对于Java或Scala中的calc大型稀疏矩阵，我可以使用什么库？

Question

当我使用大型稀疏矩阵时，最好使用压缩矩阵，如CCS、CRS等。

我试着用ScalaNLP，la4j，colc来计算100,000*100,000稀疏矩阵。有一些问题。

1. 微风(ScalaNLP/Scalala)

- it gives me the `CSCMatrix` type which can have 100,000\*100,000 size.
- but the problem is it's under development.
- so we cannot calc element-wise product of `CSCMatrix` with `CSCMatrix`, like `csc1 :* csc2`.
- and also you cannot add CSCMatrixes to each other.

​

1. la4j

- It has CCSMatrix and CRSMatrix.
- but when creating (new CCSMatrixFactory).createMatrix(100000, 100000), it occur OutOfMemoryError.
- The matrix should be zeros, so it should not use large memory spaces.

​

1. colc

- It has SparseDoubleMatrix2D.
- but when creating matrix like new SparseDoubleMatrix2d(100000, 100000), it says IllegalArgumentException: matrix too large.

​

对于计算大型稀疏矩阵，我可以使用哪些库？你能给我举个例子吗？

Answer 1

我对Breeze很好奇，所以我查了一下消息来源。这有点混乱，因为操作符都是由println样式的代码生成(！)发出的。但我想到了这个：

import breeze.linalg.operators.{BinaryOp, OpMulScalar}

object CSCMatrixExtraOps {
  abstract class CSCMatrixCanMulM_M[@specialized (Int, Float, Long, Double) A]
    extends BinaryOp[CSCMatrix[A], CSCMatrix[A], OpMulScalar, CSCMatrix[A]] {

    protected def times(a: A, b: A): A

    protected def zeros  (rows: Int, cols: Int): CSCMatrix[A]
    protected def builder(rows: Int, cols: Int, sz: Int): CSCMatrix.Builder[A]

    final def apply(a: CSCMatrix[A], b: CSCMatrix[A]): CSCMatrix[A] = {
      val rows  = a.rows
      val cols  = a.cols
      require(rows == b.rows, "Matrices must have same number of rows!")
      require(cols == b.cols, "Matrices must have same number of cols!")

      if (cols == 0) return zeros(rows, cols)

      val res     = builder(rows, cols, math.min(a.activeSize, b.activeSize))
      var ci      = 0
      var acpStop = a.colPtrs(0)
      var bcpStop = b.colPtrs(0)
      while (ci < cols) {
        val ci1 = ci + 1
        var acp = acpStop
        var bcp = bcpStop
        acpStop = a.colPtrs(ci1)
        bcpStop = b.colPtrs(ci1)
        while (acp < acpStop && bcp < bcpStop) {
          val ari = a.rowIndices(acp)
          val bri = b.rowIndices(bcp)
          if (ari == bri) {
            val v = times(a.data(acp), b.data(bcp))
            res.add(ari, ci, v)
            acp += 1
            bcp += 1
          } else if (ari < bri) {
            acp += 1
          } else /* ari > bri */ {
            bcp += 1
          }
        }
        ci = ci1
      }

      res.result()
    }
  }

  implicit object CSCMatrixCanMulM_M_Int extends CSCMatrixCanMulM_M[Int] {
    protected def times(a: Int, b: Int) = a * b
    protected def zeros(rows: Int, cols: Int) = CSCMatrix.zeros(rows, cols)
    protected def builder(rows: Int, cols: Int, sz: Int) = 
      new CSCMatrix.Builder(rows, cols, sz)
  }

  implicit object CSCMatrixCanMulM_M_Double extends CSCMatrixCanMulM_M[Double] {
    protected def times(a: Double, b: Double) = a * b
    protected def zeros(rows: Int, cols: Int) = CSCMatrix.zeros(rows, cols)
    protected def builder(rows: Int, cols: Int, sz: Int) = 
      new CSCMatrix.Builder(rows, cols, sz)
  }
}

示例：

import breeze.linalg._
import CSCMatrixExtraOps._

val m1 = CSCMatrix((0, 0, 0), (0, 5, 0), (0, 0, 10), (0, 13, 0))
val m2 = CSCMatrix((0, 0, 0), (0, 5, 0), (0, 0, 10), (13, 0, 0))
(m1 :* m2).toDenseMatrix

结果：

0  0   0    
0  25  0    
0  0   100  
0  0   0    

Answer 2

我是la4j库的作者。让我给你几个建议。因此，当您创建一个新的CRS/CCS矩阵时，la4j只为它分配32长数组(这是一个默认的最小大小)。因此，它不能抛出OOM错误(我刚刚检查过)：

Matrix a = Matrices.CRS_FACTORY.createMatrix(100000, 100000);

但是，最好使用公共构造函数：

Matrix a = new CCSMatrix(100000, 100000);

无论如何，如果仍然收到此错误，请尝试使用-Xmx1024m -Xms512m扩展堆大小。

你所说的“矩阵应该是零，所以它不应该使用大内存空间”是什么意思。我不确定我是否正确理解了它。

顺便说一句，使用la4j：0.4.0的最后一个版本。可能，您发现的问题是由这个拉扯请求解决的。