Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
145K views

Jblas - Fast Matrix Computations For Java

This document describes jblas, a Java library for linear algebra that provides matrix classes and computational routines by wrapping BLAS and LAPACK and calling native code, allowing Java code to achieve near-native performance for operations like matrix multiplication. It discusses the main features and classes, how to create and access matrices, perform basic arithmetic and machine learning algorithms like kernel ridge regression, and outlines areas for future improvement.

Uploaded by

Mikio Braun
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
145K views

Jblas - Fast Matrix Computations For Java

This document describes jblas, a Java library for linear algebra that provides matrix classes and computational routines by wrapping BLAS and LAPACK and calling native code, allowing Java code to achieve near-native performance for operations like matrix multiplication. It discusses the main features and classes, how to create and access matrices, perform basic arithmetic and machine learning algorithms like kernel ridge regression, and outlines areas for future improvement.

Uploaded by

Mikio Braun
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 19

fast linear algebra for Java http://jblas.

org
Mikio L. Braun mikio@cs.tu-berlin.de TU Berlin, Franklinstr. 28/29, 10587 Berlin

June 23, 2010

Overview
Main features: Matrix library for Java based on native BLAS and LAPACK. Support for single and double precision oats, real and complex matrices. Vectors and two-dimensional dense matrices only. For large matrix practically native performance (for example, 10 GFLOPS on a 2GHz Intel Core2Duo processor for matrix-matrix multiplication). Precompiled fat jar for Linux (Intel), Mac OS, and Windows. Parses FORTRAN code to automatically generate JNI stubs.

Benchmarks

Structure

Main classes FloatMatrix, DoubleMatrix real matrices. ComplexFloatMatrix, ComplexDoubleMatrix complex matrices. Computational routines as static methods Eigen eigendecomposition Solve solving linear equations Singular singular value decomposition Decompose decompositions like LU, Cholesky, ... Geometry centering, normalizing, ...

Background on Native Calls in Java

Native calls are expensive: Arrays are always copied on native calls. For large arrays lots of cache misses. Doesnt make sense to call native code if operation is O(1)! What about direct buers? Direct buers arent garbage collected properly! Only makes sense to have a xed number of direct buers used for copying directly use JNI with arrays. Nevertheless, Java is as fast as C for simple things like vector addition!

Matrix creation

DoubleMatrix a = new DoubleMatrix(10, 5); // 10 * 5 matrix DoubleMatrix x = new DoubleMatrix(10); // vector of length 10 DoubleMatrix y = DoubleMatrix.zeros(10, 5); DoubleMatrix z = DoubleMatrix.ones(3, 4); DoubleMatrix g = DoubleMatrix.randn(3, 4); DoubleMatrix u = DoubleMatrix.rand(3); Dimensions: rows, columns, length Duplicating and copying: dup(), copy() Transposing: transpose()

Element access
DoubleMatrix a = new DoubleMatrix(10, 5); // accessing elements by row and column a.put(3, 2, 10.0); a.get(2, 3); // accessing element in linear order a.get(15); a.put(20, 1.0); // for example, for implementing elementwise operations: for (int i = 0; i < a.length; i++) a.put(i, a.get(i) * 3);

Accessing rows and columns: putRow(), getRow(), ... Pass a buer object to prevent spurious object creation: DoubleMatrix a = DoubleMatrix.randn(10, 100); DoubleMatrix buf = new DoubleMatrix(10); for (int i = 0; i < a.columns; i++) { a.getColumn(i, buf); // do something with buf }

Simple Arithmetic
DoubleMatrix a = new DoubleMatrix(3, 3, 1, 2, 3, 4, 5, 6, 7, 8, 9); DoubleMatrix x = new DoubleMatrix(3, 1, 10, 11, 12); DoubleMatrix y = a.mmul(x); DoubleMatrix z = x.add(y); Supported Operations: Basic arithmetics: add, sub, mul, mmul, div, dot Element-wise logical operations: and, or, xor, not, lt, le, gt, ge, eq, ne Rows and vectors to a matrix: addRowVector, addColumnVector, ...

Machine Learning: Kernel Ridge Regression with Gaussian Kernel


Lets implement the following: Noisy sinc data set learned with Kernel Ridge Regression with a Gaussian Kernel. Sinc Function sinc(x) = sin(x)/x 1 x =0 x = 0.

DoubleMatrix sinc(DoubleMatrix x) { return sin(x).div(x); } Not safe, what about x = 0? DoubleMatrix safeSinc(DoubleMatrix x) { return sin(x).div(x.add(x.eq(0))); }

Computing a sinc data set

X uniformly from -4, 4


2 Y = sinc(x) +

N (0, 1). DoubleMatrix[] sincDataset(int n, double noise) { DoubleMatrix X = rand(n).mul(8).sub(4); DoubleMatrix Y = sinc(X) .add( randn(n).mul(noise) ); return new DoubleMatrix[] {X, Y}; }

Gaussian kernel

k(x, z) = exp

x z w

DoubleMatrix gaussianKernel(double w, DoubleMatrix X, DoubleMatrix Z) { DoubleMatrix d = Geometry.pairwiseSquaredDistances(X.transpose(), Z.transpose()); return exp(d.div(w).neg()); }

Kernel Ridge Regression

KRR learns a normal kernel model


n

f (x) =
i=1

k(x, Xi )i

with = (K + I )1 Y , where K is the kernel matrix Kij = k(Xi , Xj ).

DoubleMatrix learnKRR(DoubleMatrix X, DoubleMatrix Y, double w, double lambda) { int n = X.rows; DoubleMatrix K = gaussianKernel(w, X, X); K.addi(eye(n).muli(lambda)); DoubleMatrix alpha = Solve.solveSymmetric(K, Y); return alpha; } DoubleMatrix predictKRR(DoubleMatrix XE, DoubleMatrix X, double w, DoubleMatrix alpha) { DoubleMatrix K = gaussianKernel(w, XE, X); return K.mmul(alpha); }

Computing the mean-squared error

1 n

(Yi Yi )2
i=1

double mse(DoubleMatrix Y1, DoubleMatrix Y2) { DoubleMatrix diff = Y1.sub(Y2); return pow(diff, 2).mean(); }

Conjugate Gradients
1. r b Ax 2. p r 3. repeat 4. 5. 6. 7. 8. 9. 10.
rT r p T Ap

x x + p r r Ap if r is suciently small, exit loop


r Tr rT r

p r + p r r .

11. end repeat

DoubleMatrix cg(DoubleMatrix A, DoubleMatrix b, DoubleMatrix x, double thresh) { int n = x.length; DoubleMatrix r = b.sub(A.mmul(x)); // 1 DoubleMatrix p = r.dup(); // 2 double alpha = 0, beta = 0; DoubleMatrix r2 = zeros(n), Ap = zeros(n); while (true) { // 3 A.mmuli(p, Ap); alpha = r.dot(r) / p.dot(Ap); // 4 x.addi(p.mul(alpha)); // 5 r.subi(Ap.mul(alpha), r2); // 6 double error = r2.norm2(); // 7 System.out.printf("Residual error = %f\n", error); if (error < thresh) break; beta = r2.dot(r2) / r.dot(r); // 8 r2.addi(p.mul(beta), p); // 9 DoubleMatrix temp = r; // 10 r = r2; r2 = temp; }

Outlook

Better integration of dierent matrix classes. Better performance through caching. Support for sparse matrices and more compact storage schemes.

You might also like