Repository: spark Updated Branches: refs/heads/master 8d72341ab -> 02b55de3d
[SPARK-4409][MLlib] Additional Linear Algebra Utils Addition of a very limited number of local matrix manipulation and generation methods that would be helpful in the further development for algorithms on top of BlockMatrix (SPARK-3974), such as Randomized SVD, and Multi Model Training (SPARK-1486). The proposed methods for addition are: For `Matrix` - map: maps the values in the matrix with a given function. Produces a new matrix. - update: the values in the matrix are updated with a given function. Occurs in place. Factory methods for `DenseMatrix`: - *zeros: Generate a matrix consisting of zeros - *ones: Generate a matrix consisting of ones - *eye: Generate an identity matrix - *rand: Generate a matrix consisting of i.i.d. uniform random numbers - *randn: Generate a matrix consisting of i.i.d. gaussian random numbers - *diag: Generate a diagonal matrix from a supplied vector *These methods already exist in the factory methods for `Matrices`, however for cases where we require a `DenseMatrix`, you constantly have to add `.asInstanceOf[DenseMatrix]` everywhere, which makes the code "dirtier". I propose moving these functions to factory methods for `DenseMatrix` where the putput will be a `DenseMatrix` and the factory methods for `Matrices` will call these functions directly and output a generic `Matrix`. Factory methods for `SparseMatrix`: - speye: Identity matrix in sparse format. Saves a ton of memory when dimensions are large, especially in Multi Model Training, where each row requires being multiplied by a scalar. - sprand: Generate a sparse matrix with a given density consisting of i.i.d. uniform random numbers. - sprandn: Generate a sparse matrix with a given density consisting of i.i.d. gaussian random numbers. - diag: Generate a diagonal matrix from a supplied vector, but is memory efficient, because it just stores the diagonal. Again, very helpful in Multi Model Training. Factory methods for `Matrices`: - Include all the factory methods given above, but return a generic `Matrix` rather than `SparseMatrix` or `DenseMatrix`. - horzCat: Horizontally concatenate matrices to form one larger matrix. Very useful in both Multi Model Training, and for the repartitioning of BlockMatrix. - vertCat: Vertically concatenate matrices to form one larger matrix. Very useful for the repartitioning of BlockMatrix. The names for these methods were selected from MATLAB Author: Burak Yavuz <brk...@gmail.com> Author: Xiangrui Meng <m...@databricks.com> Closes #3319 from brkyvz/SPARK-4409 and squashes the following commits: b0354f6 [Burak Yavuz] [SPARK-4409] Incorporated mengxr's code 04c4829 [Burak Yavuz] Merge pull request #1 from mengxr/SPARK-4409 80cfa29 [Xiangrui Meng] minor changes ecc937a [Xiangrui Meng] update sprand 4e95e24 [Xiangrui Meng] simplify fromCOO implementation 10a63a6 [Burak Yavuz] [SPARK-4409] Fourth pass of code review f62d6c7 [Burak Yavuz] [SPARK-4409] Modified genRandMatrix 3971c93 [Burak Yavuz] [SPARK-4409] Third pass of code review 75239f8 [Burak Yavuz] [SPARK-4409] Second pass of code review e4bd0c0 [Burak Yavuz] [SPARK-4409] Modified horzcat and vertcat 65c562e [Burak Yavuz] [SPARK-4409] Hopefully fixed Java Test d8be7bc [Burak Yavuz] [SPARK-4409] Organized imports 065b531 [Burak Yavuz] [SPARK-4409] First pass after code review a8120d2 [Burak Yavuz] [SPARK-4409] Finished updates to API according to SPARK-4614 f798c82 [Burak Yavuz] [SPARK-4409] Updated API according to SPARK-4614 c75f3cd [Burak Yavuz] [SPARK-4409] Added JavaAPI Tests, and fixed a couple of bugs d662f9d [Burak Yavuz] [SPARK-4409] Modified according to remote repo 83dfe37 [Burak Yavuz] [SPARK-4409] Scalastyle error fixed a14c0da [Burak Yavuz] [SPARK-4409] Initial commit to add methods Project: http://git-wip-us.apache.org/repos/asf/spark/repo Commit: http://git-wip-us.apache.org/repos/asf/spark/commit/02b55de3 Tree: http://git-wip-us.apache.org/repos/asf/spark/tree/02b55de3 Diff: http://git-wip-us.apache.org/repos/asf/spark/diff/02b55de3 Branch: refs/heads/master Commit: 02b55de3dce9a1fef806be13e5cefa0f39ea2fcc Parents: 8d72341 Author: Burak Yavuz <brk...@gmail.com> Authored: Mon Dec 29 13:24:26 2014 -0800 Committer: Xiangrui Meng <m...@databricks.com> Committed: Mon Dec 29 13:24:26 2014 -0800 ---------------------------------------------------------------------- .../apache/spark/mllib/linalg/Matrices.scala | 570 +++++++++++++++++-- .../spark/mllib/linalg/JavaMatricesSuite.java | 163 ++++++ .../spark/mllib/linalg/MatricesSuite.scala | 172 +++++- .../apache/spark/mllib/util/TestingUtils.scala | 6 +- 4 files changed, 868 insertions(+), 43 deletions(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/spark/blob/02b55de3/mllib/src/main/scala/org/apache/spark/mllib/linalg/Matrices.scala ---------------------------------------------------------------------- diff --git a/mllib/src/main/scala/org/apache/spark/mllib/linalg/Matrices.scala b/mllib/src/main/scala/org/apache/spark/mllib/linalg/Matrices.scala index 327366a..5a7281e 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/linalg/Matrices.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/linalg/Matrices.scala @@ -17,9 +17,11 @@ package org.apache.spark.mllib.linalg -import java.util.{Random, Arrays} +import java.util.{Arrays, Random} -import breeze.linalg.{Matrix => BM, DenseMatrix => BDM, CSCMatrix => BSM} +import scala.collection.mutable.{ArrayBuilder => MArrayBuilder, HashSet => MHashSet, ArrayBuffer} + +import breeze.linalg.{CSCMatrix => BSM, DenseMatrix => BDM, Matrix => BM} /** * Trait for a local matrix. @@ -80,6 +82,16 @@ sealed trait Matrix extends Serializable { /** A human readable representation of the matrix */ override def toString: String = toBreeze.toString() + + /** Map the values of this matrix using a function. Generates a new matrix. Performs the + * function on only the backing array. For example, an operation such as addition or + * subtraction will only be performed on the non-zero values in a `SparseMatrix`. */ + private[mllib] def map(f: Double => Double): Matrix + + /** Update all the values of this matrix using the function f. Performed in-place on the + * backing array. For example, an operation such as addition or subtraction will only be + * performed on the non-zero values in a `SparseMatrix`. */ + private[mllib] def update(f: Double => Double): Matrix } /** @@ -123,6 +135,122 @@ class DenseMatrix(val numRows: Int, val numCols: Int, val values: Array[Double]) } override def copy = new DenseMatrix(numRows, numCols, values.clone()) + + private[mllib] def map(f: Double => Double) = new DenseMatrix(numRows, numCols, values.map(f)) + + private[mllib] def update(f: Double => Double): DenseMatrix = { + val len = values.length + var i = 0 + while (i < len) { + values(i) = f(values(i)) + i += 1 + } + this + } + + /** Generate a `SparseMatrix` from the given `DenseMatrix`. */ + def toSparse(): SparseMatrix = { + val spVals: MArrayBuilder[Double] = new MArrayBuilder.ofDouble + val colPtrs: Array[Int] = new Array[Int](numCols + 1) + val rowIndices: MArrayBuilder[Int] = new MArrayBuilder.ofInt + var nnz = 0 + var j = 0 + while (j < numCols) { + var i = 0 + val indStart = j * numRows + while (i < numRows) { + val v = values(indStart + i) + if (v != 0.0) { + rowIndices += i + spVals += v + nnz += 1 + } + i += 1 + } + j += 1 + colPtrs(j) = nnz + } + new SparseMatrix(numRows, numCols, colPtrs, rowIndices.result(), spVals.result()) + } +} + +/** + * Factory methods for [[org.apache.spark.mllib.linalg.DenseMatrix]]. + */ +object DenseMatrix { + + /** + * Generate a `DenseMatrix` consisting of zeros. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @return `DenseMatrix` with size `numRows` x `numCols` and values of zeros + */ + def zeros(numRows: Int, numCols: Int): DenseMatrix = + new DenseMatrix(numRows, numCols, new Array[Double](numRows * numCols)) + + /** + * Generate a `DenseMatrix` consisting of ones. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @return `DenseMatrix` with size `numRows` x `numCols` and values of ones + */ + def ones(numRows: Int, numCols: Int): DenseMatrix = + new DenseMatrix(numRows, numCols, Array.fill(numRows * numCols)(1.0)) + + /** + * Generate an Identity Matrix in `DenseMatrix` format. + * @param n number of rows and columns of the matrix + * @return `DenseMatrix` with size `n` x `n` and values of ones on the diagonal + */ + def eye(n: Int): DenseMatrix = { + val identity = DenseMatrix.zeros(n, n) + var i = 0 + while (i < n) { + identity.update(i, i, 1.0) + i += 1 + } + identity + } + + /** + * Generate a `DenseMatrix` consisting of i.i.d. uniform random numbers. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @param rng a random number generator + * @return `DenseMatrix` with size `numRows` x `numCols` and values in U(0, 1) + */ + def rand(numRows: Int, numCols: Int, rng: Random): DenseMatrix = { + new DenseMatrix(numRows, numCols, Array.fill(numRows * numCols)(rng.nextDouble())) + } + + /** + * Generate a `DenseMatrix` consisting of i.i.d. gaussian random numbers. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @param rng a random number generator + * @return `DenseMatrix` with size `numRows` x `numCols` and values in N(0, 1) + */ + def randn(numRows: Int, numCols: Int, rng: Random): DenseMatrix = { + new DenseMatrix(numRows, numCols, Array.fill(numRows * numCols)(rng.nextGaussian())) + } + + /** + * Generate a diagonal matrix in `DenseMatrix` format from the supplied values. + * @param vector a `Vector` that will form the values on the diagonal of the matrix + * @return Square `DenseMatrix` with size `values.length` x `values.length` and `values` + * on the diagonal + */ + def diag(vector: Vector): DenseMatrix = { + val n = vector.size + val matrix = DenseMatrix.zeros(n, n) + val values = vector.toArray + var i = 0 + while (i < n) { + matrix.update(i, i, values(i)) + i += 1 + } + matrix + } } /** @@ -156,6 +284,8 @@ class SparseMatrix( require(colPtrs.length == numCols + 1, "The length of the column indices should be the " + s"number of columns + 1. Currently, colPointers.length: ${colPtrs.length}, " + s"numCols: $numCols") + require(values.length == colPtrs.last, "The last value of colPtrs must equal the number of " + + s"elements. values.length: ${values.length}, colPtrs.last: ${colPtrs.last}") override def toArray: Array[Double] = { val arr = new Array[Double](numRows * numCols) @@ -188,7 +318,7 @@ class SparseMatrix( private[mllib] def update(i: Int, j: Int, v: Double): Unit = { val ind = index(i, j) - if (ind == -1){ + if (ind == -1) { throw new NoSuchElementException("The given row and column indices correspond to a zero " + "value. Only non-zero elements in Sparse Matrices can be updated.") } else { @@ -197,6 +327,192 @@ class SparseMatrix( } override def copy = new SparseMatrix(numRows, numCols, colPtrs, rowIndices, values.clone()) + + private[mllib] def map(f: Double => Double) = + new SparseMatrix(numRows, numCols, colPtrs, rowIndices, values.map(f)) + + private[mllib] def update(f: Double => Double): SparseMatrix = { + val len = values.length + var i = 0 + while (i < len) { + values(i) = f(values(i)) + i += 1 + } + this + } + + /** Generate a `DenseMatrix` from the given `SparseMatrix`. */ + def toDense(): DenseMatrix = { + new DenseMatrix(numRows, numCols, toArray) + } +} + +/** + * Factory methods for [[org.apache.spark.mllib.linalg.SparseMatrix]]. + */ +object SparseMatrix { + + /** + * Generate a `SparseMatrix` from Coordinate List (COO) format. Input must be an array of + * (i, j, value) tuples. Entries that have duplicate values of i and j are + * added together. Tuples where value is equal to zero will be omitted. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @param entries Array of (i, j, value) tuples + * @return The corresponding `SparseMatrix` + */ + def fromCOO(numRows: Int, numCols: Int, entries: Iterable[(Int, Int, Double)]): SparseMatrix = { + val sortedEntries = entries.toSeq.sortBy(v => (v._2, v._1)) + val numEntries = sortedEntries.size + if (sortedEntries.nonEmpty) { + // Since the entries are sorted by column index, we only need to check the first and the last. + for (col <- Seq(sortedEntries.head._2, sortedEntries.last._2)) { + require(col >= 0 && col < numCols, s"Column index out of range [0, $numCols): $col.") + } + } + val colPtrs = new Array[Int](numCols + 1) + val rowIndices = MArrayBuilder.make[Int] + rowIndices.sizeHint(numEntries) + val values = MArrayBuilder.make[Double] + values.sizeHint(numEntries) + var nnz = 0 + var prevCol = 0 + var prevRow = -1 + var prevVal = 0.0 + // Append a dummy entry to include the last one at the end of the loop. + (sortedEntries.view :+ (numRows, numCols, 1.0)).foreach { case (i, j, v) => + if (v != 0) { + if (i == prevRow && j == prevCol) { + prevVal += v + } else { + if (prevVal != 0) { + require(prevRow >= 0 && prevRow < numRows, + s"Row index out of range [0, $numRows): $prevRow.") + nnz += 1 + rowIndices += prevRow + values += prevVal + } + prevRow = i + prevVal = v + while (prevCol < j) { + colPtrs(prevCol + 1) = nnz + prevCol += 1 + } + } + } + } + new SparseMatrix(numRows, numCols, colPtrs, rowIndices.result(), values.result()) + } + + /** + * Generate an Identity Matrix in `SparseMatrix` format. + * @param n number of rows and columns of the matrix + * @return `SparseMatrix` with size `n` x `n` and values of ones on the diagonal + */ + def speye(n: Int): SparseMatrix = { + new SparseMatrix(n, n, (0 to n).toArray, (0 until n).toArray, Array.fill(n)(1.0)) + } + + /** + * Generates the skeleton of a random `SparseMatrix` with a given random number generator. + * The values of the matrix returned are undefined. + */ + private def genRandMatrix( + numRows: Int, + numCols: Int, + density: Double, + rng: Random): SparseMatrix = { + require(numRows > 0, s"numRows must be greater than 0 but got $numRows") + require(numCols > 0, s"numCols must be greater than 0 but got $numCols") + require(density >= 0.0 && density <= 1.0, + s"density must be a double in the range 0.0 <= d <= 1.0. Currently, density: $density") + val size = numRows.toLong * numCols + val expected = size * density + assert(expected < Int.MaxValue, + "The expected number of nonzeros cannot be greater than Int.MaxValue.") + val nnz = math.ceil(expected).toInt + if (density == 0.0) { + new SparseMatrix(numRows, numCols, new Array[Int](numCols + 1), Array[Int](), Array[Double]()) + } else if (density == 1.0) { + val colPtrs = Array.tabulate(numCols + 1)(j => j * numRows) + val rowIndices = Array.tabulate(size.toInt)(idx => idx % numRows) + new SparseMatrix(numRows, numCols, colPtrs, rowIndices, new Array[Double](numRows * numCols)) + } else if (density < 0.34) { + // draw-by-draw, expected number of iterations is less than 1.5 * nnz + val entries = MHashSet[(Int, Int)]() + while (entries.size < nnz) { + entries += ((rng.nextInt(numRows), rng.nextInt(numCols))) + } + SparseMatrix.fromCOO(numRows, numCols, entries.map(v => (v._1, v._2, 1.0))) + } else { + // selection-rejection method + var idx = 0L + var numSelected = 0 + var j = 0 + val colPtrs = new Array[Int](numCols + 1) + val rowIndices = new Array[Int](nnz) + while (j < numCols && numSelected < nnz) { + var i = 0 + while (i < numRows && numSelected < nnz) { + if (rng.nextDouble() < 1.0 * (nnz - numSelected) / (size - idx)) { + rowIndices(numSelected) = i + numSelected += 1 + } + i += 1 + idx += 1 + } + colPtrs(j + 1) = numSelected + j += 1 + } + new SparseMatrix(numRows, numCols, colPtrs, rowIndices, new Array[Double](nnz)) + } + } + + /** + * Generate a `SparseMatrix` consisting of i.i.d. uniform random numbers. The number of non-zero + * elements equal the ceiling of `numRows` x `numCols` x `density` + * + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @param density the desired density for the matrix + * @param rng a random number generator + * @return `SparseMatrix` with size `numRows` x `numCols` and values in U(0, 1) + */ + def sprand(numRows: Int, numCols: Int, density: Double, rng: Random): SparseMatrix = { + val mat = genRandMatrix(numRows, numCols, density, rng) + mat.update(i => rng.nextDouble()) + } + + /** + * Generate a `SparseMatrix` consisting of i.i.d. gaussian random numbers. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @param density the desired density for the matrix + * @param rng a random number generator + * @return `SparseMatrix` with size `numRows` x `numCols` and values in N(0, 1) + */ + def sprandn(numRows: Int, numCols: Int, density: Double, rng: Random): SparseMatrix = { + val mat = genRandMatrix(numRows, numCols, density, rng) + mat.update(i => rng.nextGaussian()) + } + + /** + * Generate a diagonal matrix in `SparseMatrix` format from the supplied values. + * @param vector a `Vector` that will form the values on the diagonal of the matrix + * @return Square `SparseMatrix` with size `values.length` x `values.length` and non-zero + * `values` on the diagonal + */ + def diag(vector: Vector): SparseMatrix = { + val n = vector.size + vector match { + case sVec: SparseVector => + SparseMatrix.fromCOO(n, n, sVec.indices.zip(sVec.values).map(v => (v._1, v._1, v._2))) + case dVec: DenseVector => + val entries = dVec.values.zipWithIndex + val nnzVals = entries.filter(v => v._1 != 0.0) + SparseMatrix.fromCOO(n, n, nnzVals.map(v => (v._2, v._2, v._1))) + } + } } /** @@ -256,72 +572,250 @@ object Matrices { * Generate a `DenseMatrix` consisting of zeros. * @param numRows number of rows of the matrix * @param numCols number of columns of the matrix - * @return `DenseMatrix` with size `numRows` x `numCols` and values of zeros + * @return `Matrix` with size `numRows` x `numCols` and values of zeros */ - def zeros(numRows: Int, numCols: Int): Matrix = - new DenseMatrix(numRows, numCols, new Array[Double](numRows * numCols)) + def zeros(numRows: Int, numCols: Int): Matrix = DenseMatrix.zeros(numRows, numCols) /** * Generate a `DenseMatrix` consisting of ones. * @param numRows number of rows of the matrix * @param numCols number of columns of the matrix - * @return `DenseMatrix` with size `numRows` x `numCols` and values of ones + * @return `Matrix` with size `numRows` x `numCols` and values of ones */ - def ones(numRows: Int, numCols: Int): Matrix = - new DenseMatrix(numRows, numCols, Array.fill(numRows * numCols)(1.0)) + def ones(numRows: Int, numCols: Int): Matrix = DenseMatrix.ones(numRows, numCols) /** - * Generate an Identity Matrix in `DenseMatrix` format. + * Generate a dense Identity Matrix in `Matrix` format. * @param n number of rows and columns of the matrix - * @return `DenseMatrix` with size `n` x `n` and values of ones on the diagonal + * @return `Matrix` with size `n` x `n` and values of ones on the diagonal */ - def eye(n: Int): Matrix = { - val identity = Matrices.zeros(n, n) - var i = 0 - while (i < n){ - identity.update(i, i, 1.0) - i += 1 - } - identity - } + def eye(n: Int): Matrix = DenseMatrix.eye(n) + + /** + * Generate a sparse Identity Matrix in `Matrix` format. + * @param n number of rows and columns of the matrix + * @return `Matrix` with size `n` x `n` and values of ones on the diagonal + */ + def speye(n: Int): Matrix = SparseMatrix.speye(n) /** * Generate a `DenseMatrix` consisting of i.i.d. uniform random numbers. * @param numRows number of rows of the matrix * @param numCols number of columns of the matrix * @param rng a random number generator - * @return `DenseMatrix` with size `numRows` x `numCols` and values in U(0, 1) + * @return `Matrix` with size `numRows` x `numCols` and values in U(0, 1) */ - def rand(numRows: Int, numCols: Int, rng: Random): Matrix = { - new DenseMatrix(numRows, numCols, Array.fill(numRows * numCols)(rng.nextDouble())) - } + def rand(numRows: Int, numCols: Int, rng: Random): Matrix = + DenseMatrix.rand(numRows, numCols, rng) + + /** + * Generate a `SparseMatrix` consisting of i.i.d. gaussian random numbers. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @param density the desired density for the matrix + * @param rng a random number generator + * @return `Matrix` with size `numRows` x `numCols` and values in U(0, 1) + */ + def sprand(numRows: Int, numCols: Int, density: Double, rng: Random): Matrix = + SparseMatrix.sprand(numRows, numCols, density, rng) /** * Generate a `DenseMatrix` consisting of i.i.d. gaussian random numbers. * @param numRows number of rows of the matrix * @param numCols number of columns of the matrix * @param rng a random number generator - * @return `DenseMatrix` with size `numRows` x `numCols` and values in N(0, 1) + * @return `Matrix` with size `numRows` x `numCols` and values in N(0, 1) */ - def randn(numRows: Int, numCols: Int, rng: Random): Matrix = { - new DenseMatrix(numRows, numCols, Array.fill(numRows * numCols)(rng.nextGaussian())) - } + def randn(numRows: Int, numCols: Int, rng: Random): Matrix = + DenseMatrix.randn(numRows, numCols, rng) + + /** + * Generate a `SparseMatrix` consisting of i.i.d. gaussian random numbers. + * @param numRows number of rows of the matrix + * @param numCols number of columns of the matrix + * @param density the desired density for the matrix + * @param rng a random number generator + * @return `Matrix` with size `numRows` x `numCols` and values in N(0, 1) + */ + def sprandn(numRows: Int, numCols: Int, density: Double, rng: Random): Matrix = + SparseMatrix.sprandn(numRows, numCols, density, rng) /** * Generate a diagonal matrix in `DenseMatrix` format from the supplied values. * @param vector a `Vector` tat will form the values on the diagonal of the matrix - * @return Square `DenseMatrix` with size `values.length` x `values.length` and `values` + * @return Square `Matrix` with size `values.length` x `values.length` and `values` * on the diagonal */ - def diag(vector: Vector): Matrix = { - val n = vector.size - val matrix = Matrices.eye(n) - val values = vector.toArray - var i = 0 - while (i < n) { - matrix.update(i, i, values(i)) - i += 1 + def diag(vector: Vector): Matrix = DenseMatrix.diag(vector) + + /** + * Horizontally concatenate a sequence of matrices. The returned matrix will be in the format + * the matrices are supplied in. Supplying a mix of dense and sparse matrices will result in + * a sparse matrix. If the Array is empty, an empty `DenseMatrix` will be returned. + * @param matrices array of matrices + * @return a single `Matrix` composed of the matrices that were horizontally concatenated + */ + def horzcat(matrices: Array[Matrix]): Matrix = { + if (matrices.isEmpty) { + return new DenseMatrix(0, 0, Array[Double]()) + } else if (matrices.size == 1) { + return matrices(0) + } + val numRows = matrices(0).numRows + var hasSparse = false + var numCols = 0 + matrices.foreach { mat => + require(numRows == mat.numRows, "The number of rows of the matrices in this sequence, " + + "don't match!") + mat match { + case sparse: SparseMatrix => hasSparse = true + case dense: DenseMatrix => // empty on purpose + case _ => throw new IllegalArgumentException("Unsupported matrix format. Expected " + + s"SparseMatrix or DenseMatrix. Instead got: ${mat.getClass}") + } + numCols += mat.numCols + } + if (!hasSparse) { + new DenseMatrix(numRows, numCols, matrices.flatMap(_.toArray)) + } else { + var startCol = 0 + val entries: Array[(Int, Int, Double)] = matrices.flatMap { + case spMat: SparseMatrix => + var j = 0 + val colPtrs = spMat.colPtrs + val rowIndices = spMat.rowIndices + val values = spMat.values + val data = new Array[(Int, Int, Double)](values.length) + val nCols = spMat.numCols + while (j < nCols) { + var idx = colPtrs(j) + while (idx < colPtrs(j + 1)) { + val i = rowIndices(idx) + val v = values(idx) + data(idx) = (i, j + startCol, v) + idx += 1 + } + j += 1 + } + startCol += nCols + data + case dnMat: DenseMatrix => + val data = new ArrayBuffer[(Int, Int, Double)]() + var j = 0 + val nCols = dnMat.numCols + val nRows = dnMat.numRows + val values = dnMat.values + while (j < nCols) { + var i = 0 + val indStart = j * nRows + while (i < nRows) { + val v = values(indStart + i) + if (v != 0.0) { + data.append((i, j + startCol, v)) + } + i += 1 + } + j += 1 + } + startCol += nCols + data + } + SparseMatrix.fromCOO(numRows, numCols, entries) + } + } + + /** + * Vertically concatenate a sequence of matrices. The returned matrix will be in the format + * the matrices are supplied in. Supplying a mix of dense and sparse matrices will result in + * a sparse matrix. If the Array is empty, an empty `DenseMatrix` will be returned. + * @param matrices array of matrices + * @return a single `Matrix` composed of the matrices that were vertically concatenated + */ + def vertcat(matrices: Array[Matrix]): Matrix = { + if (matrices.isEmpty) { + return new DenseMatrix(0, 0, Array[Double]()) + } else if (matrices.size == 1) { + return matrices(0) + } + val numCols = matrices(0).numCols + var hasSparse = false + var numRows = 0 + matrices.foreach { mat => + require(numCols == mat.numCols, "The number of rows of the matrices in this sequence, " + + "don't match!") + mat match { + case sparse: SparseMatrix => + hasSparse = true + case dense: DenseMatrix => + case _ => throw new IllegalArgumentException("Unsupported matrix format. Expected " + + s"SparseMatrix or DenseMatrix. Instead got: ${mat.getClass}") + } + numRows += mat.numRows + + } + if (!hasSparse) { + val allValues = new Array[Double](numRows * numCols) + var startRow = 0 + matrices.foreach { mat => + var j = 0 + val nRows = mat.numRows + val values = mat.toArray + while (j < numCols) { + var i = 0 + val indStart = j * numRows + startRow + val subMatStart = j * nRows + while (i < nRows) { + allValues(indStart + i) = values(subMatStart + i) + i += 1 + } + j += 1 + } + startRow += nRows + } + new DenseMatrix(numRows, numCols, allValues) + } else { + var startRow = 0 + val entries: Array[(Int, Int, Double)] = matrices.flatMap { + case spMat: SparseMatrix => + var j = 0 + val colPtrs = spMat.colPtrs + val rowIndices = spMat.rowIndices + val values = spMat.values + val data = new Array[(Int, Int, Double)](values.length) + while (j < numCols) { + var idx = colPtrs(j) + while (idx < colPtrs(j + 1)) { + val i = rowIndices(idx) + val v = values(idx) + data(idx) = (i + startRow, j, v) + idx += 1 + } + j += 1 + } + startRow += spMat.numRows + data + case dnMat: DenseMatrix => + val data = new ArrayBuffer[(Int, Int, Double)]() + var j = 0 + val nCols = dnMat.numCols + val nRows = dnMat.numRows + val values = dnMat.values + while (j < nCols) { + var i = 0 + val indStart = j * nRows + while (i < nRows) { + val v = values(indStart + i) + if (v != 0.0) { + data.append((i + startRow, j, v)) + } + i += 1 + } + j += 1 + } + startRow += nRows + data + } + SparseMatrix.fromCOO(numRows, numCols, entries) } - matrix } } http://git-wip-us.apache.org/repos/asf/spark/blob/02b55de3/mllib/src/test/java/org/apache/spark/mllib/linalg/JavaMatricesSuite.java ---------------------------------------------------------------------- diff --git a/mllib/src/test/java/org/apache/spark/mllib/linalg/JavaMatricesSuite.java b/mllib/src/test/java/org/apache/spark/mllib/linalg/JavaMatricesSuite.java new file mode 100644 index 0000000..704d484 --- /dev/null +++ b/mllib/src/test/java/org/apache/spark/mllib/linalg/JavaMatricesSuite.java @@ -0,0 +1,163 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.mllib.linalg; + +import static org.junit.Assert.*; +import org.junit.Test; + +import java.io.Serializable; +import java.util.Random; + +public class JavaMatricesSuite implements Serializable { + + @Test + public void randMatrixConstruction() { + Random rng = new Random(24); + Matrix r = Matrices.rand(3, 4, rng); + rng.setSeed(24); + DenseMatrix dr = DenseMatrix.rand(3, 4, rng); + assertArrayEquals(r.toArray(), dr.toArray(), 0.0); + + rng.setSeed(24); + Matrix rn = Matrices.randn(3, 4, rng); + rng.setSeed(24); + DenseMatrix drn = DenseMatrix.randn(3, 4, rng); + assertArrayEquals(rn.toArray(), drn.toArray(), 0.0); + + rng.setSeed(24); + Matrix s = Matrices.sprand(3, 4, 0.5, rng); + rng.setSeed(24); + SparseMatrix sr = SparseMatrix.sprand(3, 4, 0.5, rng); + assertArrayEquals(s.toArray(), sr.toArray(), 0.0); + + rng.setSeed(24); + Matrix sn = Matrices.sprandn(3, 4, 0.5, rng); + rng.setSeed(24); + SparseMatrix srn = SparseMatrix.sprandn(3, 4, 0.5, rng); + assertArrayEquals(sn.toArray(), srn.toArray(), 0.0); + } + + @Test + public void identityMatrixConstruction() { + Matrix r = Matrices.eye(2); + DenseMatrix dr = DenseMatrix.eye(2); + SparseMatrix sr = SparseMatrix.speye(2); + assertArrayEquals(r.toArray(), dr.toArray(), 0.0); + assertArrayEquals(sr.toArray(), dr.toArray(), 0.0); + assertArrayEquals(r.toArray(), new double[]{1.0, 0.0, 0.0, 1.0}, 0.0); + } + + @Test + public void diagonalMatrixConstruction() { + Vector v = Vectors.dense(1.0, 0.0, 2.0); + Vector sv = Vectors.sparse(3, new int[]{0, 2}, new double[]{1.0, 2.0}); + + Matrix m = Matrices.diag(v); + Matrix sm = Matrices.diag(sv); + DenseMatrix d = DenseMatrix.diag(v); + DenseMatrix sd = DenseMatrix.diag(sv); + SparseMatrix s = SparseMatrix.diag(v); + SparseMatrix ss = SparseMatrix.diag(sv); + + assertArrayEquals(m.toArray(), sm.toArray(), 0.0); + assertArrayEquals(d.toArray(), sm.toArray(), 0.0); + assertArrayEquals(d.toArray(), sd.toArray(), 0.0); + assertArrayEquals(sd.toArray(), s.toArray(), 0.0); + assertArrayEquals(s.toArray(), ss.toArray(), 0.0); + assertArrayEquals(s.values(), ss.values(), 0.0); + assert(s.values().length == 2); + assert(ss.values().length == 2); + assert(s.colPtrs().length == 4); + assert(ss.colPtrs().length == 4); + } + + @Test + public void zerosMatrixConstruction() { + Matrix z = Matrices.zeros(2, 2); + Matrix one = Matrices.ones(2, 2); + DenseMatrix dz = DenseMatrix.zeros(2, 2); + DenseMatrix done = DenseMatrix.ones(2, 2); + + assertArrayEquals(z.toArray(), new double[]{0.0, 0.0, 0.0, 0.0}, 0.0); + assertArrayEquals(dz.toArray(), new double[]{0.0, 0.0, 0.0, 0.0}, 0.0); + assertArrayEquals(one.toArray(), new double[]{1.0, 1.0, 1.0, 1.0}, 0.0); + assertArrayEquals(done.toArray(), new double[]{1.0, 1.0, 1.0, 1.0}, 0.0); + } + + @Test + public void sparseDenseConversion() { + int m = 3; + int n = 2; + double[] values = new double[]{1.0, 2.0, 4.0, 5.0}; + double[] allValues = new double[]{1.0, 2.0, 0.0, 0.0, 4.0, 5.0}; + int[] colPtrs = new int[]{0, 2, 4}; + int[] rowIndices = new int[]{0, 1, 1, 2}; + + SparseMatrix spMat1 = new SparseMatrix(m, n, colPtrs, rowIndices, values); + DenseMatrix deMat1 = new DenseMatrix(m, n, allValues); + + SparseMatrix spMat2 = deMat1.toSparse(); + DenseMatrix deMat2 = spMat1.toDense(); + + assertArrayEquals(spMat1.toArray(), spMat2.toArray(), 0.0); + assertArrayEquals(deMat1.toArray(), deMat2.toArray(), 0.0); + } + + @Test + public void concatenateMatrices() { + int m = 3; + int n = 2; + + Random rng = new Random(42); + SparseMatrix spMat1 = SparseMatrix.sprand(m, n, 0.5, rng); + rng.setSeed(42); + DenseMatrix deMat1 = DenseMatrix.rand(m, n, rng); + Matrix deMat2 = Matrices.eye(3); + Matrix spMat2 = Matrices.speye(3); + Matrix deMat3 = Matrices.eye(2); + Matrix spMat3 = Matrices.speye(2); + + Matrix spHorz = Matrices.horzcat(new Matrix[]{spMat1, spMat2}); + Matrix deHorz1 = Matrices.horzcat(new Matrix[]{deMat1, deMat2}); + Matrix deHorz2 = Matrices.horzcat(new Matrix[]{spMat1, deMat2}); + Matrix deHorz3 = Matrices.horzcat(new Matrix[]{deMat1, spMat2}); + + assert(deHorz1.numRows() == 3); + assert(deHorz2.numRows() == 3); + assert(deHorz3.numRows() == 3); + assert(spHorz.numRows() == 3); + assert(deHorz1.numCols() == 5); + assert(deHorz2.numCols() == 5); + assert(deHorz3.numCols() == 5); + assert(spHorz.numCols() == 5); + + Matrix spVert = Matrices.vertcat(new Matrix[]{spMat1, spMat3}); + Matrix deVert1 = Matrices.vertcat(new Matrix[]{deMat1, deMat3}); + Matrix deVert2 = Matrices.vertcat(new Matrix[]{spMat1, deMat3}); + Matrix deVert3 = Matrices.vertcat(new Matrix[]{deMat1, spMat3}); + + assert(deVert1.numRows() == 5); + assert(deVert2.numRows() == 5); + assert(deVert3.numRows() == 5); + assert(spVert.numRows() == 5); + assert(deVert1.numCols() == 2); + assert(deVert2.numCols() == 2); + assert(deVert3.numCols() == 2); + assert(spVert.numCols() == 2); + } +} http://git-wip-us.apache.org/repos/asf/spark/blob/02b55de3/mllib/src/test/scala/org/apache/spark/mllib/linalg/MatricesSuite.scala ---------------------------------------------------------------------- diff --git a/mllib/src/test/scala/org/apache/spark/mllib/linalg/MatricesSuite.scala b/mllib/src/test/scala/org/apache/spark/mllib/linalg/MatricesSuite.scala index 322a0e9..a35d0fe 100644 --- a/mllib/src/test/scala/org/apache/spark/mllib/linalg/MatricesSuite.scala +++ b/mllib/src/test/scala/org/apache/spark/mllib/linalg/MatricesSuite.scala @@ -43,9 +43,9 @@ class MatricesSuite extends FunSuite { test("sparse matrix construction") { val m = 3 - val n = 2 + val n = 4 val values = Array(1.0, 2.0, 4.0, 5.0) - val colPtrs = Array(0, 2, 4) + val colPtrs = Array(0, 2, 2, 4, 4) val rowIndices = Array(1, 2, 1, 2) val mat = Matrices.sparse(m, n, colPtrs, rowIndices, values).asInstanceOf[SparseMatrix] assert(mat.numRows === m) @@ -53,6 +53,13 @@ class MatricesSuite extends FunSuite { assert(mat.values.eq(values), "should not copy data") assert(mat.colPtrs.eq(colPtrs), "should not copy data") assert(mat.rowIndices.eq(rowIndices), "should not copy data") + + val entries: Array[(Int, Int, Double)] = Array((2, 2, 3.0), (1, 0, 1.0), (2, 0, 2.0), + (1, 2, 2.0), (2, 2, 2.0), (1, 2, 2.0), (0, 0, 0.0)) + + val mat2 = SparseMatrix.fromCOO(m, n, entries) + assert(mat.toBreeze === mat2.toBreeze) + assert(mat2.values.length == 4) } test("sparse matrix construction with wrong number of elements") { @@ -117,6 +124,142 @@ class MatricesSuite extends FunSuite { assert(sparseMat.values(2) === 10.0) } + test("toSparse, toDense") { + val m = 3 + val n = 2 + val values = Array(1.0, 2.0, 4.0, 5.0) + val allValues = Array(1.0, 2.0, 0.0, 0.0, 4.0, 5.0) + val colPtrs = Array(0, 2, 4) + val rowIndices = Array(0, 1, 1, 2) + + val spMat1 = new SparseMatrix(m, n, colPtrs, rowIndices, values) + val deMat1 = new DenseMatrix(m, n, allValues) + + val spMat2 = deMat1.toSparse() + val deMat2 = spMat1.toDense() + + assert(spMat1.toBreeze === spMat2.toBreeze) + assert(deMat1.toBreeze === deMat2.toBreeze) + } + + test("map, update") { + val m = 3 + val n = 2 + val values = Array(1.0, 2.0, 4.0, 5.0) + val allValues = Array(1.0, 2.0, 0.0, 0.0, 4.0, 5.0) + val colPtrs = Array(0, 2, 4) + val rowIndices = Array(0, 1, 1, 2) + + val spMat1 = new SparseMatrix(m, n, colPtrs, rowIndices, values) + val deMat1 = new DenseMatrix(m, n, allValues) + val deMat2 = deMat1.map(_ * 2) + val spMat2 = spMat1.map(_ * 2) + deMat1.update(_ * 2) + spMat1.update(_ * 2) + + assert(spMat1.toArray === spMat2.toArray) + assert(deMat1.toArray === deMat2.toArray) + } + + test("horzcat, vertcat, eye, speye") { + val m = 3 + val n = 2 + val values = Array(1.0, 2.0, 4.0, 5.0) + val allValues = Array(1.0, 2.0, 0.0, 0.0, 4.0, 5.0) + val colPtrs = Array(0, 2, 4) + val rowIndices = Array(0, 1, 1, 2) + + val spMat1 = new SparseMatrix(m, n, colPtrs, rowIndices, values) + val deMat1 = new DenseMatrix(m, n, allValues) + val deMat2 = Matrices.eye(3) + val spMat2 = Matrices.speye(3) + val deMat3 = Matrices.eye(2) + val spMat3 = Matrices.speye(2) + + val spHorz = Matrices.horzcat(Array(spMat1, spMat2)) + val spHorz2 = Matrices.horzcat(Array(spMat1, deMat2)) + val spHorz3 = Matrices.horzcat(Array(deMat1, spMat2)) + val deHorz1 = Matrices.horzcat(Array(deMat1, deMat2)) + + val deHorz2 = Matrices.horzcat(Array[Matrix]()) + + assert(deHorz1.numRows === 3) + assert(spHorz2.numRows === 3) + assert(spHorz3.numRows === 3) + assert(spHorz.numRows === 3) + assert(deHorz1.numCols === 5) + assert(spHorz2.numCols === 5) + assert(spHorz3.numCols === 5) + assert(spHorz.numCols === 5) + assert(deHorz2.numRows === 0) + assert(deHorz2.numCols === 0) + assert(deHorz2.toArray.length === 0) + + assert(deHorz1.toBreeze.toDenseMatrix === spHorz2.toBreeze.toDenseMatrix) + assert(spHorz2.toBreeze === spHorz3.toBreeze) + assert(spHorz(0, 0) === 1.0) + assert(spHorz(2, 1) === 5.0) + assert(spHorz(0, 2) === 1.0) + assert(spHorz(1, 2) === 0.0) + assert(spHorz(1, 3) === 1.0) + assert(spHorz(2, 4) === 1.0) + assert(spHorz(1, 4) === 0.0) + assert(deHorz1(0, 0) === 1.0) + assert(deHorz1(2, 1) === 5.0) + assert(deHorz1(0, 2) === 1.0) + assert(deHorz1(1, 2) == 0.0) + assert(deHorz1(1, 3) === 1.0) + assert(deHorz1(2, 4) === 1.0) + assert(deHorz1(1, 4) === 0.0) + + intercept[IllegalArgumentException] { + Matrices.horzcat(Array(spMat1, spMat3)) + } + + intercept[IllegalArgumentException] { + Matrices.horzcat(Array(deMat1, spMat3)) + } + + val spVert = Matrices.vertcat(Array(spMat1, spMat3)) + val deVert1 = Matrices.vertcat(Array(deMat1, deMat3)) + val spVert2 = Matrices.vertcat(Array(spMat1, deMat3)) + val spVert3 = Matrices.vertcat(Array(deMat1, spMat3)) + val deVert2 = Matrices.vertcat(Array[Matrix]()) + + assert(deVert1.numRows === 5) + assert(spVert2.numRows === 5) + assert(spVert3.numRows === 5) + assert(spVert.numRows === 5) + assert(deVert1.numCols === 2) + assert(spVert2.numCols === 2) + assert(spVert3.numCols === 2) + assert(spVert.numCols === 2) + assert(deVert2.numRows === 0) + assert(deVert2.numCols === 0) + assert(deVert2.toArray.length === 0) + + assert(deVert1.toBreeze.toDenseMatrix === spVert2.toBreeze.toDenseMatrix) + assert(spVert2.toBreeze === spVert3.toBreeze) + assert(spVert(0, 0) === 1.0) + assert(spVert(2, 1) === 5.0) + assert(spVert(3, 0) === 1.0) + assert(spVert(3, 1) === 0.0) + assert(spVert(4, 1) === 1.0) + assert(deVert1(0, 0) === 1.0) + assert(deVert1(2, 1) === 5.0) + assert(deVert1(3, 0) === 1.0) + assert(deVert1(3, 1) === 0.0) + assert(deVert1(4, 1) === 1.0) + + intercept[IllegalArgumentException] { + Matrices.vertcat(Array(spMat1, spMat2)) + } + + intercept[IllegalArgumentException] { + Matrices.vertcat(Array(deMat1, spMat2)) + } + } + test("zeros") { val mat = Matrices.zeros(2, 3).asInstanceOf[DenseMatrix] assert(mat.numRows === 2) @@ -162,4 +305,29 @@ class MatricesSuite extends FunSuite { assert(mat.numCols === 2) assert(mat.values.toSeq === Seq(1.0, 0.0, 0.0, 2.0)) } + + test("sprand") { + val rng = mock[Random] + when(rng.nextInt(4)).thenReturn(0, 1, 1, 3, 2, 2, 0, 1, 3, 0) + when(rng.nextDouble()).thenReturn(1.0, 2.0, 3.0, 4.0, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0) + val mat = SparseMatrix.sprand(4, 4, 0.25, rng) + assert(mat.numRows === 4) + assert(mat.numCols === 4) + assert(mat.rowIndices.toSeq === Seq(3, 0, 2, 1)) + assert(mat.values.toSeq === Seq(1.0, 2.0, 3.0, 4.0)) + val mat2 = SparseMatrix.sprand(2, 3, 1.0, rng) + assert(mat2.rowIndices.toSeq === Seq(0, 1, 0, 1, 0, 1)) + assert(mat2.colPtrs.toSeq === Seq(0, 2, 4, 6)) + } + + test("sprandn") { + val rng = mock[Random] + when(rng.nextInt(4)).thenReturn(0, 1, 1, 3, 2, 2, 0, 1, 3, 0) + when(rng.nextGaussian()).thenReturn(1.0, 2.0, 3.0, 4.0) + val mat = SparseMatrix.sprandn(4, 4, 0.25, rng) + assert(mat.numRows === 4) + assert(mat.numCols === 4) + assert(mat.rowIndices.toSeq === Seq(3, 0, 2, 1)) + assert(mat.values.toSeq === Seq(1.0, 2.0, 3.0, 4.0)) + } } http://git-wip-us.apache.org/repos/asf/spark/blob/02b55de3/mllib/src/test/scala/org/apache/spark/mllib/util/TestingUtils.scala ---------------------------------------------------------------------- diff --git a/mllib/src/test/scala/org/apache/spark/mllib/util/TestingUtils.scala b/mllib/src/test/scala/org/apache/spark/mllib/util/TestingUtils.scala index 30b906a..e957fa5 100644 --- a/mllib/src/test/scala/org/apache/spark/mllib/util/TestingUtils.scala +++ b/mllib/src/test/scala/org/apache/spark/mllib/util/TestingUtils.scala @@ -178,17 +178,17 @@ object TestingUtils { implicit class MatrixWithAlmostEquals(val x: Matrix) { /** - * When the difference of two vectors are within eps, returns true; otherwise, returns false. + * When the difference of two matrices are within eps, returns true; otherwise, returns false. */ def ~=(r: CompareMatrixRightSide): Boolean = r.fun(x, r.y, r.eps) /** - * When the difference of two vectors are within eps, returns false; otherwise, returns true. + * When the difference of two matrices are within eps, returns false; otherwise, returns true. */ def !~=(r: CompareMatrixRightSide): Boolean = !r.fun(x, r.y, r.eps) /** - * Throws exception when the difference of two vectors are NOT within eps; + * Throws exception when the difference of two matrices are NOT within eps; * otherwise, returns true. */ def ~==(r: CompareMatrixRightSide): Boolean = { --------------------------------------------------------------------- To unsubscribe, e-mail: commits-unsubscr...@spark.apache.org For additional commands, e-mail: commits-h...@spark.apache.org
