cern.colt.matrix.tobject.algo.ObjectPartitioning Maven / Gradle / Ivy
Show all versions of parallelcolt Show documentation
package cern.colt.matrix.tobject.algo;
/*
Copyright (C) 1999 CERN - European Organization for Nuclear Research.
Permission to use, copy, modify, distribute and sell this software and its documentation for any purpose
is hereby granted without fee, provided that the above copyright notice appear in all copies and
that both that copyright notice and this permission notice appear in supporting documentation.
CERN makes no representations about the suitability of this software for any purpose.
It is provided "as is" without expressed or implied warranty.
*/
import cern.colt.Swapper;
import cern.colt.function.tint.IntComparator;
import cern.colt.matrix.tobject.ObjectMatrix1D;
import cern.colt.matrix.tobject.ObjectMatrix2D;
/**
* Given some interval boundaries, partitions matrices such that cell values
* falling into an interval are placed next to each other.
*
* Performance
*
* Partitioning into two intervals is O( N ). Partitioning into k
* intervals is O( N * log(k)). Constants factors are minimized.
*
* @see cern.colt.Partitioning "Partitioning arrays (provides more
* documentation)"
*
* @author [email protected]
* @version 1.0, 09/24/99
*/
public class ObjectPartitioning extends Object {
/**
* Makes this class non instantiable, but still let's others inherit from
* it.
*/
protected ObjectPartitioning() {
}
/**
* Same as
* {@link cern.colt.Partitioning#partition(int[],int,int,int[],int,int,int[])}
* except that it synchronously partitions the rows of the given
* matrix by the values of the given matrix column; This is essentially the
* same as partitioning a list of composite objects by some instance
* variable; In other words, two entire rows of the matrix are swapped,
* whenever two column values indicate so.
*
* Let's say, a "row" is an "object" (tuple, d-dimensional point). A
* "column" is the list of "object" values of a given variable (field,
* dimension). A "matrix" is a list of "objects" (tuples, points).
*
* Now, rows (objects, tuples) are partially sorted according to their
* values in one given variable (dimension). Two entire rows of the matrix
* are swapped, whenever two column values indicate so.
*
* Note that arguments are not checked for validity.
*
* Example:
*
*
* 8 x 3 matrix:
23, 22, 21
20, 19, 18
17, 16, 15
14, 13, 12
11, 10, 9
8, 7, 6
5, 4, 3
2, 1, 0
*
*
* column = 0;
rowIndexes = {0,1,2,..,matrix.rows()-1};
rowFrom = 0;
rowTo = matrix.rows()-1;
splitters = {5,10,12}
c = 0;
d = splitters.length-1;
partition(matrix,rowIndexes,rowFrom,rowTo,column,splitters,c,d,splitIndexes);
==>
splitIndexes == {0, 2, 3}
rowIndexes == {7, 6, 5, 4, 0, 1, 2, 3}
*
*
* The matrix IS NOT REORDERED.
* Here is how it would look
* like, if it would be reordered
* accoring to rowIndexes.
* 8 x 3 matrix:
2, 1, 0
5, 4, 3
8, 7, 6
11, 10, 9
23, 22, 21
20, 19, 18
17, 16, 15
14, 13, 12
*
*
*
* @param matrix
* the matrix to be partitioned.
* @param rowIndexes
* the index of the i-th row; is modified by this method to
* reflect partitioned indexes.
* @param rowFrom
* the index of the first row (inclusive).
* @param rowTo
* the index of the last row (inclusive).
* @param column
* the index of the column to partition on.
* @param splitters
* the values at which the rows shall be split into intervals.
* Must be sorted ascending and must not contain multiple
* identical values. These preconditions are not checked; be sure
* that they are met.
*
* @param splitFrom
* the index of the first splitter element to be considered.
* @param splitTo
* the index of the last splitter element to be considered. The
* method considers the splitter elements
* splitters[splitFrom] .. splitters[splitTo].
*
* @param splitIndexes
* a list into which this method fills the indexes of rows
* delimiting intervals. Upon return
* splitIndexes[splitFrom..splitTo] will be set
* accordingly. Therefore, must satisfy
* splitIndexes.length >= splitters.length.
*/
public static void partition(ObjectMatrix2D matrix, int[] rowIndexes, int rowFrom, int rowTo, int column,
final Object[] splitters, int splitFrom, int splitTo, int[] splitIndexes) {
if (rowFrom < 0 || rowTo >= matrix.rows() || rowTo >= rowIndexes.length)
throw new IllegalArgumentException();
if (column < 0 || column >= matrix.columns())
throw new IllegalArgumentException();
if (splitFrom < 0 || splitTo >= splitters.length)
throw new IllegalArgumentException();
if (splitIndexes.length < splitters.length)
throw new IllegalArgumentException();
// this one knows how to swap two row indexes (a,b)
final int[] g = rowIndexes;
Swapper swapper = new Swapper() {
public void swap(int b, int c) {
int tmp = g[b];
g[b] = g[c];
g[c] = tmp;
}
};
// compare splitter[a] with columnView[rowIndexes[b]]
final ObjectMatrix1D columnView = matrix.viewColumn(column);
IntComparator comp = new IntComparator() {
public int compare(int a, int b) {
Comparable av = (Comparable) (splitters[a]);
Comparable bv = (Comparable) (columnView.getQuick(g[b]));
int r = av.compareTo(bv);
return r < 0 ? -1 : (r == 0 ? 0 : 1);
}
};
// compare columnView[rowIndexes[a]] with columnView[rowIndexes[b]]
IntComparator comp2 = new IntComparator() {
public int compare(int a, int b) {
Comparable av = (Comparable) (columnView.getQuick(g[a]));
Comparable bv = (Comparable) (columnView.getQuick(g[b]));
int r = av.compareTo(bv);
return r < 0 ? -1 : (r == 0 ? 0 : 1);
}
};
// compare splitter[a] with splitter[b]
IntComparator comp3 = new IntComparator() {
public int compare(int a, int b) {
Comparable av = (Comparable) (splitters[a]);
Comparable bv = (Comparable) (splitters[b]);
int r = av.compareTo(bv);
return r < 0 ? -1 : (r == 0 ? 0 : 1);
}
};
// generic partitioning does the main work of reordering row indexes
cern.colt.Partitioning.genericPartition(rowFrom, rowTo, splitFrom, splitTo, splitIndexes, comp, comp2, comp3,
swapper);
}
/**
* Same as
* {@link cern.colt.Partitioning#partition(int[],int,int,int[],int,int,int[])}
* except that it synchronously partitions the rows of the given
* matrix by the values of the given matrix column; This is essentially the
* same as partitioning a list of composite objects by some instance
* variable; In other words, two entire rows of the matrix are swapped,
* whenever two column values indicate so.
*
* Let's say, a "row" is an "object" (tuple, d-dimensional point). A
* "column" is the list of "object" values of a given variable (field,
* dimension). A "matrix" is a list of "objects" (tuples, points).
*
* Now, rows (objects, tuples) are partially sorted according to their
* values in one given variable (dimension). Two entire rows of the matrix
* are swapped, whenever two column values indicate so.
*
* Note that arguments are not checked for validity.
*
* Example:
*
*
* 8 x 3 matrix:
23, 22, 21
20, 19, 18
17, 16, 15
14, 13, 12
11, 10, 9
8, 7, 6
5, 4, 3
2, 1, 0
* column = 0;
splitters = {5,10,12}
partition(matrix,column,splitters,splitIndexes);
==>
splitIndexes == {0, 2, 3}
*
* The matrix IS NOT REORDERED.
* The new VIEW IS REORDERED:
* 8 x 3 matrix:
2, 1, 0
5, 4, 3
8, 7, 6
11, 10, 9
23, 22, 21
20, 19, 18
17, 16, 15
14, 13, 12
*
* @param matrix
* the matrix to be partitioned.
* @param column
* the index of the column to partition on.
* @param splitters
* the values at which the rows shall be split into intervals.
* Must be sorted ascending and must not contain multiple
* identical values. These preconditions are not checked; be sure
* that they are met.
*
* @param splitIndexes
* a list into which this method fills the indexes of rows
* delimiting intervals. Therefore, must satisfy
* splitIndexes.length >= splitters.length.
*
* @return a new matrix view having rows partitioned by the given column and
* splitters.
*/
public static ObjectMatrix2D partition(ObjectMatrix2D matrix, int column, final Object[] splitters,
int[] splitIndexes) {
int rowFrom = 0;
int rowTo = matrix.rows() - 1;
int splitFrom = 0;
int splitTo = splitters.length - 1;
int[] rowIndexes = new int[matrix.rows()]; // row indexes to reorder
// instead of matrix itself
for (int i = rowIndexes.length; --i >= 0;)
rowIndexes[i] = i;
partition(matrix, rowIndexes, rowFrom, rowTo, column, splitters, splitFrom, splitTo, splitIndexes);
// take all columns in the original order
int[] columnIndexes = new int[matrix.columns()];
for (int i = columnIndexes.length; --i >= 0;)
columnIndexes[i] = i;
// view the matrix according to the reordered row indexes
return matrix.viewSelection(rowIndexes, columnIndexes);
}
/**
* Same as {@link #partition(int[],int,int,int[],int,int,int[])} except that
* it synchronously partitions the rows of the given matrix by the
* values of the given matrix column; This is essentially the same as
* partitioning a list of composite objects by some instance variable; In
* other words, two entire rows of the matrix are swapped, whenever two
* column values indicate so.
*
* Let's say, a "row" is an "object" (tuple, d-dimensional point). A
* "column" is the list of "object" values of a given variable (field,
* dimension). A "matrix" is a list of "objects" (tuples, points).
*
* Now, rows (objects, tuples) are partially sorted according to their
* values in one given variable (dimension). Two entire rows of the matrix
* are swapped, whenever two column values indicate so.
*
* Of course, the column must not be a column of a different matrix. More
* formally, there must hold:
* There exists an i such that
* matrix.viewColumn(i)==column.
*
* Note that arguments are not checked for validity.
*
* Example:
*
*
* 8 x 3 matrix:
23, 22, 21
20, 19, 18
17, 16, 15
14, 13, 12
11, 10, 9
8, 7, 6
5, 4, 3
2, 1, 0
*
*
* column = matrix.viewColumn(0);
a = 0;
b = column.size()-1;
splitters={5,10,12}
c=0;
d=splitters.length-1;
partition(matrix,column,a,b,splitters,c,d,splitIndexes);
==>
splitIndexes == {0, 2, 3}
*
*
* 8 x 3 matrix:
2, 1, 0
5, 4, 3
8, 7, 6
11, 10, 9
23, 22, 21
20, 19, 18
17, 16, 15
14, 13, 12
*
*
*/
private static void xPartitionOld(ObjectMatrix2D matrix, ObjectMatrix1D column, int from, int to,
Object[] splitters, int splitFrom, int splitTo, int[] splitIndexes) {
/*
* Object splitter; // int, Object --> template type dependent
*
* if (splitFrom>splitTo) return; // nothing to do if (from>to) { // all
* bins are empty from--; for (int i = splitFrom; i<=splitTo; )
* splitIndexes[i++] = from; return; } // Choose a partition (pivot)
* index, m // Ideally, the pivot should be the median, because a median
* splits a list into two equal sized sublists. // However, computing
* the median is expensive, so we use an approximation. int medianIndex;
* if (splitFrom==splitTo) { // we don't really have a choice
* medianIndex = splitFrom; } else { // we do have a choice int m =
* (from+to) / 2; // Small arrays, middle element int len = to-from+1;
* if (len > SMALL) { int l = from; int n = to; if (len > MEDIUM) { //
* Big arrays, pseudomedian of 9 int s = len/8; l = med3(column, l, l+s,
* l+2*s); m = med3(column, m-s, m, m+s); n = med3(column, n-2*s, n-s,
* n); } m = med3(column, l, m, n); // Mid-size, pseudomedian of 3 } //
* Find the splitter closest to the pivot, i.e. the splitter that best
* splits the list into two equal sized sublists. medianIndex =
* cern.colt.Sorting.binarySearchFromTo(splitters,column.getQuick(m),splitFrom,splitTo);
* if (medianIndex < 0) medianIndex = -medianIndex - 1; // not found if
* (medianIndex > splitTo) medianIndex = splitTo; // not found, one past
* the end } splitter = splitters[medianIndex]; // Partition the list
* according to the splitter, i.e. // Establish invariant: list[i] <
* splitter <= list[j] for i=from..medianIndex and j=medianIndex+1 .. to
* int splitIndex = xPartitionOld(matrix,column,from,to,splitter);
* splitIndexes[medianIndex] = splitIndex; // Optimization: Handle
* special cases to cut down recursions. if (splitIndex < from) { // no
* element falls into this bin // all bins with splitters[i] <= splitter
* are empty int i = medianIndex-1; while (i>=splitFrom && (!(splitter <
* splitters[i]))) splitIndexes[i--] = splitIndex; splitFrom =
* medianIndex+1; } else if (splitIndex >= to) { // all elements fall
* into this bin // all bins with splitters[i] >= splitter are empty int
* i = medianIndex+1; while (i<=splitTo && (!(splitter >
* splitters[i]))) splitIndexes[i++] = splitIndex; splitTo =
* medianIndex-1; } // recursively partition left half if (splitFrom <=
* medianIndex-1) { xPartitionOld(matrix, column, from, splitIndex,
* splitters, splitFrom, medianIndex-1, splitIndexes); } // recursively
* partition right half if (medianIndex+1 <= splitTo) {
* xPartitionOld(matrix, column, splitIndex+1, to, splitters,
* medianIndex+1, splitTo, splitIndexes); }
*/
}
/**
* Same as {@link #partition(int[],int,int,int)} except that it
* synchronously partitions the rows of the given matrix by the
* values of the given matrix column; This is essentially the same as
* partitioning a list of composite objects by some instance variable; In
* other words, two entire rows of the matrix are swapped, whenever two
* column values indicate so.
*
* Let's say, a "row" is an "object" (tuple, d-dimensional point). A
* "column" is the list of "object" values of a given variable (field,
* dimension). A "matrix" is a list of "objects" (tuples, points).
*
* Now, rows (objects, tuples) are partially sorted according to their
* values in one given variable (dimension). Two entire rows of the matrix
* are swapped, whenever two column values indicate so.
*
* Of course, the column must not be a column of a different matrix. More
* formally, there must hold:
* There exists an i such that
* matrix.viewColumn(i)==column.
*
* Note that arguments are not checked for validity.
*/
private static int xPartitionOld(ObjectMatrix2D matrix, ObjectMatrix1D column, int from, int to, Object splitter) {
/*
* Object element; // int, Object --> template type dependent for (int
* i=from-1; ++i<=to; ) { element = column.getQuick(i); if (element <
* splitter) { // swap x[i] with x[from] matrix.swapRows(i,from);
* from++; } } return from-1;
*/
return 0;
}
}