Fix conversions from non-canonical scipy.sparse arrays. (#861)

Author: hameerabbasi

sparse/numba_backend/_compressed/compressed.py

@@ -209,11 +209,14 @@ def from_coo(cls, x, compressed_axes=None, idx_dtype=None):
 
     @classmethod
     def from_scipy_sparse(cls, x, /, *, fill_value=None):
-        if x.format == "csc":
-            return cls((x.data, x.indices, x.indptr), shape=x.shape, compressed_axes=(1,), fill_value=fill_value)
-
-        x = x.asformat("csr")
-        return cls((x.data, x.indices, x.indptr), shape=x.shape, compressed_axes=(0,), fill_value=fill_value)
+        is_csc = x.format == "csc"
+        ca = (1,) if is_csc else (0,)
+        if not is_csc:
+            x = x.asformat("csr")
+        if not x.has_canonical_format:
+            x.eliminate_zeros()
+            x.sum_duplicates()
+        return cls((x.data, x.indices, x.indptr), shape=x.shape, compressed_axes=ca, fill_value=fill_value)
 
     @classmethod
     def from_iter(cls, x, shape=None, compressed_axes=None, fill_value=None, idx_dtype=None):
@@ -903,6 +906,10 @@ def __init__(self, arg, shape=None, compressed_axes=class_compressed_axes, prune
     @classmethod
     def from_scipy_sparse(cls, x, /, *, fill_value=None):
         x = x.asformat("csr", copy=False)
+        if not x.has_canonical_format:
+            x.eliminate_zeros()
+            x.sum_duplicates()
+
         return cls((x.data, x.indices, x.indptr), shape=x.shape, fill_value=fill_value)
 
     def transpose(self, axes: None = None, copy: bool = False) -> Union["CSC", "CSR"]:
@@ -935,6 +942,10 @@ def __init__(self, arg, shape=None, compressed_axes=class_compressed_axes, prune
     @classmethod
     def from_scipy_sparse(cls, x, /, *, fill_value=None):
         x = x.asformat("csc", copy=False)
+        if not x.has_canonical_format:
+            x.eliminate_zeros()
+            x.sum_duplicates()
+
         return cls((x.data, x.indices, x.indptr), shape=x.shape, fill_value=fill_value)
 
     def transpose(self, axes: None = None, copy: bool = False) -> Union["CSC", "CSR"]:

sparse/numba_backend/_coo/core.py

@@ -447,6 +447,10 @@ def from_scipy_sparse(cls, x, /, *, fill_value=None):
         True
         """
         x = x.asformat("coo")
+        if not x.has_canonical_format:
+            x.eliminate_zeros()
+            x.sum_duplicates()
+
         coords = np.empty((2, x.nnz), dtype=x.row.dtype)
         coords[0, :] = x.row
         coords[1, :] = x.col

sparse/numba_backend/tests/test_conversion.py

@@ -3,6 +3,9 @@
 
 import pytest
 
+import numpy as np
+import scipy.sparse as sps
+
 FORMATS_ND = [
     sparse.COO,
     sparse.DOK,
@@ -38,3 +41,20 @@ def test_conversion_scalar(format1, format2):
     x = sparse.random((), format=format1, fill_value=0.5)
     y = x.asformat(format2)
     assert_eq(x, y)
+
+
+def test_non_canonical_conversion():
+    """
+    Regression test for gh-602.
+
+    Adapted from https://github.com/LiberTEM/sparseconverter/blob/4cfc0ee2ad4c37b07742db8f3643bcbd858a4e85/src/sparseconverter/__init__.py#L154-L183
+    """
+    data = np.array((2.0, 1.0, 3.0, 3.0, 1.0))
+    indices = np.array((1, 0, 0, 1, 1), dtype=int)
+    indptr = np.array((0, 2, 5), dtype=int)
+
+    x = sps.csr_matrix((data, indices, indptr), shape=(2, 2))
+    ref = np.array(((1.0, 2.0), (3.0, 4.0)))
+
+    gcxs_check = sparse.GCXS(x)
+    assert np.all(gcxs_check[:1].todense() == ref[:1]) and np.all(gcxs_check[1:].todense() == ref[1:])