scikit-learn/sklearn/utils/sparsetools/_graph_tools.pyx

461 lines
15 KiB
Cython

"""
Tools and utilities for working with compressed sparse graphs
"""
# Author: Jake Vanderplas -- <vanderplas@astro.washington.edu>
# License: BSD, (C) 2012
import numpy as np
cimport numpy as np
from scipy.sparse import csr_matrix, isspmatrix,\
isspmatrix_csr, isspmatrix_csc, isspmatrix_lil
DTYPE = np.float64
ctypedef np.float64_t DTYPE_t
ITYPE = np.int32
ctypedef np.int32_t ITYPE_t
# EPS is the precision of DTYPE
cdef DTYPE_t DTYPE_EPS = 1E-15
# NULL_IDX is the index used in predecessor matrices to store a non-path
cdef ITYPE_t NULL_IDX = -9999
def csgraph_from_masked(graph):
"""
csgraph_from_masked(graph)
Construct a CSR-format graph from a masked array.
.. versionadded:: 0.11.0
Parameters
----------
graph : MaskedArray
Input graph. Shape should be (n_nodes, n_nodes).
Returns
-------
csgraph : csr_matrix
Compressed sparse representation of graph,
"""
# check that graph is a square matrix
graph = np.ma.asarray(graph)
if graph.ndim != 2:
raise ValueError("graph should have two dimensions")
N = graph.shape[0]
if graph.shape[1] != N:
raise ValueError("graph should be a square array")
# construct the csr matrix using graph and mask
if np.ma.is_masked(graph):
data = graph.compressed()
mask = ~graph.mask
else:
data = graph.data
mask = np.ones(graph.shape, dtype='bool')
data = np.asarray(data, dtype=DTYPE, order='c')
idx_grid = np.empty((N, N), dtype=ITYPE)
idx_grid[:] = np.arange(N, dtype=ITYPE)
indices = np.asarray(idx_grid[mask], dtype=ITYPE, order='c')
indptr = np.zeros(N + 1, dtype=ITYPE)
indptr[1:] = mask.sum(1).cumsum()
return csr_matrix((data, indices, indptr), (N, N))
def csgraph_masked_from_dense(graph,
null_value=0,
nan_null=True,
infinity_null=True,
copy=True):
"""
csgraph_masked_from_dense(graph, null_value=0, nan_null=True,
infinity_null=True, copy=True)
Construct a masked array graph representation from a dense matrix.
.. versionadded:: 0.11.0
Parameters
----------
graph : array_like
Input graph. Shape should be (n_nodes, n_nodes).
null_value : float or None (optional)
Value that denotes non-edges in the graph. Default is zero.
infinity_null : bool
If True (default), then infinite entries (both positive and negative)
are treated as null edges.
nan_null : bool
If True (default), then NaN entries are treated as non-edges
Returns
-------
csgraph : MaskedArray
masked array representation of graph
"""
graph = np.array(graph, copy=copy)
# check that graph is a square matrix
if graph.ndim != 2:
raise ValueError("graph should have two dimensions")
N = graph.shape[0]
if graph.shape[1] != N:
raise ValueError("graph should be a square array")
# check whether null_value is infinity or NaN
if null_value is not None:
null_value = DTYPE(null_value)
if np.isnan(null_value):
nan_null = True
null_value = None
elif np.isinf(null_value):
infinity_null = True
null_value = None
# flag all the null edges
if null_value is None:
mask = np.zeros(graph.shape, dtype='bool')
graph = np.ma.masked_array(graph, mask, copy=False)
else:
graph = np.ma.masked_values(graph, null_value, copy=False)
if infinity_null:
graph.mask |= np.isinf(graph)
if nan_null:
graph.mask |= np.isnan(graph)
return graph
def csgraph_from_dense(graph,
null_value=0,
nan_null=True,
infinity_null=True):
"""
csgraph_from_dense(graph, null_value=0, nan_null=True, infinity_null=True)
Construct a CSR-format sparse graph from a dense matrix.
.. versionadded:: 0.11.0
Parameters
----------
graph : array_like
Input graph. Shape should be (n_nodes, n_nodes).
null_value : float or None (optional)
Value that denotes non-edges in the graph. Default is zero.
infinity_null : bool
If True (default), then infinite entries (both positive and negative)
are treated as null edges.
nan_null : bool
If True (default), then NaN entries are treated as non-edges
Returns
-------
csgraph : csr_matrix
Compressed sparse representation of graph,
"""
return csgraph_from_masked(csgraph_masked_from_dense(graph,
null_value,
nan_null,
infinity_null))
def csgraph_to_dense(csgraph, null_value=0):
"""
csgraph_to_dense(csgraph, null_value=0)
Convert a sparse graph representation to a dense representation
.. versionadded:: 0.11.0
Parameters
----------
csgraph : csr_matrix, csc_matrix, or lil_matrix
Sparse representation of a graph.
null_value : float, optional
The value used to indicate null edges in the dense representation.
Default is 0.
Returns
-------
graph : ndarray
The dense representation of the sparse graph.
Notes
-----
For normal sparse graph representations, calling csgraph_to_dense with
null_value=0 produces an equivalent result to using dense format
conversions in the main sparse package. When the sparse representations
have repeated values, however, the results will differ. The tools in
scipy.sparse will add repeating values to obtain a final value. This
function will select the minimum among repeating values to obtain a
final value. For example, here we'll create a two-node directed sparse
graph with multiple edges from node 0 to node 1, of weights 2 and 3.
This illustrates the difference in behavior:
>>> from scipy.sparse import csr_matrix
>>> data = np.array([2, 3])
>>> indices = np.array([1, 1])
>>> indptr = np.array([0, 2, 2])
>>> M = csr_matrix((data, indices, indptr), shape=(2, 2))
>>> M.toarray()
array([[0, 5],
[0, 0]])
>>> csgraph_to_dense(M)
array([[0, 2],
[0, 0]])
The reason for this difference is to allow a compressed sparse graph to
represent multiple edges between any two nodes. As most sparse graph
algorithms are concerned with the single lowest-cost edge between any
two nodes, the default scipy.sparse behavior of summming multiple weights
does not make sense in this context.
The other reason for using this routine is to allow for graphs with
zero-weight edges. Let's look at the example of a two-node directed
graph, connected by an edge of weight zero:
>>> from scipy.sparse import csr_matrix
>>> data = np.array([0.0])
>>> indices = np.array([1])
>>> indptr = np.array([0, 2, 2])
>>> M = csr_matrix((data, indices, indptr), shape=(2, 2))
>>> M.toarray()
array([[0, 0],
[0, 0]])
>>> csgraph_to_dense(M, np.inf)
array([[ Inf, 0.],
[ Inf, Inf]])
In the first case, the zero-weight edge gets lost in the dense
representation. In the second case, we can choose a different null value
and see the true form of the graph.
"""
# Allow only csr, lil and csc matrices: other formats when converted to csr
# combine duplicated edges: we don't want this to happen in the background.
if isspmatrix_csc(csgraph) or isspmatrix_lil(csgraph):
csgraph = csgraph.tocsr()
elif not isspmatrix_csr(csgraph):
raise ValueError("csgraph must be lil, csr, or csc format")
N = csgraph.shape[0]
if csgraph.shape[1] != N:
raise ValueError('csgraph should be a square matrix')
# get attribute arrays
data = np.asarray(csgraph.data, dtype=DTYPE, order='C')
indices = np.asarray(csgraph.indices, dtype=ITYPE, order='C')
indptr = np.asarray(csgraph.indptr, dtype=ITYPE, order='C')
# create the output array
graph = np.empty(csgraph.shape, dtype=DTYPE)
graph.fill(np.inf)
_populate_graph(data, indices, indptr, graph, null_value)
return graph
def csgraph_to_masked(csgraph):
"""
csgraph_to_masked(csgraph)
Convert a sparse graph representation to a masked array representation
.. versionadded:: 0.11.0
Parameters
----------
csgraph : csr_matrix, csc_matrix, or lil_matrix
Sparse representation of a graph.
Returns
-------
graph : MaskedArray
The masked dense representation of the sparse graph.
"""
return np.ma.masked_invalid(csgraph_to_dense(csgraph, np.nan))
cdef void _populate_graph(np.ndarray[DTYPE_t, ndim=1, mode='c'] data,
np.ndarray[ITYPE_t, ndim=1, mode='c'] indices,
np.ndarray[ITYPE_t, ndim=1, mode='c'] indptr,
np.ndarray[DTYPE_t, ndim=2, mode='c'] graph,
DTYPE_t null_value):
# data, indices, indptr are the csr attributes of the sparse input.
# on input, graph should be filled with infinities, and should be
# of size [N, N], which is also the size of the sparse matrix
cdef unsigned int N = graph.shape[0]
cdef np.ndarray null_flag = np.ones((N, N), dtype=bool, order='C')
cdef np.npy_bool* null_ptr = <np.npy_bool*> null_flag.data
cdef unsigned int row, col, i
for row from 0 <= row < N:
for i from indptr[row] <= i < indptr[row + 1]:
col = indices[i]
null_ptr[col] = 0
# in case of multiple edges, we'll choose the smallest
if data[i] < graph[row, col]:
graph[row, col] = data[i]
null_ptr += N
graph[null_flag] = null_value
def reconstruct_path(csgraph, predecessors, directed=True):
"""
reconstruct_path(csgraph, predecessors, directed=True)
Construct a tree from a graph and a predecessor list.
.. versionadded:: 0.11.0
Parameters
----------
csgraph : array_like or sparse matrix
The N x N matrix representing the directed or undirected graph
from which the predecessors are drawn.
predecessors : array_like, one dimension
The length-N array of indices of predecessors for the tree. The
index of the parent of node i is given by predecessors[i].
directed : bool, optional
If True (default), then operate on a directed graph: only move from
point i to point j along paths csgraph[i, j].
If False, then operate on an undirected graph: the algorithm can
progress from point i to j along csgraph[i, j] or csgraph[j, i].
Returns
-------
cstree : csr matrix
The N x N directed compressed-sparse representation of the tree drawn
from csgraph which is encoded by the predecessor list.
"""
from _validation import validate_graph
csgraph = validate_graph(csgraph, directed, dense_output=False)
N = csgraph.shape[0]
nnull = (predecessors < 0).sum()
indices = np.argsort(predecessors)[nnull:].astype(ITYPE)
pind = predecessors[indices]
indptr = pind.searchsorted(np.arange(N + 1)).astype(ITYPE)
if directed == True:
data = csgraph[pind, indices]
else:
data1 = csgraph[pind, indices]
data2 = csgraph[indices, pind]
data1[data1 == 0] = np.inf
data2[data2 == 0] = np.inf
data = np.minimum(data1, data2)
data = np.asarray(data).ravel()
return csr_matrix((data, indices, indptr), shape=(N, N))
def construct_dist_matrix(graph,
predecessors,
directed=True,
null_value=np.inf):
"""
construct_dist_matrix(graph, predecessors, directed=True, null_value=np.inf)
Construct distance matrix from a predecessor matrix
.. versionadded:: 0.11.0
Parameters
----------
graph : array_like or sparse
The N x N matrix representation of a directed or undirected graph.
If dense, then non-edges are indicated by zeros or infinities.
predecessors : array_like
The N x N matrix of predecessors of each node (see Notes below).
directed : bool, optional
If True (default), then operate on a directed graph: only move from
point i to point j along paths csgraph[i, j].
If False, then operate on an undirected graph: the algorithm can
progress from point i to j along csgraph[i, j] or csgraph[j, i].
null_value : bool, optional
value to use for distances between unconnected nodes. Default is
np.inf
Returns
-------
dist_matrix : ndarray
The N x N matrix of distances between nodes along the path specified
by the predecessor matrix. If no path exists, the distance is zero.
Notes
-----
The predecessor matrix is of the form returned by
:func:`graph_shortest_path`. Row i of the predecessor matrix contains
information on the shortest paths from point i: each entry
predecessors[i, j] gives the index of the previous node in the path from
point i to point j. If no path exists between point i and j, then
predecessors[i, j] = -9999
"""
from _validation import validate_graph
graph = validate_graph(graph, directed, dtype=DTYPE,
csr_output=False,
copy_if_dense=not directed)
predecessors = np.asarray(predecessors)
if predecessors.shape != graph.shape:
raise ValueError("graph and predecessors must have the same shape")
dist_matrix = np.zeros(graph.shape, dtype=DTYPE)
_construct_dist_matrix(graph, predecessors, dist_matrix,
directed, null_value)
return dist_matrix
cdef void _construct_dist_matrix(np.ndarray[DTYPE_t, ndim=2] graph,
np.ndarray[ITYPE_t, ndim=2] pred,
np.ndarray[DTYPE_t, ndim=2] dist,
int directed,
DTYPE_t null_value):
# All matrices should be size N x N
# note that graph will be modified if directed == False
# dist should be all zero on entry
global NULL_IDX
cdef int i, j, k1, k2, N, null_path
N = graph.shape[0]
#------------------------------------------
# symmetrize matrix if necessary
if not directed:
graph[graph == 0] = np.inf
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]
#------------------------------------------
for i from 0 <= i < N:
for j from 0 <= j < N:
null_path = True
k2 = j
while k2 != i:
k1 = pred[i, k2]
if k1 == NULL_IDX:
break
dist[i, j] += graph[k1, k2]
null_path = False
k2 = k1
if null_path and i != j:
dist[i, j] = null_value