor/hnswlib/space_l2.h +281 -0
- data/vendor/hnswlib/visited_list_pool.h +79 -0
- data/vendor/irlba/irlba.hpp +575 -0
- data/vendor/irlba/lanczos.hpp +212 -0
- data/vendor/irlba/parallel.hpp +474 -0
- data/vendor/irlba/utils.hpp +224 -0
- data/vendor/irlba/wrappers.hpp +228 -0
- data/vendor/kmeans/Base.hpp +75 -0
- data/vendor/kmeans/Details.hpp +79 -0
- data/vendor/kmeans/HartiganWong.hpp +492 -0
- data/vendor/kmeans/InitializeKmeansPP.hpp +144 -0
- data/vendor/kmeans/InitializeNone.hpp +44 -0
- data/vendor/kmeans/InitializePCAPartition.hpp +309 -0
- data/vendor/kmeans/InitializeRandom.hpp +91 -0
- data/vendor/kmeans/Kmeans.hpp +161 -0
- data/vendor/kmeans/Lloyd.hpp +134 -0
- data/vendor/kmeans/MiniBatch.hpp +269 -0
- data/vendor/kmeans/QuickSearch.hpp +179 -0
- data/vendor/kmeans/compute_centroids.hpp +32 -0
- data/vendor/kmeans/compute_wcss.hpp +27 -0
- data/vendor/kmeans/is_edge_case.hpp +42 -0
- data/vendor/kmeans/random.hpp +55 -0
- data/vendor/knncolle/Annoy/Annoy.hpp +193 -0
- data/vendor/knncolle/BruteForce/BruteForce.hpp +120 -0
- data/vendor/knncolle/Hnsw/Hnsw.hpp +225 -0
- data/vendor/knncolle/Kmknn/Kmknn.hpp +286 -0
- data/vendor/knncolle/VpTree/VpTree.hpp +256 -0
- data/vendor/knncolle/knncolle.hpp +34 -0
- data/vendor/knncolle/utils/Base.hpp +100 -0
- data/vendor/knncolle/utils/NeighborQueue.hpp +94 -0
- data/vendor/knncolle/utils/distances.hpp +98 -0
- data/vendor/knncolle/utils/find_nearest_neighbors.hpp +112 -0
- data/vendor/powerit/PowerIterations.hpp +157 -0
- data/vendor/umappp/NeighborList.hpp +37 -0
- data/vendor/umappp/Umap.hpp +662 -0
- data/vendor/umappp/combine_neighbor_sets.hpp +95 -0
- data/vendor/umappp/find_ab.hpp +157 -0
- data/vendor/umappp/neighbor_similarities.hpp +136 -0
- data/vendor/umappp/optimize_layout.hpp +285 -0
- data/vendor/umappp/spectral_init.hpp +181 -0
- data/vendor/umappp/umappp.hpp +13 -0
- metadata +465 -0
@@ -0,0 +1,492 @@
|
|
1
|
+
#ifndef KMEANS_HARTIGAN_WONG_HPP
|
2
|
+
#define KMEANS_HARTIGAN_WONG_HPP
|
3
|
+
|
4
|
+
#include <vector>
|
5
|
+
#include <algorithm>
|
6
|
+
#include <numeric>
|
7
|
+
#include <cstdint>
|
8
|
+
#include <stdexcept>
|
9
|
+
#include <limits>
|
10
|
+
|
11
|
+
#include "Base.hpp"
|
12
|
+
#include "Details.hpp"
|
13
|
+
#include "compute_centroids.hpp"
|
14
|
+
#include "compute_wcss.hpp"
|
15
|
+
#include "is_edge_case.hpp"
|
16
|
+
|
17
|
+
/**
|
18
|
+
* @file HartiganWong.hpp
|
19
|
+
*
|
20
|
+
* @brief Implements the Hartigan-Wong algorithm for k-means clustering.
|
21
|
+
*/
|
22
|
+
|
23
|
+
namespace kmeans {
|
24
|
+
|
25
|
+
/**
|
26
|
+
* @brief Implements the Hartigan-Wong algorithm for k-means clustering.
|
27
|
+
*
|
28
|
+
* The Hartigan-Wong algorithm performs several iterations of transferring points between clusters,
|
29
|
+
* involving a computationally expensive "optimal transfer" step that checks each observation against each cluster for the lowest squared distance;
|
30
|
+
* followed by a cheaper "quick transfer" step, which iterates between the best and second-best cluster choices for each observation.
|
31
|
+
* The latter enables rapid exploration of the local solution space without the unnecessary expense of repeatedly comparing to all clusters for all observations.
|
32
|
+
* In addition, each distance calculation to a cluster accounts for the shift in the means when a point is transferred.
|
33
|
+
* The algorithm terminates when no observation wishes to transfer between clusters.
|
34
|
+
*
|
35
|
+
* This implementation is derived from the Fortran code underlying the `kmeans` function in the **stats** R package,
|
36
|
+
* which in turn is derived from Hartigan and Wong (1979).
|
37
|
+
*
|
38
|
+
* @tparam DATA_t Floating-point type for the data and centroids.
|
39
|
+
* @tparam CLUSTER_t Integer type for the cluster assignments.
|
40
|
+
* @tparam INDEX_t Integer type for the observation index.
|
41
|
+
* This should have a maximum positive value that is at least 50 times greater than the maximum expected number of observations.
|
42
|
+
*
|
43
|
+
* @see
|
44
|
+
* Hartigan, J. A. and Wong, M. A. (1979).
|
45
|
+
* Algorithm AS 136: A K-means clustering algorithm.
|
46
|
+
* _Applied Statistics_, 28, 100-108.
|
47
|
+
*/
|
48
|
+
template<typename DATA_t = double, typename CLUSTER_t = int, typename INDEX_t = int>
|
49
|
+
class HartiganWong : public Refine<DATA_t, CLUSTER_t, INDEX_t> {
|
50
|
+
int num_dim;
|
51
|
+
INDEX_t num_obs;
|
52
|
+
const DATA_t* data_ptr;
|
53
|
+
|
54
|
+
CLUSTER_t num_centers;
|
55
|
+
DATA_t* centers_ptr;
|
56
|
+
|
57
|
+
// Array arguments in the same order as supplied to R's kmns_ function.
|
58
|
+
CLUSTER_t * ic1;
|
59
|
+
std::vector<CLUSTER_t> ic2;
|
60
|
+
std::vector<INDEX_t> nc;
|
61
|
+
std::vector<DATA_t> an1, an2;
|
62
|
+
std::vector<INDEX_t> ncp;
|
63
|
+
std::vector<DATA_t> d;
|
64
|
+
std::vector<uint8_t> itran;
|
65
|
+
std::vector<INDEX_t> live;
|
66
|
+
|
67
|
+
private:
|
68
|
+
static constexpr double big = 1e30; // Define BIG to be a very large positive number
|
69
|
+
|
70
|
+
/* The following comparisons to 'ncp' are wrapped in functions to account
|
71
|
+
* for the fact that we need to shift all the 'ncp' values by two to give
|
72
|
+
* some space for the error codes when dealing with unsigned integers in
|
73
|
+
* 'INDEX_t'. All interactions with 'ncp' should occur via these utilities.
|
74
|
+
*/
|
75
|
+
static constexpr INDEX_t ncp_init = 0;
|
76
|
+
static constexpr INDEX_t ncp_unchanged = 1;
|
77
|
+
static constexpr INDEX_t ncp_shift = 2;
|
78
|
+
|
79
|
+
void initialize_ncp() {
|
80
|
+
std::fill(ncp.begin(), ncp.end(), ncp_init);
|
81
|
+
}
|
82
|
+
|
83
|
+
void reset_ncp() {
|
84
|
+
std::fill(ncp.begin(), ncp.end(), ncp_unchanged);
|
85
|
+
}
|
86
|
+
|
87
|
+
void set_ncp(INDEX_t obs, INDEX_t val) {
|
88
|
+
ncp[obs] = val + ncp_shift;
|
89
|
+
}
|
90
|
+
|
91
|
+
bool unchanged_ncp(INDEX_t obs) const {
|
92
|
+
return ncp[obs] == ncp_unchanged;
|
93
|
+
}
|
94
|
+
|
95
|
+
bool lt_ncp(INDEX_t obs, INDEX_t val) const {
|
96
|
+
return ncp[obs] > val + ncp_shift;
|
97
|
+
}
|
98
|
+
|
99
|
+
bool le_ncp(INDEX_t obs, INDEX_t val) const {
|
100
|
+
return ncp[obs] >= val + ncp_shift;
|
101
|
+
}
|
102
|
+
|
103
|
+
public:
|
104
|
+
/**
|
105
|
+
* @brief Default parameter values for `HartiganWong`.
|
106
|
+
*/
|
107
|
+
struct Defaults {
|
108
|
+
/**
|
109
|
+
* See `HartiganWong::set_max_iterations()`.
|
110
|
+
*/
|
111
|
+
static constexpr int max_iterations = 10;
|
112
|
+
};
|
113
|
+
|
114
|
+
private:
|
115
|
+
int maxiter = Defaults::max_iterations;
|
116
|
+
|
117
|
+
public:
|
118
|
+
/**
|
119
|
+
* @param m Maximum number of iterations.
|
120
|
+
* More iterations increase the opportunity for convergence at the cost of more computational time.
|
121
|
+
*
|
122
|
+
* @return A reference to this `HartiganWong` object.
|
123
|
+
*/
|
124
|
+
HartiganWong& set_max_iterations(int m = Defaults::max_iterations) {
|
125
|
+
maxiter = m;
|
126
|
+
return *this;
|
127
|
+
}
|
128
|
+
|
129
|
+
public:
|
130
|
+
Details<DATA_t, INDEX_t> run(int ndim, INDEX_t nobs, const DATA_t* data, CLUSTER_t ncenters, DATA_t* centers, CLUSTER_t* clusters) {
|
131
|
+
num_dim = ndim;
|
132
|
+
num_obs = nobs;
|
133
|
+
data_ptr = data;
|
134
|
+
num_centers = ncenters;
|
135
|
+
centers_ptr = centers;
|
136
|
+
ic1 = clusters;
|
137
|
+
|
138
|
+
// Sizes taken from the .Fortran() call in stats::kmeans().
|
139
|
+
ic2.resize(num_obs);
|
140
|
+
nc.resize(num_centers);
|
141
|
+
an1.resize(num_centers);
|
142
|
+
an2.resize(num_centers);
|
143
|
+
d.resize(num_obs);
|
144
|
+
|
145
|
+
/* ITRAN(L) = 1 if cluster L is updated in the quick-transfer stage,
|
146
|
+
* = 0 otherwise
|
147
|
+
* In the optimal-transfer stage, NCP(L) stores the step at which
|
148
|
+
* cluster L is last updated.
|
149
|
+
* In the quick-transfer stage, NCP(L) stores the step at which
|
150
|
+
* cluster L is last updated plus M.
|
151
|
+
*/
|
152
|
+
ncp.resize(num_centers);
|
153
|
+
itran.resize(num_centers);
|
154
|
+
live.resize(num_centers);
|
155
|
+
|
156
|
+
if (is_edge_case(num_obs, num_centers)) {
|
157
|
+
return process_edge_case(num_dim, num_obs, data_ptr, num_centers, centers_ptr, ic1);
|
158
|
+
}
|
159
|
+
|
160
|
+
/* For each point I, find its two closest centres, IC1(I) and
|
161
|
+
* IC2(I). Assign it to IC1(I).
|
162
|
+
*/
|
163
|
+
#pragma omp parallel for
|
164
|
+
for (INDEX_t obs = 0; obs < num_obs; ++obs) {
|
165
|
+
auto& best = ic1[obs];
|
166
|
+
best = 0;
|
167
|
+
DATA_t best_dist = squared_distance_from_cluster(obs, best);
|
168
|
+
|
169
|
+
auto& second = ic2[obs];
|
170
|
+
second = 1;
|
171
|
+
DATA_t second_dist = squared_distance_from_cluster(obs, second);
|
172
|
+
|
173
|
+
if (best_dist > second_dist) {
|
174
|
+
std::swap(best, second);
|
175
|
+
std::swap(best_dist, second_dist);
|
176
|
+
}
|
177
|
+
|
178
|
+
for (CLUSTER_t cen = 2; cen < num_centers; ++cen) {
|
179
|
+
DATA_t candidate_dist = squared_distance_from_cluster(obs, cen);
|
180
|
+
if (candidate_dist < second_dist) {
|
181
|
+
second_dist = candidate_dist;
|
182
|
+
second = cen;
|
183
|
+
if (candidate_dist < best_dist) {
|
184
|
+
std::swap(best_dist, second_dist);
|
185
|
+
std::swap(best, second);
|
186
|
+
}
|
187
|
+
}
|
188
|
+
}
|
189
|
+
}
|
190
|
+
|
191
|
+
/* Update cluster centres to be the average of points contained
|
192
|
+
* within them.
|
193
|
+
* NC(L) := #{units in cluster L}, L = 1..K
|
194
|
+
*/
|
195
|
+
std::fill(nc.begin(), nc.end(), 0);
|
196
|
+
for (INDEX_t obs = 0; obs < num_obs; ++obs) {
|
197
|
+
++nc[ic1[obs]];
|
198
|
+
}
|
199
|
+
|
200
|
+
compute_centroids(num_dim, num_obs, data_ptr, num_centers, centers_ptr, ic1, nc);
|
201
|
+
|
202
|
+
// Check to see if there is any empty cluster at this stage
|
203
|
+
for (CLUSTER_t cen = 0; cen < num_centers; ++cen) {
|
204
|
+
if (nc[cen] == 0) {
|
205
|
+
return Details<DATA_t, INDEX_t>(0, 1); // i.e., ifault = 1 here.
|
206
|
+
}
|
207
|
+
|
208
|
+
/* Initialize AN1, AN2.
|
209
|
+
* AN1(L) = NC(L) / (NC(L) - 1)
|
210
|
+
* AN2(L) = NC(L) / (NC(L) + 1)
|
211
|
+
*/
|
212
|
+
const DATA_t num = nc[cen];
|
213
|
+
an2[cen] = num / (num + 1);
|
214
|
+
an1[cen] = (num > 1 ? num / (num - 1) : big);
|
215
|
+
}
|
216
|
+
|
217
|
+
INDEX_t indx = 0;
|
218
|
+
if (std::numeric_limits<INDEX_t>::max() / 50 < num_obs) {
|
219
|
+
throw std::runtime_error("too many observations for the index integer type");
|
220
|
+
}
|
221
|
+
INDEX_t imaxqtr = num_obs * 50; // default derived from stats::kmeans()
|
222
|
+
|
223
|
+
initialize_ncp();
|
224
|
+
std::fill(itran.begin(), itran.end(), true);
|
225
|
+
std::fill(live.begin(), live.end(), 0);
|
226
|
+
int iter = 0;
|
227
|
+
int ifault = 0;
|
228
|
+
|
229
|
+
for (iter = 1; iter <= maxiter; ++iter) {
|
230
|
+
|
231
|
+
/* OPtimal-TRAnsfer stage: there is only one pass through the data.
|
232
|
+
* Each point is re-allocated, if necessary, to the cluster that will
|
233
|
+
* induce the maximum reduction in within-cluster sum of squares.
|
234
|
+
*/
|
235
|
+
optimal_transfer(indx);
|
236
|
+
|
237
|
+
// Stop if no transfer took place in the last M optimal transfer steps.
|
238
|
+
if (indx == num_obs) {
|
239
|
+
break;
|
240
|
+
}
|
241
|
+
|
242
|
+
/* Quick-TRANSfer stage: Each point is tested in turn to see if it should
|
243
|
+
* be re-allocated to the cluster to which it is most likely to be
|
244
|
+
* transferred, IC2(I), from its present cluster, IC1(I).
|
245
|
+
* Loop through the data until no further change is to take place.
|
246
|
+
*/
|
247
|
+
quick_transfer(indx, imaxqtr);
|
248
|
+
|
249
|
+
if (imaxqtr < 0) {
|
250
|
+
ifault = 4;
|
251
|
+
break;
|
252
|
+
}
|
253
|
+
|
254
|
+
// If there are only two clusters, there is no need to re-enter the optimal transfer stage.
|
255
|
+
if (num_centers == 2) {
|
256
|
+
break;
|
257
|
+
}
|
258
|
+
|
259
|
+
// NCP has to be reset before entering optimal_transfer().
|
260
|
+
reset_ncp();
|
261
|
+
}
|
262
|
+
|
263
|
+
/* Since the specified number of iterations has been exceeded, set
|
264
|
+
* IFAULT = 2. This may indicate unforeseen looping.
|
265
|
+
*/
|
266
|
+
if (iter == maxiter + 1) {
|
267
|
+
ifault = 2;
|
268
|
+
}
|
269
|
+
|
270
|
+
compute_centroids(num_dim, num_obs, data_ptr, num_centers, centers_ptr, ic1, nc);
|
271
|
+
return Details(
|
272
|
+
std::move(nc),
|
273
|
+
compute_wcss(num_dim, num_obs, data_ptr, num_centers, centers_ptr, ic1),
|
274
|
+
iter,
|
275
|
+
ifault
|
276
|
+
);
|
277
|
+
}
|
278
|
+
|
279
|
+
private:
|
280
|
+
#ifdef DEBUG
|
281
|
+
template<class T>
|
282
|
+
void print_vector(const T& vec, const char* msg) {
|
283
|
+
std::cout << msg << std::endl;
|
284
|
+
for (auto c : vec) {
|
285
|
+
std::cout << c << " ";
|
286
|
+
}
|
287
|
+
std::cout << std::endl;
|
288
|
+
}
|
289
|
+
#endif
|
290
|
+
|
291
|
+
DATA_t squared_distance_from_cluster(INDEX_t pt, CLUSTER_t clust) const {
|
292
|
+
const DATA_t* acopy = data_ptr + pt * num_dim;
|
293
|
+
const DATA_t* ccopy = centers_ptr + clust * num_dim;
|
294
|
+
DATA_t output = 0;
|
295
|
+
for (int dim = 0; dim < num_dim; ++dim, ++acopy, ++ccopy) {
|
296
|
+
output += (*acopy - *ccopy) * (*acopy - *ccopy);
|
297
|
+
}
|
298
|
+
return output;
|
299
|
+
}
|
300
|
+
|
301
|
+
private:
|
302
|
+
/* ALGORITHM AS 136.1 APPL. STATIST. (1979) VOL.28, NO.1
|
303
|
+
* This is the OPtimal TRAnsfer stage.
|
304
|
+
* ----------------------
|
305
|
+
* Each point is re-allocated, if necessary, to the cluster that
|
306
|
+
* will induce a maximum reduction in the within-cluster sum of
|
307
|
+
* squares.
|
308
|
+
*/
|
309
|
+
void optimal_transfer(INDEX_t& indx) {
|
310
|
+
/* If cluster L is updated in the last quick-transfer stage, it
|
311
|
+
* belongs to the live set throughout this stage. Otherwise, at
|
312
|
+
* each step, it is not in the live set if it has not been updated
|
313
|
+
* in the last M optimal transfer steps. (AL: M being a synonym for
|
314
|
+
* the number of observations, here and in other functions.)
|
315
|
+
*/
|
316
|
+
for (CLUSTER_t cen = 0; cen < num_centers; ++cen) {
|
317
|
+
if (itran[cen]) {
|
318
|
+
live[cen] = num_obs; // AL: using 0-index, so no need for +1.
|
319
|
+
}
|
320
|
+
}
|
321
|
+
|
322
|
+
for (INDEX_t obs = 0; obs < num_obs; ++obs) {
|
323
|
+
++indx;
|
324
|
+
auto l1 = ic1[obs];
|
325
|
+
|
326
|
+
// If point I is the only member of cluster L1, no transfer.
|
327
|
+
if (nc[l1] != 1) {
|
328
|
+
// If L1 has not yet been updated in this stage, no need to re-compute D(I).
|
329
|
+
if (!unchanged_ncp(l1)) {
|
330
|
+
d[obs] = squared_distance_from_cluster(obs, l1) * an1[l1];
|
331
|
+
}
|
332
|
+
|
333
|
+
// Find the cluster with minimum R2.
|
334
|
+
auto l2 = ic2[obs];
|
335
|
+
auto ll = l2;
|
336
|
+
DATA_t r2 = squared_distance_from_cluster(obs, l2) * an2[l2];
|
337
|
+
|
338
|
+
for (CLUSTER_t cen = 0; cen < num_centers; ++cen) {
|
339
|
+
/* If I >= LIVE(L1), then L1 is not in the live set. If this is
|
340
|
+
* true, we only need to consider clusters that are in the live
|
341
|
+
* set for possible transfer of point I. Otherwise, we need to
|
342
|
+
* consider all possible clusters.
|
343
|
+
*/
|
344
|
+
if (obs >= live[l1] && obs >= live[cen] || cen == l1 || cen == ll) {
|
345
|
+
continue;
|
346
|
+
}
|
347
|
+
|
348
|
+
DATA_t rr = r2 / an2[cen];
|
349
|
+
DATA_t dc = squared_distance_from_cluster(obs, cen);
|
350
|
+
if (dc < rr) {
|
351
|
+
r2 = dc * an2[cen];
|
352
|
+
l2 = cen;
|
353
|
+
}
|
354
|
+
}
|
355
|
+
|
356
|
+
if (r2 >= d[obs]) {
|
357
|
+
// If no transfer is necessary, L2 is the new IC2(I).
|
358
|
+
ic2[obs] = l2;
|
359
|
+
|
360
|
+
} else {
|
361
|
+
/* Update cluster centres, LIVE, NCP, AN1 & AN2 for clusters L1 and
|
362
|
+
* L2, and update IC1(I) & IC2(I).
|
363
|
+
*/
|
364
|
+
indx = 0;
|
365
|
+
live[l1] = num_obs + obs;
|
366
|
+
live[l2] = num_obs + obs;
|
367
|
+
set_ncp(l1, obs);
|
368
|
+
set_ncp(l2, obs);
|
369
|
+
|
370
|
+
transfer_point(obs, l1, l2);
|
371
|
+
}
|
372
|
+
}
|
373
|
+
|
374
|
+
if (indx == num_obs) {
|
375
|
+
return;
|
376
|
+
}
|
377
|
+
}
|
378
|
+
|
379
|
+
for (CLUSTER_t cen = 0; cen < num_centers; ++cen) {
|
380
|
+
itran[cen] = false;
|
381
|
+
|
382
|
+
// LIVE(L) has to be decreased by M before re-entering OPTRA.
|
383
|
+
// This means that if I >= LIVE(L1) in the next OPTRA call,
|
384
|
+
// the last update must be >= M steps ago, as we effectively
|
385
|
+
// 'lapped' the previous update for this cluster.
|
386
|
+
live[cen] -= num_obs;
|
387
|
+
}
|
388
|
+
|
389
|
+
return;
|
390
|
+
}
|
391
|
+
|
392
|
+
private:
|
393
|
+
/* ALGORITHM AS 136.2 APPL. STATIST. (1979) VOL.28, NO.1
|
394
|
+
* This is the Quick TRANsfer stage.
|
395
|
+
* --------------------
|
396
|
+
* IC1(I) is the cluster which point I belongs to.
|
397
|
+
* IC2(I) is the cluster which point I is most likely to be
|
398
|
+
* transferred to.
|
399
|
+
*
|
400
|
+
* For each point I, IC1(I) & IC2(I) are switched, if necessary, to
|
401
|
+
* reduce within-cluster sum of squares. The cluster centres are
|
402
|
+
* updated after each step.
|
403
|
+
*/
|
404
|
+
void quick_transfer (INDEX_t& indx, INDEX_t& imaxqtr) {
|
405
|
+
INDEX_t icoun = 0;
|
406
|
+
INDEX_t istep = 0;
|
407
|
+
|
408
|
+
while (1) {
|
409
|
+
for (INDEX_t obs = 0; obs < num_obs; ++obs) {
|
410
|
+
++icoun;
|
411
|
+
auto l1 = ic1[obs];
|
412
|
+
|
413
|
+
// point I is the only member of cluster L1, no transfer.
|
414
|
+
if (nc[l1] != 1) {
|
415
|
+
|
416
|
+
/* NCP(L) is equal to the step at which cluster L is last updated plus M.
|
417
|
+
* (AL: M is the notation for the number of observations, a.k.a. 'num_obs').
|
418
|
+
*
|
419
|
+
* If ISTEP > NCP(L1), no need to re-compute distance from point I to
|
420
|
+
* cluster L1. Note that if cluster L1 is last updated exactly M
|
421
|
+
* steps ago, we still need to compute the distance from point I to
|
422
|
+
* cluster L1.
|
423
|
+
*/
|
424
|
+
if (le_ncp(l1, istep)) {
|
425
|
+
d[obs] = squared_distance_from_cluster(obs, l1) * an1[l1];
|
426
|
+
}
|
427
|
+
|
428
|
+
// If ISTEP >= both NCP(L1) & NCP(L2) there will be no transfer of point I at this step.
|
429
|
+
auto l2 = ic2[obs];
|
430
|
+
if (lt_ncp(l1, istep) || lt_ncp(l2, istep)) {
|
431
|
+
if (squared_distance_from_cluster(obs, l2) < d[obs] / an2[l2]) {
|
432
|
+
/* Update cluster centres, NCP, NC, ITRAN, AN1 & AN2 for clusters
|
433
|
+
* L1 & L2. Also update IC1(I) & IC2(I). Note that if any
|
434
|
+
* updating occurs in this stage, INDX is set back to 0.
|
435
|
+
*/
|
436
|
+
icoun = 0;
|
437
|
+
indx = 0;
|
438
|
+
|
439
|
+
itran[l1] = true;
|
440
|
+
itran[l2] = true;
|
441
|
+
set_ncp(l1, istep + num_obs);
|
442
|
+
set_ncp(l2, istep + num_obs);
|
443
|
+
transfer_point(obs, l1, l2);
|
444
|
+
}
|
445
|
+
}
|
446
|
+
}
|
447
|
+
|
448
|
+
// If no re-allocation took place in the last M steps, return.
|
449
|
+
if (icoun == num_obs) {
|
450
|
+
return;
|
451
|
+
}
|
452
|
+
|
453
|
+
// AL: incrementing ISTEP after checks against NCP(L1), to avoid off-by-one
|
454
|
+
// errors after switching to zero-indexing for the observations.
|
455
|
+
++istep;
|
456
|
+
if (istep >= imaxqtr) {
|
457
|
+
imaxqtr = -1;
|
458
|
+
return;
|
459
|
+
}
|
460
|
+
}
|
461
|
+
}
|
462
|
+
}
|
463
|
+
|
464
|
+
private:
|
465
|
+
void transfer_point(INDEX_t obs, CLUSTER_t l1, CLUSTER_t l2) {
|
466
|
+
const DATA_t al1 = nc[l1], alw = al1 - 1;
|
467
|
+
const DATA_t al2 = nc[l2], alt = al2 + 1;
|
468
|
+
|
469
|
+
auto copy1 = centers_ptr + l1 * num_dim;
|
470
|
+
auto copy2 = centers_ptr + l2 * num_dim;
|
471
|
+
auto acopy = data_ptr + obs * num_dim;
|
472
|
+
for (int dim = 0; dim < num_dim; ++dim, ++copy1, ++copy2, ++acopy) {
|
473
|
+
*copy1 = (*copy1 * al1 - *acopy) / alw;
|
474
|
+
*copy2 = (*copy2 * al2 + *acopy) / alt;
|
475
|
+
}
|
476
|
+
|
477
|
+
--nc[l1];
|
478
|
+
++nc[l2];
|
479
|
+
|
480
|
+
an2[l1] = alw / al1;
|
481
|
+
an1[l1] = (alw > 1 ? alw / (alw - 1) : big);
|
482
|
+
an1[l2] = alt / al2;
|
483
|
+
an2[l2] = alt / (alt + 1);
|
484
|
+
|
485
|
+
ic1[obs] = l2;
|
486
|
+
ic2[obs] = l1;
|
487
|
+
}
|
488
|
+
};
|
489
|
+
|
490
|
+
}
|
491
|
+
|
492
|
+
#endif
|
@@ -0,0 +1,144 @@
|
|
1
|
+
#ifndef KMEANS_INITIALIZE_KMEANSPP_HPP
|
2
|
+
#define KMEANS_INITIALIZE_KMEANSPP_HPP
|
3
|
+
|
4
|
+
#include <vector>
|
5
|
+
#include <algorithm>
|
6
|
+
#include <cstdint>
|
7
|
+
|
8
|
+
#include "Base.hpp"
|
9
|
+
#include "InitializeRandom.hpp"
|
10
|
+
#include "random.hpp"
|
11
|
+
|
12
|
+
/**
|
13
|
+
* @file InitializeKmeansPP.hpp
|
14
|
+
*
|
15
|
+
* @brief Class for **kmeans++** initialization.
|
16
|
+
*/
|
17
|
+
|
18
|
+
namespace kmeans {
|
19
|
+
|
20
|
+
/**
|
21
|
+
* @brief Implements the **k-means++** initialization of Arthur and Vassilvitskii (2007).
|
22
|
+
*
|
23
|
+
* This approach involves the selection of starting points via iterations of weighted sampling,
|
24
|
+
* where the sampling probability for each point is proportional to the squared distance to the closest starting point that was chosen in any of the previous iterations.
|
25
|
+
* The aim is to obtain well-separated starting points to encourage the formation of suitable clusters.
|
26
|
+
*
|
27
|
+
* @tparam DATA_t Floating-point type for the data and centroids.
|
28
|
+
* @tparam CLUSTER_t Integer type for the cluster index.
|
29
|
+
* @tparam INDEX_t Integer type for the observation index.
|
30
|
+
*
|
31
|
+
* @see
|
32
|
+
* Arthur, D. and Vassilvitskii, S. (2007).
|
33
|
+
* k-means++: the advantages of careful seeding.
|
34
|
+
* _Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms_, 1027-1035.
|
35
|
+
*/
|
36
|
+
template<typename DATA_t = double, typename CLUSTER_t = int, typename INDEX_t = int>
|
37
|
+
class InitializeKmeansPP : public Initialize<DATA_t, CLUSTER_t, INDEX_t> {
|
38
|
+
public:
|
39
|
+
/**
|
40
|
+
* @brief Default parameter settings.
|
41
|
+
*/
|
42
|
+
struct Defaults {
|
43
|
+
/**
|
44
|
+
* See `set_seed()` for more details.
|
45
|
+
*/
|
46
|
+
static constexpr uint64_t seed = 6523u;
|
47
|
+
};
|
48
|
+
|
49
|
+
/**
|
50
|
+
* @param Random seed to use to construct the PRNG prior to sampling.
|
51
|
+
*
|
52
|
+
* @return A reference to this `InitializeKmeansPP` object.
|
53
|
+
*/
|
54
|
+
InitializeKmeansPP& set_seed(uint64_t s = Defaults::seed) {
|
55
|
+
seed = s;
|
56
|
+
return *this;
|
57
|
+
}
|
58
|
+
private:
|
59
|
+
uint64_t seed = Defaults::seed;
|
60
|
+
|
61
|
+
public:
|
62
|
+
/**
|
63
|
+
* @cond
|
64
|
+
*/
|
65
|
+
std::vector<INDEX_t> run(int ndim, INDEX_t nobs, const DATA_t* data, CLUSTER_t ncenters) {
|
66
|
+
std::vector<DATA_t> mindist(nobs, 1);
|
67
|
+
std::vector<DATA_t> cumulative(nobs);
|
68
|
+
std::vector<INDEX_t> sofar;
|
69
|
+
sofar.reserve(ncenters);
|
70
|
+
std::mt19937_64 eng(seed);
|
71
|
+
|
72
|
+
for (CLUSTER_t cen = 0; cen < ncenters; ++cen) {
|
73
|
+
INDEX_t counter = 0;
|
74
|
+
if (!sofar.empty()) {
|
75
|
+
auto last = sofar.back();
|
76
|
+
|
77
|
+
#pragma omp parallel for
|
78
|
+
for (INDEX_t obs = 0; obs < nobs; ++obs) {
|
79
|
+
if (mindist[obs]) {
|
80
|
+
const DATA_t* acopy = data + obs * ndim;
|
81
|
+
const DATA_t* scopy = data + last * ndim;
|
82
|
+
DATA_t r2 = 0;
|
83
|
+
for (int dim = 0; dim < ndim; ++dim, ++acopy, ++scopy) {
|
84
|
+
r2 += (*acopy - *scopy) * (*acopy - *scopy);
|
85
|
+
}
|
86
|
+
|
87
|
+
if (cen == 1 || r2 < mindist[obs]) {
|
88
|
+
mindist[obs] = r2;
|
89
|
+
}
|
90
|
+
}
|
91
|
+
}
|
92
|
+
} else {
|
93
|
+
counter = nobs;
|
94
|
+
}
|
95
|
+
|
96
|
+
cumulative[0] = mindist[0];
|
97
|
+
for (INDEX_t i = 1; i < nobs; ++i) {
|
98
|
+
cumulative[i] = cumulative[i-1] + mindist[i];
|
99
|
+
}
|
100
|
+
|
101
|
+
const auto total = cumulative.back();
|
102
|
+
if (total == 0) { // a.k.a. only duplicates left.
|
103
|
+
break;
|
104
|
+
}
|
105
|
+
|
106
|
+
auto chosen_id = weighted_sample(cumulative, mindist, nobs, eng);
|
107
|
+
mindist[chosen_id] = 0;
|
108
|
+
sofar.push_back(chosen_id);
|
109
|
+
}
|
110
|
+
|
111
|
+
return sofar;
|
112
|
+
}
|
113
|
+
/**
|
114
|
+
* @endcond
|
115
|
+
*/
|
116
|
+
|
117
|
+
public:
|
118
|
+
/*
|
119
|
+
* @param ndim Number of dimensions.
|
120
|
+
* @param nobs Number of observations.
|
121
|
+
* @param data Pointer to an array where the dimensions are rows and the observations are columns.
|
122
|
+
* Data should be stored in column-major format.
|
123
|
+
* @param ncenters Number of centers to pick.
|
124
|
+
* @param[out] centers Pointer to a `ndim`-by-`ncenters` array where columns are cluster centers and rows are dimensions.
|
125
|
+
* On output, this will contain the final centroid locations for each cluster.
|
126
|
+
* Data should be stored in column-major order.
|
127
|
+
* @param clusters Ignored in this method.
|
128
|
+
*
|
129
|
+
* @return `centers` is filled with the new cluster centers.
|
130
|
+
* The number of filled centers is returned, see `Initializer::run()`.
|
131
|
+
*/
|
132
|
+
CLUSTER_t run(int ndim, INDEX_t nobs, const DATA_t* data, CLUSTER_t ncenters, DATA_t* centers, CLUSTER_t* clusters) {
|
133
|
+
if (!nobs) {
|
134
|
+
return 0;
|
135
|
+
}
|
136
|
+
auto sofar = run(ndim, nobs, data, ncenters);
|
137
|
+
copy_into_array(sofar, ndim, data, centers);
|
138
|
+
return sofar.size();
|
139
|
+
}
|
140
|
+
};
|
141
|
+
|
142
|
+
}
|
143
|
+
|
144
|
+
#endif
|
@@ -0,0 +1,44 @@
|
|
1
|
+
#ifndef KMEANS_INITIALIZE_NONE_HPP
|
2
|
+
#define KMEANS_INITIALIZE_NONE_HPP
|
3
|
+
|
4
|
+
#include "Base.hpp"
|
5
|
+
#include <algorithm>
|
6
|
+
|
7
|
+
/**
|
8
|
+
* @file InitializeNone.hpp
|
9
|
+
*
|
10
|
+
* @brief Class for no initialization.
|
11
|
+
*/
|
12
|
+
|
13
|
+
namespace kmeans {
|
14
|
+
|
15
|
+
/**
|
16
|
+
* @brief Perform "initialization" by just using the input cluster centers.
|
17
|
+
*
|
18
|
+
* @tparam DATA_t Floating-point type for the data and centroids.
|
19
|
+
* @tparam CLUSTER_t Integer type for the cluster index.
|
20
|
+
* @tparam INDEX_t Integer type for the observation index.
|
21
|
+
*/
|
22
|
+
template<typename DATA_t = double, typename CLUSTER_t = int, typename INDEX_t = int>
|
23
|
+
class InitializeNone : public Initialize<DATA_t, CLUSTER_t, INDEX_t> {
|
24
|
+
public:
|
25
|
+
/*
|
26
|
+
* @param ndim Number of dimensions.
|
27
|
+
* @param nobs Number of observations.
|
28
|
+
* @param data Pointer to an array where the dimensions are rows and the observations are columns.
|
29
|
+
* Data should be stored in column-major format.
|
30
|
+
* @param ncenters Number of centers to pick.
|
31
|
+
* @param centers Pointer to a `ndim`-by-`ncenters` array where columns are cluster centers and rows are dimensions.
|
32
|
+
* This is left unchanged.
|
33
|
+
* @param clusters Ignored in this method.
|
34
|
+
*
|
35
|
+
* @return The smaller of `ncenters` and `nobs` is returned, see `Initialize::run()`.
|
36
|
+
*/
|
37
|
+
CLUSTER_t run(int ndim, INDEX_t nobs, const DATA_t* data, CLUSTER_t ncenters, DATA_t* centers, CLUSTER_t* clusters) {
|
38
|
+
return std::min(nobs, static_cast<INDEX_t>(ncenters));
|
39
|
+
}
|
40
|
+
};
|
41
|
+
|
42
|
+
}
|
43
|
+
|
44
|
+
#endif
|