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 +1,95 @@
|
|
1
|
+
#ifndef UMAPPP_COMBINE_NEIGHBOR_SETS_HPP
|
2
|
+
#define UMAPPP_COMBINE_NEIGHBOR_SETS_HPP
|
3
|
+
|
4
|
+
#include <vector>
|
5
|
+
#include <limits>
|
6
|
+
#include <cmath>
|
7
|
+
#include <algorithm>
|
8
|
+
#include <numeric>
|
9
|
+
|
10
|
+
#include "NeighborList.hpp"
|
11
|
+
|
12
|
+
namespace umappp {
|
13
|
+
|
14
|
+
template<typename Float>
|
15
|
+
void combine_neighbor_sets(NeighborList<Float>& x, Float mix_ratio = 1) {
|
16
|
+
std::vector<size_t> last(x.size());
|
17
|
+
std::vector<size_t> original(x.size());
|
18
|
+
|
19
|
+
for (size_t i = 0; i < x.size(); ++i) {
|
20
|
+
auto& current = x[i];
|
21
|
+
std::sort(current.begin(), current.end()); // sorting by ID, see below.
|
22
|
+
original[i] = x[i].size();
|
23
|
+
}
|
24
|
+
|
25
|
+
for (size_t first = 0; first < x.size(); ++first) {
|
26
|
+
auto& current = x[first];
|
27
|
+
const int desired = first;
|
28
|
+
|
29
|
+
// Looping through the neighbors and searching for self in each
|
30
|
+
// neighbor's neighbors. As each inner vector in 'x' is sorted,
|
31
|
+
// this should only require a single pass through the entire set of
|
32
|
+
// neighbors as we do not need to search previously searched hits.
|
33
|
+
for (auto& y : current) {
|
34
|
+
auto& target = x[y.first];
|
35
|
+
auto& curlast = last[y.first];
|
36
|
+
const auto& limits = original[y.first];
|
37
|
+
while (curlast < limits && target[curlast].first < desired) {
|
38
|
+
++curlast;
|
39
|
+
}
|
40
|
+
|
41
|
+
if (curlast < limits && target[curlast].first == desired) {
|
42
|
+
if (desired < y.first) { // don't average it twice.
|
43
|
+
Float product = y.second * target[curlast].second;
|
44
|
+
Float prob_final;
|
45
|
+
|
46
|
+
if (mix_ratio == 1) {
|
47
|
+
prob_final = y.second + target[curlast].second - product;
|
48
|
+
} else if (mix_ratio == 0) {
|
49
|
+
prob_final = product;
|
50
|
+
} else {
|
51
|
+
prob_final = mix_ratio * (y.second + target[curlast].second - product) + (1 - mix_ratio) * product;
|
52
|
+
}
|
53
|
+
|
54
|
+
y.second = prob_final;
|
55
|
+
target[curlast].second = prob_final;
|
56
|
+
}
|
57
|
+
} else {
|
58
|
+
if (mix_ratio == 1) {
|
59
|
+
target.emplace_back(desired, y.second);
|
60
|
+
} else if (mix_ratio == 0) {
|
61
|
+
y.second = 0; // mark for deletion.
|
62
|
+
} else {
|
63
|
+
y.second *= mix_ratio;
|
64
|
+
target.emplace_back(desired, y.second);
|
65
|
+
}
|
66
|
+
}
|
67
|
+
}
|
68
|
+
}
|
69
|
+
|
70
|
+
// Removing zero probabilities.
|
71
|
+
if (mix_ratio == 0) {
|
72
|
+
for (auto& current : x) {
|
73
|
+
typename std::remove_reference<decltype(current)>::type replacement;
|
74
|
+
replacement.reserve(current.size());
|
75
|
+
for (const auto& y : current) {
|
76
|
+
if (y.second) {
|
77
|
+
replacement.push_back(y);
|
78
|
+
}
|
79
|
+
}
|
80
|
+
std::swap(current, replacement);
|
81
|
+
}
|
82
|
+
}
|
83
|
+
|
84
|
+
// Sorting everything by index to be more cache-friendly. Also,
|
85
|
+
// irlba::ParallelSparseMatrix needs increasing inserts.
|
86
|
+
for (auto& current : x) {
|
87
|
+
std::sort(current.begin(), current.end());
|
88
|
+
}
|
89
|
+
|
90
|
+
return;
|
91
|
+
}
|
92
|
+
|
93
|
+
}
|
94
|
+
|
95
|
+
#endif
|
@@ -0,0 +1,157 @@
|
|
1
|
+
#ifndef UMAPPP_FIND_AB_HPP
|
2
|
+
#define UMAPPP_FIND_AB_HPP
|
3
|
+
|
4
|
+
#include <cmath>
|
5
|
+
#include <vector>
|
6
|
+
#include <iostream>
|
7
|
+
|
8
|
+
namespace umappp {
|
9
|
+
|
10
|
+
/*
|
11
|
+
* Derivatives were obtained using the following code in R:
|
12
|
+
*
|
13
|
+
* > delta <- expression((1/(1+ a * x^(2*b)) - y)^2)
|
14
|
+
* > D(delta, "a")
|
15
|
+
* > D(delta, "b")
|
16
|
+
* > D(D(delta, "a"), "a")
|
17
|
+
* > D(D(delta, "b"), "a")
|
18
|
+
* > D(D(delta, "b"), "b")
|
19
|
+
*
|
20
|
+
* To test, evaluate with:
|
21
|
+
*
|
22
|
+
* > a <- 1.55461
|
23
|
+
* > b <- 0.743147
|
24
|
+
* > s <- 1
|
25
|
+
* > m <- 0.05
|
26
|
+
* > x <- seq(0, 3 * s, length.out=301)[-1]
|
27
|
+
* > y <- pmin(1, exp(-(x - m)/s))
|
28
|
+
* > sum(eval(delta))
|
29
|
+
*/
|
30
|
+
|
31
|
+
template<typename Float>
|
32
|
+
std::pair<Float, Float> find_ab(Float spread, Float min_dist, Float grid = 300, Float limit = 0.5, int iter = 50, Float tol = 1e-6) {
|
33
|
+
Float x_half = std::log(limit) * -spread + min_dist;
|
34
|
+
Float d_half = limit / -spread;
|
35
|
+
|
36
|
+
// Compute the x and y coordinates of the expected distance curve.
|
37
|
+
std::vector<Float> grid_x(grid), grid_y(grid), log_x(grid);
|
38
|
+
const Float delta = spread * 3 / grid;
|
39
|
+
for (int g = 0; g < grid; ++g) {
|
40
|
+
grid_x[g] = (g + 1) * delta; // +1 to avoid meaningless least squares result at x = 0, where both curves have y = 1 (and also the derivative w.r.t. b is not defined).
|
41
|
+
log_x[g] = std::log(grid_x[g]);
|
42
|
+
grid_y[g] = (grid_x[g] <= min_dist ? 1 : std::exp(- (grid_x[g] - min_dist) / spread));
|
43
|
+
}
|
44
|
+
|
45
|
+
// Starting estimates.
|
46
|
+
Float b = - d_half * x_half / (1 / limit - 1) / (2 * limit * limit);
|
47
|
+
Float a = (1 / limit - 1) / std::pow(x_half, 2 * b);
|
48
|
+
|
49
|
+
std::vector<Float> observed_y(grid), xpow(grid);
|
50
|
+
auto compute_ss = [&](Float A, Float B) -> Float {
|
51
|
+
for (int g = 0; g < grid; ++g) {
|
52
|
+
xpow[g] = std::pow(grid_x[g], 2 * B);
|
53
|
+
observed_y[g] = 1 / (1 + A * xpow[g]);
|
54
|
+
}
|
55
|
+
|
56
|
+
Float ss = 0;
|
57
|
+
for (int g = 0; g < grid; ++g) {
|
58
|
+
ss += (grid_y[g] - observed_y[g]) * (grid_y[g] - observed_y[g]);
|
59
|
+
}
|
60
|
+
|
61
|
+
return ss;
|
62
|
+
};
|
63
|
+
Float ss = compute_ss(a, b);
|
64
|
+
|
65
|
+
for (int it = 0; it < iter; ++it) {
|
66
|
+
// Computing the first and second derivatives of the sum of squared differences.
|
67
|
+
Float da = 0, db = 0, daa = 0, dab = 0, dbb = 0;
|
68
|
+
for (int g = 0; g < grid; ++g) {
|
69
|
+
const Float& x = grid_x[g];
|
70
|
+
const Float& gy = grid_y[g];
|
71
|
+
const Float& oy = observed_y[g];
|
72
|
+
|
73
|
+
const Float& x2b = xpow[g];
|
74
|
+
const Float logx2 = log_x[g] * 2;
|
75
|
+
const Float delta = oy - gy;
|
76
|
+
|
77
|
+
// -(2 * (x^(2 * b)/(1 + a * x^(2 * b))^2 * (1/(1 + a * x^(2 * b)) - y)))
|
78
|
+
da += -2 * x2b * oy * oy * delta;
|
79
|
+
|
80
|
+
// -(2 * (a * (x^(2 * b) * (log(x) * 2))/(1 + a * x^(2 * b))^2 * (1/(1 + a * x^(2 * b)) - y)))
|
81
|
+
db += -2 * a * x2b * logx2 * oy * oy * delta;
|
82
|
+
|
83
|
+
// 2 * (
|
84
|
+
// x^(2 * b)/(1 + a * x^(2 * b))^2 * (x^(2 * b)/(1 + a * x^(2 * b))^2)
|
85
|
+
// + x^(2 * b) * (2 * (x^(2 * b) * (1 + a * x^(2 * b))))/((1 + a * x^(2 * b))^2)^2 * (1/(1 + a * x^(2 * b)) - y)
|
86
|
+
// )
|
87
|
+
daa += 2 * (
|
88
|
+
x2b * oy * oy * x2b * oy * oy
|
89
|
+
+ x2b * 2 * x2b * oy * oy * oy * delta
|
90
|
+
);
|
91
|
+
|
92
|
+
//-(2 *
|
93
|
+
// (
|
94
|
+
// (
|
95
|
+
// (x^(2 * b) * (log(x) * 2))/(1 + a * x^(2 * b))^2
|
96
|
+
// - a * (x^(2 * b) * (log(x) * 2)) * (2 * (x^(2 * b) * (1 + a * x^(2 * b))))/((1 + a * x^(2 * b))^2)^2
|
97
|
+
// )
|
98
|
+
// * (1/(1 + a * x^(2 * b)) - y)
|
99
|
+
// - a * (x^(2 * b) * (log(x) * 2))/(1 + a * x^(2 * b))^2 * (x^(2 * b)/(1 + a * x^(2 * b))^2)
|
100
|
+
// )
|
101
|
+
//)
|
102
|
+
dab += -2 * (
|
103
|
+
(
|
104
|
+
x2b * logx2 * oy * oy
|
105
|
+
- a * x2b * logx2 * 2 * x2b * oy * oy * oy
|
106
|
+
) * delta
|
107
|
+
- a * x2b * logx2 * oy * oy * x2b * oy * oy
|
108
|
+
);
|
109
|
+
|
110
|
+
// -(2 *
|
111
|
+
// (
|
112
|
+
// (
|
113
|
+
// a * (x^(2 * b) * (log(x) * 2) * (log(x) * 2))/(1 + a * x^(2 * b))^2
|
114
|
+
// - a * (x^(2 * b) * (log(x) * 2)) * (2 * (a * (x^(2 * b) * (log(x) * 2)) * (1 + a * x^(2 * b))))/((1 + a * x^(2 * b))^2)^2
|
115
|
+
// )
|
116
|
+
// * (1/(1 + a * x^(2 * b)) - y)
|
117
|
+
// - a * (x^(2 * b) * (log(x) * 2))/(1 + a * x^(2 * b))^2 * (a * (x^(2 * b) * (log(x) * 2))/(1 + a * x^(2 * b))^2)
|
118
|
+
// )
|
119
|
+
// )
|
120
|
+
dbb += -2 * (
|
121
|
+
(
|
122
|
+
(a * x2b * logx2 * logx2 * oy * oy)
|
123
|
+
- (a * x2b * logx2 * 2 * a * x2b * logx2 * oy * oy * oy)
|
124
|
+
) * delta
|
125
|
+
- a * x2b * logx2 * oy * oy * a * x2b * logx2 * oy * oy
|
126
|
+
);
|
127
|
+
}
|
128
|
+
|
129
|
+
// Applying the Newton iterations with damping.
|
130
|
+
Float determinant = daa * dbb - dab * dab;
|
131
|
+
const Float delta_a = (da * dbb - dab * db) / determinant;
|
132
|
+
const Float delta_b = (- da * dab + daa * db) / determinant;
|
133
|
+
|
134
|
+
Float ss_next = 0;
|
135
|
+
Float factor = 1;
|
136
|
+
for (int inner = 0; inner < 10; ++inner, factor /= 2) {
|
137
|
+
ss_next = compute_ss(a - factor * delta_a, b - factor * delta_b);
|
138
|
+
if (ss_next < ss) {
|
139
|
+
break;
|
140
|
+
}
|
141
|
+
}
|
142
|
+
|
143
|
+
if (ss && 1 - ss_next/ss > tol) {
|
144
|
+
a -= factor * delta_a;
|
145
|
+
b -= factor * delta_b;
|
146
|
+
ss = ss_next;
|
147
|
+
} else {
|
148
|
+
break;
|
149
|
+
}
|
150
|
+
}
|
151
|
+
|
152
|
+
return std::make_pair(a, b);
|
153
|
+
}
|
154
|
+
|
155
|
+
}
|
156
|
+
|
157
|
+
#endif
|
@@ -0,0 +1,136 @@
|
|
1
|
+
#ifndef UMAPPP_NEIGHBOR_SIMILARITIES_HPP
|
2
|
+
#define UMAPPP_NEIGHBOR_SIMILARITIES_HPP
|
3
|
+
|
4
|
+
#include <vector>
|
5
|
+
#include <limits>
|
6
|
+
#include <cmath>
|
7
|
+
#include <algorithm>
|
8
|
+
#include <numeric>
|
9
|
+
|
10
|
+
#include "NeighborList.hpp"
|
11
|
+
|
12
|
+
namespace umappp {
|
13
|
+
|
14
|
+
template<typename Float>
|
15
|
+
void neighbor_similarities(
|
16
|
+
NeighborList<Float>& x,
|
17
|
+
Float local_connectivity = 1.0,
|
18
|
+
Float bandwidth = 1.0,
|
19
|
+
int max_iter = 64,
|
20
|
+
Float tol = 1e-5,
|
21
|
+
Float min_k_dist_scale = 1e-3
|
22
|
+
) {
|
23
|
+
Float grand_mean_dist = -1;
|
24
|
+
constexpr Float max_val = std::numeric_limits<Float>::max();
|
25
|
+
|
26
|
+
#pragma omp parallel
|
27
|
+
{
|
28
|
+
std::vector<Float> non_zero_distances;
|
29
|
+
|
30
|
+
#pragma omp for
|
31
|
+
for (size_t i = 0; i < x.size(); ++i) {
|
32
|
+
auto& all_neighbors = x[i];
|
33
|
+
const int n_neighbors = all_neighbors.size();
|
34
|
+
|
35
|
+
non_zero_distances.clear();
|
36
|
+
for (const auto& f : all_neighbors) {
|
37
|
+
if (f.second) {
|
38
|
+
non_zero_distances.push_back(f.second);
|
39
|
+
}
|
40
|
+
}
|
41
|
+
|
42
|
+
if (non_zero_distances.size() <= local_connectivity) {
|
43
|
+
// When this happens, 'rho' is just theoretically set to the
|
44
|
+
// maximum distance. In such cases, the weights are always just
|
45
|
+
// set to 1 in the remaining code, because no distance can be
|
46
|
+
// greater than 'rho'. If that's the case, we might as well
|
47
|
+
// save some time and compute it here.
|
48
|
+
for (int k = 0; k < n_neighbors; ++k) {
|
49
|
+
all_neighbors[k].second = 1;
|
50
|
+
}
|
51
|
+
continue;
|
52
|
+
}
|
53
|
+
|
54
|
+
// Find rho, the distance to the nearest (non-identical) neighbor,
|
55
|
+
// possibly with interpolation.
|
56
|
+
int index = std::floor(local_connectivity);
|
57
|
+
const Float interpolation = local_connectivity - index;
|
58
|
+
const Float lower = (index > 0 ? non_zero_distances[index - 1] : 0); // 'index' is 1-based, so -1.
|
59
|
+
const Float upper = non_zero_distances[index];
|
60
|
+
const Float rho = lower + interpolation * (upper - lower);
|
61
|
+
|
62
|
+
// Iterating to find a good sigma, just like how t-SNE does so for beta.
|
63
|
+
Float sigma = 1.0;
|
64
|
+
Float lo = 0.0;
|
65
|
+
Float hi = max_val;
|
66
|
+
Float sigma_best = sigma;
|
67
|
+
Float adiff_min = max_val;
|
68
|
+
const Float target = std::log2(all_neighbors.size() + 1); // include self. Dunno why, but uwot does it.
|
69
|
+
|
70
|
+
bool converged = false;
|
71
|
+
for (int iter = 0; iter < max_iter; ++iter) {
|
72
|
+
// If distance = 0, then max(distance - rho, 0) = 0 as rho >=
|
73
|
+
// 0. In which case, exp(-dist / sigma) is just 1 for each
|
74
|
+
// distance of zero, allowing us to just add these directly.
|
75
|
+
Float val = n_neighbors - non_zero_distances.size();
|
76
|
+
|
77
|
+
for (auto d : non_zero_distances) {
|
78
|
+
if (d > rho) {
|
79
|
+
val += std::exp(-(d - rho)/ sigma);
|
80
|
+
} else {
|
81
|
+
val += 1;
|
82
|
+
}
|
83
|
+
}
|
84
|
+
|
85
|
+
Float adiff = std::abs(val - target);
|
86
|
+
if (adiff < tol) {
|
87
|
+
converged = true;
|
88
|
+
break;
|
89
|
+
}
|
90
|
+
|
91
|
+
// store best sigma in case binary search fails (usually in the presence
|
92
|
+
// of multiple degenerate distances)
|
93
|
+
if (adiff < adiff_min) {
|
94
|
+
adiff_min = adiff;
|
95
|
+
sigma_best = sigma;
|
96
|
+
}
|
97
|
+
|
98
|
+
if (val > target) {
|
99
|
+
hi = sigma;
|
100
|
+
sigma = (lo + hi) / 2;
|
101
|
+
} else {
|
102
|
+
lo = sigma;
|
103
|
+
if (hi == max_val) {
|
104
|
+
sigma *= 2;
|
105
|
+
} else {
|
106
|
+
sigma = (lo + hi) / 2;
|
107
|
+
}
|
108
|
+
}
|
109
|
+
}
|
110
|
+
|
111
|
+
if (!converged) {
|
112
|
+
sigma = sigma_best;
|
113
|
+
}
|
114
|
+
|
115
|
+
// Quickly summing over the non-zero distances, then dividing
|
116
|
+
// by the total number of neighbors to obtain the mean.
|
117
|
+
Float mean_dist = std::accumulate(non_zero_distances.begin(), non_zero_distances.end(), 0.0)/n_neighbors;
|
118
|
+
sigma = std::max(min_k_dist_scale * mean_dist, sigma);
|
119
|
+
|
120
|
+
for (int k = 0; k < n_neighbors; ++k) {
|
121
|
+
Float& dist = all_neighbors[k].second;
|
122
|
+
if (dist > rho) {
|
123
|
+
dist = std::exp(-(dist - rho) / (sigma * bandwidth));
|
124
|
+
} else {
|
125
|
+
dist = 1;
|
126
|
+
}
|
127
|
+
}
|
128
|
+
}
|
129
|
+
}
|
130
|
+
|
131
|
+
return;
|
132
|
+
}
|
133
|
+
|
134
|
+
}
|
135
|
+
|
136
|
+
#endif
|
@@ -0,0 +1,285 @@
|
|
1
|
+
#ifndef UMAPPP_OPTIMIZE_LAYOUT_HPP
|
2
|
+
#define UMAPPP_OPTIMIZE_LAYOUT_HPP
|
3
|
+
|
4
|
+
#include <vector>
|
5
|
+
#include <limits>
|
6
|
+
#include <algorithm>
|
7
|
+
#include <cmath>
|
8
|
+
|
9
|
+
#include "NeighborList.hpp"
|
10
|
+
#include "aarand/aarand.hpp"
|
11
|
+
|
12
|
+
namespace umappp {
|
13
|
+
|
14
|
+
template<typename Float>
|
15
|
+
struct EpochData {
|
16
|
+
EpochData(size_t nobs) : head(nobs) {}
|
17
|
+
|
18
|
+
int total_epochs;
|
19
|
+
int current_epoch = 0;
|
20
|
+
|
21
|
+
std::vector<size_t> head;
|
22
|
+
std::vector<int> tail;
|
23
|
+
std::vector<Float> epochs_per_sample;
|
24
|
+
|
25
|
+
std::vector<Float> epoch_of_next_sample;
|
26
|
+
std::vector<Float> epoch_of_next_negative_sample;
|
27
|
+
Float negative_sample_rate;
|
28
|
+
};
|
29
|
+
|
30
|
+
template<typename Float>
|
31
|
+
EpochData<Float> similarities_to_epochs(const NeighborList<Float>& p, int num_epochs, Float negative_sample_rate) {
|
32
|
+
Float maxed = 0;
|
33
|
+
size_t count = 0;
|
34
|
+
for (const auto& x : p) {
|
35
|
+
count += x.size();
|
36
|
+
for (const auto& y : x) {
|
37
|
+
maxed = std::max(maxed, y.second);
|
38
|
+
}
|
39
|
+
}
|
40
|
+
|
41
|
+
EpochData<Float> output(p.size());
|
42
|
+
output.total_epochs = num_epochs;
|
43
|
+
output.tail.reserve(count);
|
44
|
+
output.epochs_per_sample.reserve(count);
|
45
|
+
const Float limit = maxed / num_epochs;
|
46
|
+
|
47
|
+
size_t last = 0;
|
48
|
+
for (size_t i = 0; i < p.size(); ++i) {
|
49
|
+
const auto& x = p[i];
|
50
|
+
for (const auto& y : x) {
|
51
|
+
if (y.second >= limit) {
|
52
|
+
output.tail.push_back(y.first);
|
53
|
+
output.epochs_per_sample.push_back(maxed / y.second);
|
54
|
+
++last;
|
55
|
+
}
|
56
|
+
}
|
57
|
+
output.head[i] = last;
|
58
|
+
}
|
59
|
+
|
60
|
+
// Filling in some epoch-related running statistics.
|
61
|
+
output.epoch_of_next_sample = output.epochs_per_sample;
|
62
|
+
output.epoch_of_next_negative_sample = output.epochs_per_sample;
|
63
|
+
for (auto& e : output.epoch_of_next_negative_sample) {
|
64
|
+
e /= negative_sample_rate;
|
65
|
+
}
|
66
|
+
output.negative_sample_rate = negative_sample_rate;
|
67
|
+
|
68
|
+
return output;
|
69
|
+
}
|
70
|
+
|
71
|
+
template<typename Float>
|
72
|
+
Float quick_squared_distance(const Float* left, const Float* right, int ndim) {
|
73
|
+
Float dist2 = 0;
|
74
|
+
for (int d = 0; d < ndim; ++d, ++left, ++right) {
|
75
|
+
dist2 += (*left - *right) * (*left - *right);
|
76
|
+
}
|
77
|
+
constexpr Float dist_eps = std::numeric_limits<Float>::epsilon();
|
78
|
+
return std::max(dist_eps, dist2);
|
79
|
+
}
|
80
|
+
|
81
|
+
template<typename Float>
|
82
|
+
Float clamp(Float input) {
|
83
|
+
constexpr Float min_gradient = -4;
|
84
|
+
constexpr Float max_gradient = 4;
|
85
|
+
return std::min(std::max(input, min_gradient), max_gradient);
|
86
|
+
}
|
87
|
+
|
88
|
+
template<bool batch, typename Float, class Setup, class Rng>
|
89
|
+
void optimize_sample(
|
90
|
+
size_t i,
|
91
|
+
int ndim,
|
92
|
+
Float* embedding,
|
93
|
+
Float* buffer,
|
94
|
+
Setup& setup,
|
95
|
+
Float a,
|
96
|
+
Float b,
|
97
|
+
Float gamma,
|
98
|
+
Float alpha,
|
99
|
+
Rng& rng,
|
100
|
+
Float epoch
|
101
|
+
) {
|
102
|
+
const auto& head = setup.head;
|
103
|
+
const auto& tail = setup.tail;
|
104
|
+
const auto& epochs_per_sample = setup.epochs_per_sample;
|
105
|
+
auto& epoch_of_next_sample = setup.epoch_of_next_sample;
|
106
|
+
auto& epoch_of_next_negative_sample = setup.epoch_of_next_negative_sample;
|
107
|
+
|
108
|
+
const size_t num_obs = head.size();
|
109
|
+
const Float negative_sample_rate = setup.negative_sample_rate;
|
110
|
+
|
111
|
+
size_t start = (i == 0 ? 0 : setup.head[i-1]), end = setup.head[i];
|
112
|
+
Float* left = embedding + i * ndim;
|
113
|
+
|
114
|
+
for (size_t j = start; j < end; ++j) {
|
115
|
+
if (epoch_of_next_sample[j] > epoch) {
|
116
|
+
continue;
|
117
|
+
}
|
118
|
+
|
119
|
+
Float* right = embedding + tail[j] * ndim;
|
120
|
+
Float dist2 = quick_squared_distance(left, right, ndim);
|
121
|
+
const Float pd2b = std::pow(dist2, b);
|
122
|
+
const Float grad_coef = (-2 * a * b * pd2b) / (dist2 * (a * pd2b + 1.0));
|
123
|
+
{
|
124
|
+
Float* lcopy = left;
|
125
|
+
Float* rcopy = right;
|
126
|
+
|
127
|
+
for (int d = 0; d < ndim; ++d, ++lcopy, ++rcopy) {
|
128
|
+
Float gradient = alpha * clamp(grad_coef * (*lcopy - *rcopy));
|
129
|
+
if constexpr(!batch) {
|
130
|
+
*lcopy += gradient;
|
131
|
+
*rcopy -= gradient;
|
132
|
+
} else {
|
133
|
+
// Doubling as we'll assume symmetry from the same
|
134
|
+
// force applied by the right node. This allows us to
|
135
|
+
// avoid worrying about accounting for modifications to
|
136
|
+
// the right node.
|
137
|
+
buffer[d] += 2 * gradient;
|
138
|
+
}
|
139
|
+
}
|
140
|
+
}
|
141
|
+
|
142
|
+
// Here, we divide by epochs_per_negative_sample, defined as epochs_per_sample[j] / negative_sample_rate.
|
143
|
+
const size_t num_neg_samples = (epoch - epoch_of_next_negative_sample[j]) * negative_sample_rate / epochs_per_sample[j];
|
144
|
+
|
145
|
+
for (size_t p = 0; p < num_neg_samples; ++p) {
|
146
|
+
size_t sampled = aarand::discrete_uniform(rng, num_obs);
|
147
|
+
if (sampled == i) {
|
148
|
+
continue;
|
149
|
+
}
|
150
|
+
|
151
|
+
Float* right = embedding + sampled * ndim;
|
152
|
+
Float dist2 = quick_squared_distance(left, right, ndim);
|
153
|
+
const Float grad_coef = 2 * gamma * b / ((0.001 + dist2) * (a * std::pow(dist2, b) + 1.0));
|
154
|
+
{
|
155
|
+
Float* lcopy = left;
|
156
|
+
const Float* rcopy = right;
|
157
|
+
for (int d = 0; d < ndim; ++d, ++lcopy, ++rcopy) {
|
158
|
+
Float gradient = alpha * clamp(grad_coef * (*lcopy - *rcopy));
|
159
|
+
if constexpr(!batch) {
|
160
|
+
*lcopy += gradient;
|
161
|
+
} else {
|
162
|
+
buffer[d] += gradient;
|
163
|
+
}
|
164
|
+
}
|
165
|
+
}
|
166
|
+
}
|
167
|
+
|
168
|
+
epoch_of_next_sample[j] += epochs_per_sample[j];
|
169
|
+
|
170
|
+
// The update to epoch_of_next_negative_sample involves adding
|
171
|
+
// num_neg_samples * epochs_per_negative_sample, which eventually boils
|
172
|
+
// down to setting epoch_of_next_negative_sample to 'n'.
|
173
|
+
epoch_of_next_negative_sample[j] = epoch;
|
174
|
+
}
|
175
|
+
}
|
176
|
+
|
177
|
+
template<typename Float, class Setup, class Rng>
|
178
|
+
void optimize_layout(
|
179
|
+
int ndim,
|
180
|
+
Float* embedding,
|
181
|
+
Setup& setup,
|
182
|
+
Float a,
|
183
|
+
Float b,
|
184
|
+
Float gamma,
|
185
|
+
Float initial_alpha,
|
186
|
+
Rng& rng,
|
187
|
+
int epoch_limit
|
188
|
+
) {
|
189
|
+
auto& n = setup.current_epoch;
|
190
|
+
auto num_epochs = setup.total_epochs;
|
191
|
+
auto limit_epochs = num_epochs;
|
192
|
+
if (epoch_limit> 0) {
|
193
|
+
limit_epochs = std::min(epoch_limit, num_epochs);
|
194
|
+
}
|
195
|
+
|
196
|
+
for (; n < limit_epochs; ++n) {
|
197
|
+
const Float epoch = n;
|
198
|
+
const Float alpha = initial_alpha * (1.0 - epoch / num_epochs);
|
199
|
+
for (size_t i = 0; i < setup.head.size(); ++i) {
|
200
|
+
optimize_sample<false>(i, ndim, embedding, static_cast<Float*>(NULL), setup, a, b, gamma, alpha, rng, epoch);
|
201
|
+
}
|
202
|
+
}
|
203
|
+
|
204
|
+
return;
|
205
|
+
}
|
206
|
+
|
207
|
+
template<typename Float, class Setup, class SeedFunction, class EngineFunction>
|
208
|
+
inline void optimize_layout_batched(
|
209
|
+
int ndim,
|
210
|
+
Float* embedding,
|
211
|
+
Setup& setup,
|
212
|
+
Float a,
|
213
|
+
Float b,
|
214
|
+
Float gamma,
|
215
|
+
Float initial_alpha,
|
216
|
+
SeedFunction seeder,
|
217
|
+
EngineFunction creator,
|
218
|
+
int epoch_limit,
|
219
|
+
int nthreads
|
220
|
+
) {
|
221
|
+
auto& n = setup.current_epoch;
|
222
|
+
auto num_epochs = setup.total_epochs;
|
223
|
+
auto limit_epochs = num_epochs;
|
224
|
+
if (epoch_limit > 0) {
|
225
|
+
limit_epochs = std::min(epoch_limit, num_epochs);
|
226
|
+
}
|
227
|
+
|
228
|
+
const size_t num_obs = setup.head.size();
|
229
|
+
std::vector<decltype(seeder())> seeds(num_obs);
|
230
|
+
std::vector<Float> replace_buffer(num_obs * ndim);
|
231
|
+
Float* replacement = replace_buffer.data();
|
232
|
+
bool using_replacement = false;
|
233
|
+
|
234
|
+
for (; n < limit_epochs; ++n) {
|
235
|
+
const Float epoch = n;
|
236
|
+
const Float alpha = initial_alpha * (1.0 - epoch / num_epochs);
|
237
|
+
|
238
|
+
// Fill the seeds.
|
239
|
+
for (auto& s : seeds) {
|
240
|
+
s = seeder();
|
241
|
+
}
|
242
|
+
|
243
|
+
// Input and output alternate between epochs, to avoid the need for a
|
244
|
+
// copy operation on the entire embedding at the end of each epoch.
|
245
|
+
Float* reference = (using_replacement ? replacement : embedding);
|
246
|
+
Float* output = (using_replacement ? embedding : replacement);
|
247
|
+
using_replacement = !using_replacement;
|
248
|
+
|
249
|
+
#ifndef UMAPPP_CUSTOM_PARALLEL
|
250
|
+
#pragma omp parallel num_threads(nthreads)
|
251
|
+
{
|
252
|
+
std::vector<Float> buffer(ndim);
|
253
|
+
#pragma omp for
|
254
|
+
for (size_t i = 0; i < setup.head.size(); ++i) {
|
255
|
+
#else
|
256
|
+
UMAPPP_CUSTOM_PARALLEL(setup.head.size(), [&](size_t first, size_t last) -> void {
|
257
|
+
std::vector<Float> buffer(ndim);
|
258
|
+
for (size_t i = first; i < last; ++i) {
|
259
|
+
#endif
|
260
|
+
|
261
|
+
size_t shift = i * ndim;
|
262
|
+
std::copy(reference + shift, reference + shift + ndim, buffer.data());
|
263
|
+
auto rng = creator(seeds[i]);
|
264
|
+
optimize_sample<true>(i, ndim, reference, buffer.data(), setup, a, b, gamma, alpha, rng, epoch);
|
265
|
+
std::copy(buffer.begin(), buffer.end(), output + shift);
|
266
|
+
|
267
|
+
#ifndef UMAPPP_CUSTOM_PARALLEL
|
268
|
+
}
|
269
|
+
}
|
270
|
+
#else
|
271
|
+
}
|
272
|
+
}, nthreads);
|
273
|
+
#endif
|
274
|
+
}
|
275
|
+
|
276
|
+
if (using_replacement) {
|
277
|
+
std::copy(replace_buffer.begin(), replace_buffer.end(), embedding);
|
278
|
+
}
|
279
|
+
|
280
|
+
return;
|
281
|
+
}
|
282
|
+
|
283
|
+
}
|
284
|
+
|
285
|
+
#endif
|