kdtree.py

# -*- coding: utf-8 -*-


"""A Python implemntation of a kd-tree

This package provides a simple implementation of a kd-tree in Python.
https://en.wikipedia.org/wiki/K-d_tree
"""

from __future__ import print_function

import operator
import math
from collections import deque
from functools import wraps

__author__ = u'Stefan Kögl <stefan@skoegl.net>'
__version__ = '0.9'
__website__ = 'https://github.com/stefankoegl/kdtree'
__license__ = 'ISC license'


# maps child position to its comparison operator
COMPARE_CHILD = {
    0: (operator.le, operator.sub),
    1: (operator.ge, operator.add),
}


class Node(object):
    """ A Node in a kd-tree

    A tree is represented by its root node, and every node represents
    its subtree"""

    def __init__(self, data=None, left=None, right=None):
        self.data = data
        self.left = left
        self.right = right


    @property
    def is_leaf(self):
        """ Returns True if a Node has no subnodes

        >>> Node().is_leaf
        True

        >>> Node( 1, left=Node(2) ).is_leaf
        False
        """
        return (not self.data) or \
               (all(not bool(c) for c, p in self.children))


    def preorder(self):
        """ iterator for nodes: root, left, right """

        if not self:
            return

        yield self

        if self.left:
            for x in self.left.preorder():
                yield x

        if self.right:
            for x in self.right.preorder():
                yield x


    def inorder(self):
        """ iterator for nodes: left, root, right """

        if not self:
            return

        if self.left:
            for x in self.left.inorder():
                yield x

        yield self

        if self.right:
            for x in self.right.inorder():
                yield x


    def postorder(self):
        """ iterator for nodes: left, right, root """

        if not self:
            return

        if self.left:
            for x in self.left.postorder():
                yield x

        if self.right:
            for x in self.right.postorder():
                yield x

        yield self


    @property
    def children(self):
        """
        Returns an iterator for the non-empty children of the Node

        The children are returned as (Node, pos) tuples where pos is 0 for the
        left subnode and 1 for the right.

        >>> len(list(create(dimensions=2).children))
        0

        >>> len(list(create([ (1, 2) ]).children))
        0

        >>> len(list(create([ (2, 2), (2, 1), (2, 3) ]).children))
        2
        """

        if self.left and self.left.data is not None:
            yield self.left, 0
        if self.right and self.right.data is not None:
            yield self.right, 1


    def set_child(self, index, child):
        """ Sets one of the node's children

        index 0 refers to the left, 1 to the right child """

        if index == 0:
            self.left = child
        else:
            self.right = child


    def height(self):
        """
        Returns height of the (sub)tree, without considering
        empty leaf-nodes

        >>> create(dimensions=2).height()
        0

        >>> create([ (1, 2) ]).height()
        1

        >>> create([ (1, 2), (2, 3) ]).height()
        2
        """

        min_height = int(bool(self))
        return max([min_height] + [c.height()+1 for c, p in self.children])


    def get_child_pos(self, child):
        """ Returns the position if the given child

        If the given node is the left child, 0 is returned. If its the right
        child, 1 is returned. Otherwise None """

        for c, pos in self.children:
            if child == c:
                return pos


    def __repr__(self):
        return '<%(cls)s - %(data)s>' % \
            dict(cls=self.__class__.__name__, data=repr(self.data))


    def __nonzero__(self):
        return self.data is not None

    __bool__ = __nonzero__

    def __eq__(self, other):
        if isinstance(other, tuple):
            return self.data == other
        else:
            return self.data == other.data

    def __hash__(self):
        try:
            return hash(self.data)
        except TypeError:
            # try to hash un-hashable types as tuples
            return hash(tuple(self.data))


def require_axis(f):
    """ Check if the object of the function has axis and sel_axis members """

    @wraps(f)
    def _wrapper(self, *args, **kwargs):
        if None in (self.axis, self.sel_axis):
            raise ValueError('%(func_name) requires the node %(node)s '
                    'to have an axis and a sel_axis function' %
                    dict(func_name=f.__name__, node=repr(self)))

        return f(self, *args, **kwargs)

    return _wrapper



class KDNode(Node):
    """ A Node that contains kd-tree specific data and methods """


    def __init__(self, data=None, left=None, right=None, axis=None,
            sel_axis=None, dimensions=None):
        """ Creates a new node for a kd-tree

        If the node will be used within a tree, the axis and the sel_axis
        function should be supplied.

        sel_axis(axis) is used when creating subnodes of the current node. It
        receives the axis of the parent node and returns the axis of the child
        node. """

        super(KDNode, self).__init__(data, left, right)
        self.axis = axis
        self.sel_axis = sel_axis
        self.dimensions = dimensions


    @require_axis
    def add(self, point):
        """
        Adds a point to the current node or iteratively
        descends to one of its children.

        Users should call add() only to the topmost tree.
        """

        current = self
        while True:
            check_dimensionality([point], dimensions=current.dimensions)

            # Adding has hit an empty leaf-node, add here
            if current.data is None:
                current.data = point
                break

            # split on self.axis, recurse either left or right
            if point[current.axis] < current.data[current.axis]:
                if current.left is None:
                    current.left = current.create_subnode(point)
                    break
                else:
                    current = current.left
                    #self.left.add(point)
            else:
                if current.right is None:
                    current.right = current.create_subnode(point)
                    break
                else:
                    current = current.right


    @require_axis
    def create_subnode(self, data):
        """ Creates a subnode for the current node """

        return self.__class__(data,
                axis=self.sel_axis(self.axis),
                sel_axis=self.sel_axis,
                dimensions=self.dimensions)


    @require_axis
    def find_replacement(self):
        """ Finds a replacement for the current node

        The replacement is returned as a
        (replacement-node, replacements-parent-node) tuple """

        if self.right:
            child, parent = self.right.extreme_child(min, self.axis)
        else:
            child, parent = self.left.extreme_child(max, self.axis)

        return (child, parent if parent is not None else self)


    def should_remove(self, point, node):
        """ checks if self's point (and maybe identity) matches """
        if not self.data == point:
            return False

        return (node is None) or (node is self)


    @require_axis
    def remove(self, point, node=None):
        """ Removes the node with the given point from the tree

        Returns the new root node of the (sub)tree.

        If there are multiple points matching "point", only one is removed. The
        optional "node" parameter is used for checking the identity, once the
        removeal candidate is decided."""

        # Recursion has reached an empty leaf node, nothing here to delete
        if not self:
            return

        # Recursion has reached the node to be deleted
        if self.should_remove(point, node):
            return self._remove(point)

        # Remove direct subnode
        if self.left and self.left.should_remove(point, node):
            self.left = self.left._remove(point)

        elif self.right and self.right.should_remove(point, node):
            self.right = self.right._remove(point)

        # Recurse to subtrees
        if point[self.axis] <= self.data[self.axis]:
            if self.left:
                self.left = self.left.remove(point, node)

        if point[self.axis] >= self.data[self.axis]:
            if self.right:
                self.right = self.right.remove(point, node)

        return self


    @require_axis
    def _remove(self, point):
        # we have reached the node to be deleted here

        # deleting a leaf node is trivial
        if self.is_leaf:
            self.data = None
            return self

        # we have to delete a non-leaf node here

        # find a replacement for the node (will be the new subtree-root)
        root, max_p = self.find_replacement()

        # self and root swap positions
        tmp_l, tmp_r = self.left, self.right
        self.left, self.right = root.left, root.right
        root.left, root.right = tmp_l if tmp_l is not root else self, tmp_r if tmp_r is not root else self
        self.axis, root.axis = root.axis, self.axis

        # Special-case if we have not chosen a direct child as the replacement
        if max_p is not self:
            pos = max_p.get_child_pos(root)
            max_p.set_child(pos, self)
            max_p.remove(point, self)

        else:
            root.remove(point, self)

        return root


    @property
    def is_balanced(self):
        """ Returns True if the (sub)tree is balanced

        The tree is balanced if the heights of both subtrees differ at most by
        1 """

        left_height = self.left.height() if self.left else 0
        right_height = self.right.height() if self.right else 0

        if abs(left_height - right_height) > 1:
            return False

        return all(c.is_balanced for c, _ in self.children)


    def rebalance(self):
        """
        Returns the (possibly new) root of the rebalanced tree
        """

        return create([x.data for x in self.inorder()])


    def axis_dist(self, point, axis):
        """
        Squared distance at the given axis between
        the current Node and the given point
        """
        return math.pow(self.data[axis] - point[axis], 2)


    def dist(self, point):
        """
        Squared distance between the current Node
        and the given point
        """
        r = range(len(self.data))
        return sum([self.axis_dist(point, i) for i in r])


    def search_knn(self, point, k):
        """ Return the k nearest neighbors of point

        point must be an actual point, not a node """

        results = set()

        prev = None
        current = self
        get_dist = lambda n: n.dist(point)

        # the nodes do not keep a reference to their parents
        parents = {id(current): None}

        # go down the tree as we would for inserting
        while current:
            if point[current.axis] < current.data[current.axis]:
                # left side
                parents[id(current.left)] = current
                prev = current
                current = current.left
            else:
                # right side
                parents[id(current.right)] = current
                prev = current
                current = current.right

        if not prev:
            return []

        examined = set()

        # Go up the tree, looking for better solutions
        current = prev
        while current:
            # search node and update results
            current._search_node(point, k, results, examined, get_dist)
            current = parents[id(current)]

        return sorted(results, key=get_dist)


    def _search_node(self, point, k, results, examined, get_dist):
        examined.add(self)

        # get current best
        bestNode = next(iter(sorted(results, key=get_dist)), None)
        bestDist = get_dist(bestNode) if bestNode is not None else float('inf')

        # If the current node is closer than the current best, then it
        # becomes the current best.
        nodeDist = get_dist(self)
        if nodeDist < bestDist:
            if len(results) == k and bestNode:
               results.remove(bestNode)

            results.add(self)

        # if we're equal to the current best, add it, regardless of k
        elif nodeDist == bestDist:
            results.add(self)

        # if we don't have k results yet, add it anyway
        elif len(results) < k:
            results.add(self)

        # get new best
        bestNode = next(iter(sorted(results, key=get_dist)))
        bestDist = get_dist(bestNode)

        # Check whether there could be any points on the other side of the
        # splitting plane that are closer to the search point than the current
        # best.
        for child, pos in self.children:
            if child in examined:
                continue

            examined.add(child)
            compare, combine = COMPARE_CHILD[pos]

            # Since the hyperplanes are all axis-aligned this is implemented
            # as a simple comparison to see whether the difference between the
            # splitting coordinate of the search point and current node is less
            # than the distance (overall coordinates) from the search point to
            # the current best.
            nodePoint = self.data[self.axis]
            pointPlusDist = combine(point[self.axis], bestDist)
            lineIntersects = compare(pointPlusDist, nodePoint)

            # If the hypersphere crosses the plane, there could be nearer
            # points on the other side of the plane, so the algorithm must move
            # down the other branch of the tree from the current node looking
            # for closer points, following the same recursive process as the
            # entire search.
            if lineIntersects:
                child._search_node(point, k, results, examined, get_dist)


    @require_axis
    def search_nn(self, point, best=None):
        """
        Search the nearest node of the given point

        point must be a location, not a node. The nearest node to the point is
        returned. If a location of an actual node is used, the Node with this
        location will be retuend (not its neighbor) """

        return next(iter(self.search_knn(point, 1)), None)


    @require_axis
    def search_nn_dist(self, point, distance, best=None):
        """
        Search the n nearest nodes of the given point which are within given
        distance

        point must be a location, not a node. A list containing the n nearest
        nodes to the point within the distance will be returned.
        """

        if best is None:
            best = []

        # consider the current node
        if self.dist(point) < distance:
            best.append(self)

        # sort the children, nearer one first (is this really necessairy?)
        children = sorted(self.children, key=lambda c_p1: c_p1[0].dist(point))

        for child, p in children:
            # check if child node needs to be recursed
            if self.axis_dist(point, self.axis) < math.pow(distance, 2):
                child.search_nn_dist(point, distance, best)

        return best


    @require_axis
    def is_valid(self):
        """ Checks recursively if the tree is valid

        It is valid if each node splits correctly """

        if not self:
            return True

        if self.left and self.data[self.axis] < self.left.data[self.axis]:
            return False

        if self.right and self.data[self.axis] > self.right.data[self.axis]:
            return False

        return all(c.is_valid() for c, _ in self.children) or self.is_leaf


    def extreme_child(self, sel_func, axis):
        """ Returns a child of the subtree and its parent

        The child is selected by sel_func which is either min or max
        (or a different function with similar semantics). """

        max_key = lambda child_parent: child_parent[0].data[axis]


        # we don't know our parent, so we include None
        me = [(self, None)] if self else []

        child_max = [c.extreme_child(sel_func, axis) for c, _ in self.children]
        # insert self for unknown parents
        child_max = [(c, p if p is not None else self) for c, p in child_max]

        candidates =  me + child_max

        if not candidates:
            return None, None

        return sel_func(candidates, key=max_key)



def create(point_list=None, dimensions=None, axis=0, sel_axis=None):
    """ Creates a kd-tree from a list of points

    All points in the list must be of the same dimensionality.

    If no point_list is given, an empty tree is created. The number of
    dimensions has to be given instead.

    If both a point_list and dimensions are given, the numbers must agree.

    Axis is the axis on which the root-node should split.

    sel_axis(axis) is used when creating subnodes of a node. It receives the
    axis of the parent node and returns the axis of the child node. """

    if not point_list and not dimensions:
        raise ValueError('either point_list or dimensions must be provided')

    elif point_list:
        dimensions = check_dimensionality(point_list, dimensions)

    # by default cycle through the axis
    sel_axis = sel_axis or (lambda prev_axis: (prev_axis+1) % dimensions)

    if not point_list:
        return KDNode(sel_axis=sel_axis, axis=axis, dimensions=dimensions)

    # Sort point list and choose median as pivot element
    point_list.sort(key=lambda point: point[axis])
    median = len(point_list) // 2

    loc   = point_list[median]
    left  = create(point_list[:median], dimensions, sel_axis(axis))
    right = create(point_list[median + 1:], dimensions, sel_axis(axis))
    return KDNode(loc, left, right, axis=axis, sel_axis=sel_axis)


def check_dimensionality(point_list, dimensions=None):
    dimensions = dimensions or len(point_list[0])
    for p in point_list:
        if len(p) != dimensions:
            raise ValueError('All Points in the point_list must have the same dimensionality')

    return dimensions



def level_order(tree, include_all=False):
    """ Returns an iterator over the tree in level-order

    If include_all is set to True, empty parts of the tree are filled
    with dummy entries and the iterator becomes infinite. """

    q = deque()
    q.append(tree)
    while q:
        node = q.popleft()
        yield node

        if include_all or node.left:
            q.append(node.left or node.__class__())

        if include_all or node.right:
            q.append(node.right or node.__class__())



def visualize(tree, max_level=100, node_width=10, left_padding=5):
    """ Prints the tree to stdout """

    height = min(max_level, tree.height()-1)
    max_width = pow(2, height)

    per_level = 1
    in_level  = 0
    level     = 0

    for node in level_order(tree, include_all=True):

        if in_level == 0:
            print()
            print()
            print(' '*left_padding, end=' ')

        width = int(max_width*node_width/per_level)

        node_str = (str(node.data) if node else '').center(width)
        print(node_str, end=' ')

        in_level += 1

        if in_level == per_level:
            in_level   = 0
            per_level *= 2
            level     += 1

        if level > height:
            break

    print()
    print()