Merge pull request #887 from flying-sheep/pa/sparse-array

fix: support scipy.sparse.sparray

Author: ntamas

.github/workflows/build.yml

@@ -336,7 +336,7 @@ jobs:
         run: |
           # We cannot install the test dependency group because many test dependencies cause
           # false positives in the sanitizer
-          pip install --prefer-binary networkx pytest pytest-timeout
+          pip install --prefer-binary networkx pytest pytest-timeout parameterized
           pip install -e .
 
       # Only pytest, and nothing else should be run in this section due to the presence of LD_PRELOAD.

setup.py

@@ -1011,6 +1011,7 @@ def get_tag(self):
             "networkx>=2.5",
             "pytest>=7.0.1",
             "pytest-timeout>=2.1.0",
+            "parameterized",
             "numpy>=1.19.0; platform_python_implementation != 'PyPy'",
             "pandas>=1.1.0; platform_python_implementation != 'PyPy'",
             "scipy>=1.5.0; platform_python_implementation != 'PyPy'",
@@ -1039,6 +1040,7 @@ def get_tag(self):
             "networkx>=2.5",
             "pytest>=7.0.1",
             "pytest-timeout>=2.1.0",
+            "parameterized",
         ],
         # Dependencies needed for building the documentation
         "doc": [

src/igraph/adjacency.py

@@ -86,15 +86,16 @@ def _get_adjacency(
     return Matrix(data)
 
 
-def _get_adjacency_sparse(self, attribute=None):
-    """Returns the adjacency matrix of a graph as a SciPy CSR matrix.
+def _get_adjacency_sparse(self, attribute=None, *, container="matrix"):
+    """Returns the adjacency matrix of a graph as a SciPy CSR array or matrix.
 
     @param attribute: if C{None}, returns the ordinary adjacency
       matrix. When the name of a valid edge attribute is given
       here, the matrix returned will contain the default value
       at the places where there is no edge or the value of the
       given attribute where there is an edge.
-    @return: the adjacency matrix as a C{scipy.sparse.csr_matrix}.
+    @param container: either C{"array"} or C{"matrix"}
+    @return: the adjacency matrix as a C{scipy.sparse.csr_array} or C{scipy.sparse.csr_matrix}.
     """
     try:
         from scipy import sparse
@@ -103,6 +104,10 @@ def _get_adjacency_sparse(self, attribute=None):
             "You should install scipy in order to use this function"
         ) from None
 
+    if container not in {"array", "matrix"}:
+        raise ValueError("container must be either 'array' or 'matrix'")
+    cls = sparse.csr_array if container == "array" else sparse.csr_matrix
+
     edges = self.get_edgelist()
     if attribute is None:
         weights = [1] * len(edges)
@@ -114,7 +119,7 @@ def _get_adjacency_sparse(self, attribute=None):
 
     N = self.vcount()
     r, c = zip(*edges) if edges else ([], [])
-    mtx = sparse.csr_matrix((weights, (r, c)), shape=(N, N))
+    mtx = cls((weights, (r, c)), shape=(N, N))
 
     if not self.is_directed():
         mtx = mtx + sparse.triu(mtx, 1).T + sparse.tril(mtx, -1).T

src/igraph/io/adjacency.py

@@ -4,13 +4,23 @@
 )
 
 
+def _sp_cls():
+    try:
+        from scipy import sparse
+    except ImportError:
+        return ()
+    if not hasattr(sparse, "sparray"):  # scipy < 1.11
+        return sparse.spmatrix
+    return (sparse.sparray, sparse.spmatrix)
+
+
 def _construct_graph_from_adjacency(cls, matrix, mode="directed", loops="once"):
     """Generates a graph from its adjacency matrix.
 
     @param matrix: the adjacency matrix. Possible types are:
       - a list of lists
       - a numpy 2D array or matrix (will be converted to list of lists)
-      - a scipy.sparse matrix (will be converted to a COO matrix, but not
+      - a scipy.sparse array or matrix (will be converted to COO format, but not
         to a dense matrix)
       - a pandas.DataFrame (column/row names must match, and will be used
         as vertex names).
@@ -42,17 +52,12 @@ def _construct_graph_from_adjacency(cls, matrix, mode="directed", loops="once"):
     except ImportError:
         np = None
 
-    try:
-        from scipy import sparse
-    except ImportError:
-        sparse = None
-
     try:
         import pandas as pd
     except ImportError:
         pd = None
 
-    if (sparse is not None) and isinstance(matrix, sparse.spmatrix):
+    if isinstance(matrix, _sp_cls()):
         return _graph_from_sparse_matrix(cls, matrix, mode=mode, loops=loops)
 
     if (pd is not None) and isinstance(matrix, pd.DataFrame):
@@ -117,17 +122,12 @@ def _construct_graph_from_weighted_adjacency(
     except ImportError:
         np = None
 
-    try:
-        from scipy import sparse
-    except ImportError:
-        sparse = None
-
     try:
         import pandas as pd
     except ImportError:
         pd = None
 
-    if (sparse is not None) and isinstance(matrix, sparse.spmatrix):
+    if isinstance(matrix, _sp_cls()):
         return _graph_from_weighted_sparse_matrix(
             cls,
             matrix,

src/igraph/sparse_matrix.py

@@ -2,8 +2,6 @@
 # -*- coding: utf-8 -*-
 """Implementation of Python-level sparse matrix operations."""
 
-from __future__ import with_statement
-
 __all__ = ()
 __docformat__ = "restructuredtext en"
 
@@ -65,7 +63,7 @@ def _maybe_halve_diagonal(m, condition):
 # Logic to get graph from scipy sparse matrix. This would be simple if there
 # weren't so many modes.
 def _graph_from_sparse_matrix(klass, matrix, mode="directed", loops="once"):
-    """Construct graph from sparse matrix, unweighted.
+    """Construct graph from sparse array or matrix, unweighted.
 
     @param loops: specifies how the diagonal of the matrix should be handled:
 
@@ -78,7 +76,7 @@ def _graph_from_sparse_matrix(klass, matrix, mode="directed", loops="once"):
     # matrix. The caller should make sure those conditions are met.
     from scipy import sparse
 
-    if not isinstance(matrix, sparse.coo_matrix):
+    if not isinstance(matrix, (sparse.coo_matrix, *([sparse.coo_array] if hasattr(sparse, "coo_array") else []))):
         matrix = matrix.tocoo()
 
     nvert = max(matrix.shape)
@@ -150,7 +148,7 @@ def _graph_from_sparse_matrix(klass, matrix, mode="directed", loops="once"):
 def _graph_from_weighted_sparse_matrix(
     klass, matrix, mode=ADJ_DIRECTED, attr="weight", loops="once"
 ):
-    """Construct graph from sparse matrix, weighted
+    """Construct graph from sparse array or matrix, weighted
 
     NOTE: Of course, you cannot emcompass a fully general weighted multigraph
     with a single adjacency matrix, so we don't try to do it here either.
@@ -165,7 +163,7 @@ def _graph_from_weighted_sparse_matrix(
     # matrix. The caller should make sure those conditions are met.
     from scipy import sparse
 
-    if not isinstance(matrix, sparse.coo_matrix):
+    if not isinstance(matrix, (sparse.coo_matrix, *([sparse.coo_array] if hasattr(sparse, "coo_array") else []))):
         matrix = matrix.tocoo()
 
     nvert = max(matrix.shape)

tests/test_generators.py

@@ -1,5 +1,7 @@
 import unittest
 
+from parameterized import parameterized
+
 from igraph import Graph, InternalError
 
 
@@ -747,11 +749,13 @@ def testWeightedAdjacencyNumPy(self):
         self.assertListEqual(el, [(1, 0), (0, 1), (3, 1), (0, 2), (2, 2)])
         self.assertListEqual(g.es["w0"], [2, 1, 1, 2, 2.5])
 
+    # not testing csc_* here, as the edge order comes out different
+    @parameterized.expand(["coo_array", "coo_matrix", "csr_matrix", "csr_array"])
     @unittest.skipIf(
         (sparse is None) or (np is None), "test case depends on NumPy/SciPy"
     )
-    def testSparseWeightedAdjacency(self):
-        mat = sparse.coo_matrix(
+    def testSparseWeightedAdjacency(self, cls):
+        mat = getattr(sparse, cls)(
             [[0, 1, 2, 0], [2, 0, 0, 0], [0, 0, 2.5, 0], [0, 1, 0, 0]]
         )