scikit-learn/sklearn/utils/graph_shortest_path.pyx

611 lines
20 KiB
Cython

"""
Routines for performing shortest-path graph searches
The main interface is in the function `graph_shortest_path`. This
calls cython routines that compute the shortest path using either
the Floyd-Warshall algorithm, or Dykstra's algorithm with Fibonacci Heaps.
"""
# Author: Jake Vanderplas -- <vanderplas@astro.washington.edu>
# License: BSD 3 clause, (C) 2011
import numpy as np
cimport numpy as np
from scipy.sparse import csr_matrix, isspmatrix, isspmatrix_csr
cimport cython
from libc.stdlib cimport malloc, free
np.import_array()
DTYPE = np.float64
ctypedef np.float64_t DTYPE_t
ITYPE = np.int32
ctypedef np.int32_t ITYPE_t
def graph_shortest_path(dist_matrix, directed=True, method='auto'):
"""
Perform a shortest-path graph search on a positive directed or
undirected graph.
Parameters
----------
dist_matrix : arraylike or sparse matrix, shape = (N,N)
Array of positive distances.
If vertex i is connected to vertex j, then dist_matrix[i,j] gives
the distance between the vertices.
If vertex i is not connected to vertex j, then dist_matrix[i,j] = 0
directed : boolean
if True, then find the shortest path on a directed graph: only
progress from a point to its neighbors, not the other way around.
if False, then find the shortest path on an undirected graph: the
algorithm can progress from a point to its neighbors and vice versa.
method : string ['auto'|'FW'|'D']
method to use. Options are
'auto' : attempt to choose the best method for the current problem
'FW' : Floyd-Warshall algorithm. O[N^3]
'D' : Dijkstra's algorithm with Fibonacci stacks. O[(k+log(N))N^2]
Returns
-------
G : np.ndarray, float, shape = [N,N]
G[i,j] gives the shortest distance from point i to point j
along the graph.
Notes
-----
As currently implemented, Dijkstra's algorithm does not work for
graphs with direction-dependent distances when directed == False.
i.e., if dist_matrix[i,j] and dist_matrix[j,i] are not equal and
both are nonzero, method='D' will not necessarily yield the correct
result.
Also, these routines have not been tested for graphs with negative
distances. Negative distances can lead to infinite cycles that must
be handled by specialized algorithms.
"""
if not isspmatrix_csr(dist_matrix):
dist_matrix = csr_matrix(dist_matrix)
N = dist_matrix.shape[0]
Nk = len(dist_matrix.data)
if method == 'auto':
if Nk < N * N / 4:
method = 'D'
else:
method = 'FW'
if method == 'FW':
graph = np.asarray(dist_matrix.toarray(), dtype=DTYPE, order='C')
floyd_warshall(graph, directed)
elif method == 'D':
graph = np.zeros((N, N), dtype=DTYPE, order='C')
dijkstra(dist_matrix, graph, directed)
else:
raise ValueError("unrecognized method '%s'" % method)
return graph
@cython.boundscheck(False)
cdef np.ndarray floyd_warshall(np.ndarray[DTYPE_t, ndim=2, mode='c'] graph,
int directed=0):
"""
FloydWarshall algorithm
Parameters
----------
graph : ndarray
on input, graph is the matrix of distances between connected points.
unconnected points have distance=0
on exit, graph is overwritten with the output
directed : bool, default = False
if True, then the algorithm will only traverse from a point to
its neighbors when finding the shortest path.
if False, then the algorithm will traverse all paths in both
directions.
Returns
-------
graph : ndarray
the matrix of shortest paths between points.
If no path exists, the path length is zero
"""
cdef int N = graph.shape[0]
assert graph.shape[1] == N
cdef unsigned int i, j, k, m
cdef DTYPE_t infinity = np.inf
cdef DTYPE_t sum_ijk
#initialize all distances to infinity
graph[np.where(graph == 0)] = infinity
#graph[i,i] should be zero
graph.flat[::N + 1] = 0
# for a non-directed graph, we need to symmetrize the distances
if not directed:
for i from 0 <= i < N:
for j from i + 1 <= j < N:
if graph[j, i] <= graph[i, j]:
graph[i, j] = graph[j, i]
else:
graph[j, i] = graph[i, j]
#now perform the Floyd-Warshall algorithm
for k from 0 <= k < N:
for i from 0 <= i < N:
if graph[i, k] == infinity:
continue
for j from 0 <= j < N:
sum_ijk = graph[i, k] + graph[k, j]
if sum_ijk < graph[i, j]:
graph[i, j] = sum_ijk
graph[np.where(np.isinf(graph))] = 0
return graph
@cython.boundscheck(False)
cdef np.ndarray dijkstra(dist_matrix,
np.ndarray[DTYPE_t, ndim=2] graph,
int directed=0):
"""
Dijkstra algorithm using Fibonacci Heaps
Parameters
----------
graph : array or sparse matrix
dist_matrix is the matrix of distances between connected points.
unconnected points have distance=0. It will be converted to
a csr_matrix internally
indptr :
These arrays encode a distance matrix in compressed-sparse-row
format.
graph : ndarray
on input, graph is the matrix of distances between connected points.
unconnected points have distance=0
on exit, graph is overwritten with the output
directed : bool, default = False
if True, then the algorithm will only traverse from a point to
its neighbors when finding the shortest path.
if False, then the algorithm will traverse all paths in both
directions.
Returns
-------
graph : array
the matrix of shortest paths between points.
If no path exists, the path length is zero
"""
cdef unsigned int N = graph.shape[0]
cdef unsigned int i
cdef FibonacciHeap heap
cdef FibonacciNode* nodes = <FibonacciNode*> malloc(N *
sizeof(FibonacciNode))
cdef np.ndarray distances, neighbors, indptr
cdef np.ndarray distances2, neighbors2, indptr2
if not isspmatrix_csr(dist_matrix):
dist_matrix = csr_matrix(dist_matrix)
distances = np.asarray(dist_matrix.data, dtype=DTYPE, order='C')
neighbors = np.asarray(dist_matrix.indices, dtype=ITYPE, order='C')
indptr = np.asarray(dist_matrix.indptr, dtype=ITYPE, order='C')
for i from 0 <= i < N:
initialize_node(&nodes[i], i)
heap.min_node = NULL
if directed:
for i from 0 <= i < N:
dijkstra_directed_one_row(i, neighbors, distances, indptr,
graph, &heap, nodes)
else:
#use the csr -> csc sparse matrix conversion to quickly get
# both directions of neigbors
dist_matrix_T = dist_matrix.T.tocsr()
distances2 = np.asarray(dist_matrix_T.data,
dtype=DTYPE, order='C')
neighbors2 = np.asarray(dist_matrix_T.indices,
dtype=ITYPE, order='C')
indptr2 = np.asarray(dist_matrix_T.indptr,
dtype=ITYPE, order='C')
for i from 0 <= i < N:
dijkstra_one_row(i, neighbors, distances, indptr,
neighbors2, distances2, indptr2,
graph, &heap, nodes)
free(nodes)
return graph
######################################################################
# FibonacciNode structure
# This structure and the operations on it are the nodes of the
# Fibonacci heap.
#
cdef struct FibonacciNode:
unsigned int index
unsigned int rank
unsigned int state
DTYPE_t val
FibonacciNode* parent
FibonacciNode* left_sibling
FibonacciNode* right_sibling
FibonacciNode* children
cdef void initialize_node(FibonacciNode* node,
unsigned int index,
DTYPE_t val=0):
# Assumptions: - node is a valid pointer
# - node is not currently part of a heap
node.index = index
node.val = val
node.rank = 0
node.state = 0 # 0 -> NOT_IN_HEAP
node.parent = NULL
node.left_sibling = NULL
node.right_sibling = NULL
node.children = NULL
cdef FibonacciNode* rightmost_sibling(FibonacciNode* node):
# Assumptions: - node is a valid pointer
cdef FibonacciNode* temp = node
while(temp.right_sibling):
temp = temp.right_sibling
return temp
cdef FibonacciNode* leftmost_sibling(FibonacciNode* node):
# Assumptions: - node is a valid pointer
cdef FibonacciNode* temp = node
while(temp.left_sibling):
temp = temp.left_sibling
return temp
cdef void add_child(FibonacciNode* node, FibonacciNode* new_child):
# Assumptions: - node is a valid pointer
# - new_child is a valid pointer
# - new_child is not the sibling or child of another node
new_child.parent = node
if node.children:
add_sibling(node.children, new_child)
else:
node.children = new_child
new_child.right_sibling = NULL
new_child.left_sibling = NULL
node.rank = 1
cdef void add_sibling(FibonacciNode* node, FibonacciNode* new_sibling):
# Assumptions: - node is a valid pointer
# - new_sibling is a valid pointer
# - new_sibling is not the child or sibling of another node
cdef FibonacciNode* temp = rightmost_sibling(node)
temp.right_sibling = new_sibling
new_sibling.left_sibling = temp
new_sibling.right_sibling = NULL
new_sibling.parent = node.parent
if new_sibling.parent:
new_sibling.parent.rank += 1
cdef void remove(FibonacciNode* node):
# Assumptions: - node is a valid pointer
if node.parent:
node.parent.rank -= 1
if node.left_sibling:
node.parent.children = node.left_sibling
elif node.right_sibling:
node.parent.children = node.right_sibling
else:
node.parent.children = NULL
if node.left_sibling:
node.left_sibling.right_sibling = node.right_sibling
if node.right_sibling:
node.right_sibling.left_sibling = node.left_sibling
node.left_sibling = NULL
node.right_sibling = NULL
node.parent = NULL
######################################################################
# FibonacciHeap structure
# This structure and operations on it use the FibonacciNode
# routines to implement a Fibonacci heap
ctypedef FibonacciNode* pFibonacciNode
cdef struct FibonacciHeap:
FibonacciNode* min_node
pFibonacciNode[100] roots_by_rank # maximum number of nodes is ~2^100.
cdef void insert_node(FibonacciHeap* heap,
FibonacciNode* node):
# Assumptions: - heap is a valid pointer
# - node is a valid pointer
# - node is not the child or sibling of another node
if heap.min_node:
add_sibling(heap.min_node, node)
if node.val < heap.min_node.val:
heap.min_node = node
else:
heap.min_node = node
cdef void decrease_val(FibonacciHeap* heap,
FibonacciNode* node,
DTYPE_t newval):
# Assumptions: - heap is a valid pointer
# - newval <= node.val
# - node is a valid pointer
# - node is not the child or sibling of another node
node.val = newval
if node.parent and (node.parent.val >= newval):
remove(node)
insert_node(heap, node)
elif heap.min_node.val > node.val:
heap.min_node = node
cdef void link(FibonacciHeap* heap, FibonacciNode* node):
# Assumptions: - heap is a valid pointer
# - node is a valid pointer
# - node is already within heap
cdef FibonacciNode *linknode
cdef FibonacciNode *parent
cdef FibonacciNode *child
if heap.roots_by_rank[node.rank] == NULL:
heap.roots_by_rank[node.rank] = node
else:
linknode = heap.roots_by_rank[node.rank]
heap.roots_by_rank[node.rank] = NULL
if node.val < linknode.val or node == heap.min_node:
remove(linknode)
add_child(node, linknode)
link(heap, node)
else:
remove(node)
add_child(linknode, node)
link(heap, linknode)
cdef FibonacciNode* remove_min(FibonacciHeap* heap):
# Assumptions: - heap is a valid pointer
# - heap.min_node is a valid pointer
cdef FibonacciNode *temp
cdef FibonacciNode *temp_right
cdef FibonacciNode *out
cdef unsigned int i
# make all min_node children into root nodes
if heap.min_node.children:
temp = leftmost_sibling(heap.min_node.children)
temp_right = NULL
while temp:
temp_right = temp.right_sibling
remove(temp)
add_sibling(heap.min_node, temp)
temp = temp_right
heap.min_node.children = NULL
# choose a root node other than min_node
temp = leftmost_sibling(heap.min_node)
if temp == heap.min_node:
if heap.min_node.right_sibling:
temp = heap.min_node.right_sibling
else:
out = heap.min_node
heap.min_node = NULL
return out
# remove min_node, and point heap to the new min
out = heap.min_node
remove(heap.min_node)
heap.min_node = temp
# re-link the heap
for i from 0 <= i < 100:
heap.roots_by_rank[i] = NULL
while temp:
if temp.val < heap.min_node.val:
heap.min_node = temp
temp_right = temp.right_sibling
link(heap, temp)
temp = temp_right
return out
######################################################################
# Debugging: Functions for printing the fibonacci heap
#
#cdef void print_node(FibonacciNode* node, int level=0):
# print '%s(%i,%i) %i' % (level*' ', node.index, node.val, node.rank)
# if node.children:
# print_node(leftmost_sibling(node.children), level+1)
# if node.right_sibling:
# print_node(node.right_sibling, level)
#
#
#cdef void print_heap(FibonacciHeap* heap):
# print "---------------------------------"
# if heap.min_node:
# print_node(leftmost_sibling(heap.min_node))
# else:
# print "[empty heap]"
@cython.boundscheck(False)
cdef void dijkstra_directed_one_row(
unsigned int i_node,
np.ndarray[ITYPE_t, ndim=1, mode='c'] neighbors,
np.ndarray[DTYPE_t, ndim=1, mode='c'] distances,
np.ndarray[ITYPE_t, ndim=1, mode='c'] indptr,
np.ndarray[DTYPE_t, ndim=2, mode='c'] graph,
FibonacciHeap* heap,
FibonacciNode* nodes):
"""
Calculate distances from a single point to all targets using a
directed graph.
Parameters
----------
i_node : index of source point
neighbors : array, shape = [N,]
indices of neighbors for each point
distances : array, shape = [N,]
lengths of edges to each neighbor
indptr : array, shape = (N+1,)
the neighbors of point i are given by
neighbors[indptr[i]:indptr[i+1]]
graph : array, shape = (N,N)
on return, graph[i_node] contains the path lengths from
i_node to each target
heap : the Fibonacci heap object to use
nodes : the array of nodes to use
"""
cdef unsigned int N = graph.shape[0]
cdef unsigned int i
cdef FibonacciNode *v
cdef FibonacciNode *current_neighbor
cdef DTYPE_t dist
# initialize nodes
for i from 0 <= i < N:
initialize_node(&nodes[i], i)
heap.min_node = NULL
insert_node(heap, &nodes[i_node])
while heap.min_node:
v = remove_min(heap)
v.state = 2 # 2 -> SCANNED
for i from indptr[v.index] <= i < indptr[v.index + 1]:
current_neighbor = &nodes[neighbors[i]]
if current_neighbor.state != 2: # 2 -> SCANNED
dist = distances[i]
if current_neighbor.state == 0: # 0 -> NOT_IN_HEAP
current_neighbor.state = 1 # 1 -> IN_HEAP
current_neighbor.val = v.val + dist
insert_node(heap, current_neighbor)
elif current_neighbor.val > v.val + dist:
decrease_val(heap, current_neighbor,
v.val + dist)
#v has now been scanned: add the distance to the results
graph[i_node, v.index] = v.val
@cython.boundscheck(False)
cdef void dijkstra_one_row(unsigned int i_node,
np.ndarray[ITYPE_t, ndim=1, mode='c'] neighbors1,
np.ndarray[DTYPE_t, ndim=1, mode='c'] distances1,
np.ndarray[ITYPE_t, ndim=1, mode='c'] indptr1,
np.ndarray[ITYPE_t, ndim=1, mode='c'] neighbors2,
np.ndarray[DTYPE_t, ndim=1, mode='c'] distances2,
np.ndarray[ITYPE_t, ndim=1, mode='c'] indptr2,
np.ndarray[DTYPE_t, ndim=2, mode='c'] graph,
FibonacciHeap* heap,
FibonacciNode* nodes):
"""
Calculate distances from a single point to all targets using an
undirected graph.
Parameters
----------
i_node : index of source point
neighbors[1,2] : array, shape = [N,]
indices of neighbors for each point
distances[1,2] : array, shape = [N,]
lengths of edges to each neighbor
indptr[1,2] : array, shape = (N+1,)
the neighbors of point i are given by
neighbors1[indptr1[i]:indptr1[i+1]] and
neighbors2[indptr2[i]:indptr2[i+1]]
graph : array, shape = (N,)
on return, graph[i_node] contains the path lengths from
i_node to each target
heap : the Fibonacci heap object to use
nodes : the array of nodes to use
"""
cdef unsigned int N = graph.shape[0]
cdef unsigned int i
cdef FibonacciNode *v
cdef FibonacciNode *current_neighbor
cdef DTYPE_t dist
# re-initialize nodes
# children, parent, left_sibling, right_sibling should already be NULL
# rank should already be 0, index will already be set
# we just need to re-set state and val
for i from 0 <= i < N:
nodes[i].state = 0 # 0 -> NOT_IN_HEAP
nodes[i].val = 0
insert_node(heap, &nodes[i_node])
while heap.min_node:
v = remove_min(heap)
v.state = 2 # 2 -> SCANNED
for i from indptr1[v.index] <= i < indptr1[v.index + 1]:
current_neighbor = &nodes[neighbors1[i]]
if current_neighbor.state != 2: # 2 -> SCANNED
dist = distances1[i]
if current_neighbor.state == 0: # 0 -> NOT_IN_HEAP
current_neighbor.state = 1 # 1 -> IN_HEAP
current_neighbor.val = v.val + dist
insert_node(heap, current_neighbor)
elif current_neighbor.val > v.val + dist:
decrease_val(heap, current_neighbor,
v.val + dist)
for i from indptr2[v.index] <= i < indptr2[v.index + 1]:
current_neighbor = &nodes[neighbors2[i]]
if current_neighbor.state != 2: # 2 -> SCANNED
dist = distances2[i]
if current_neighbor.state == 0: # 0 -> NOT_IN_HEAP
current_neighbor.state = 1 # 1 -> IN_HEAP
current_neighbor.val = v.val + dist
insert_node(heap, current_neighbor)
elif current_neighbor.val > v.val + dist:
decrease_val(heap, current_neighbor,
v.val + dist)
#v has now been scanned: add the distance to the results
graph[i_node, v.index] = v.val