[Sparse] Support sparse matrix dividing scalar (dmlc#5173)

* use NotImplemented * format * extend to pytorch scalar * sparse div scalar * oops * Apply suggestions from code review Co-authored-by: Mufei Li <[email protected]> Co-authored-by: Mufei Li <[email protected]>
XiaoYixin7132 · Jan 13, 2023 · acc567a · acc567a
1 parent b5c5c86
commit acc567a
Show file tree

Hide file tree

Showing 3 changed files with 60 additions and 6 deletions.
diff --git a/python/dgl/sparse/elementwise_op.py b/python/dgl/sparse/elementwise_op.py
@@ -152,8 +152,8 @@ def mul(
 
 
 def div(
-    A: Union[DiagMatrix], B: Union[DiagMatrix, Scalar]
-) -> Union[DiagMatrix]:
+    A: Union[SparseMatrix, DiagMatrix], B: Union[DiagMatrix, Scalar]
+) -> Union[SparseMatrix, DiagMatrix]:
     r"""Elementwise division for ``DiagMatrix`` and ``SparseMatrix``, equivalent
     to ``A / B``.
 
@@ -164,15 +164,15 @@ def div(
     +--------------+------------+--------------+--------+
     |  DiagMatrix  |     ✅     |      🚫      |   ✅   |
     +--------------+------------+--------------+--------+
-    | SparseMatrix |     🚫     |      🚫      |   🚫   |
+    | SparseMatrix |     🚫     |      🚫      |   ✅   |
     +--------------+------------+--------------+--------+
     |    scalar    |     🚫     |      🚫      |   🚫   |
     +--------------+------------+--------------+--------+
 
     Parameters
     ----------
-    A : DiagMatrix
-        Diagonal matrix
+    A : SparseMatrix or DiagMatrix
+        Sparse or diagonal matrix
     B : DiagMatrix or Scalar
         Diagonal matrix or scalar value
 
@@ -193,6 +193,16 @@ def div(
     >>> div(A, 2)
     DiagMatrix(val=tensor([0.5000, 1.0000, 1.5000]),
                shape=(3, 3))
+
+    >>> row = torch.tensor([1, 0, 2])
+    >>> col = torch.tensor([0, 3, 2])
+    >>> val = torch.tensor([1, 2, 3])
+    >>> A = from_coo(row, col, val, shape=(3, 4))
+    >>> A / 2
+    SparseMatrix(indices=tensor([[1, 0, 2],
+                                 [0, 3, 2]]),
+                 values=tensor([0.5000, 1.0000, 1.5000]),
+                 shape=(3, 4), nnz=3)
     """
     return A / B
 

diff --git a/python/dgl/sparse/elementwise_op_sp.py b/python/dgl/sparse/elementwise_op_sp.py
@@ -89,6 +89,40 @@ def sp_mul(A: SparseMatrix, B: Scalar) -> SparseMatrix:
     return NotImplemented
 
 
+def sp_div(A: SparseMatrix, B: Scalar) -> SparseMatrix:
+    """Elementwise division
+
+    Parameters
+    ----------
+    A : SparseMatrix
+        First operand
+    B : Scalar
+        Second operand
+
+    Returns
+    -------
+    SparseMatrix
+        Result of A / B
+
+    Examples
+    --------
+    >>> row = torch.tensor([1, 0, 2])
+    >>> col = torch.tensor([0, 3, 2])
+    >>> val = torch.tensor([1, 2, 3])
+    >>> A = from_coo(row, col, val, shape=(3, 4))
+    >>> A / 2
+    SparseMatrix(indices=tensor([[1, 0, 2],
+                                 [0, 3, 2]]),
+                 values=tensor([0.5000, 1.0000, 1.5000]),
+                 shape=(3, 4), nnz=3)
+    """
+    if is_scalar(B):
+        return val_like(A, A.val / B)
+    # Python falls back to B.__rtruediv__(A) then TypeError when NotImplemented
+    # is returned.
+    return NotImplemented
+
+
 def sp_power(A: SparseMatrix, scalar: Scalar) -> SparseMatrix:
     """Take the power of each nonzero element and return a sparse matrix with
     the result.
@@ -125,4 +159,5 @@ def sp_power(A: SparseMatrix, scalar: Scalar) -> SparseMatrix:
 SparseMatrix.__add__ = sp_add
 SparseMatrix.__mul__ = sp_mul
 SparseMatrix.__rmul__ = sp_mul
+SparseMatrix.__truediv__ = sp_div
 SparseMatrix.__pow__ = sp_power
diff --git a/tests/pytorch/sparse/test_elementwise_op_sp.py b/tests/pytorch/sparse/test_elementwise_op_sp.py
@@ -23,7 +23,7 @@ def all_close_sparse(A, row, col, val, shape):
 @pytest.mark.parametrize(
     "v_scalar", [2, 2.5, torch.tensor(2), torch.tensor(2.5)]
 )
-def test_mul_scalar(v_scalar):
+def test_muldiv_scalar(v_scalar):
     ctx = F.ctx()
     row = torch.tensor([1, 0, 2]).to(ctx)
     col = torch.tensor([0, 3, 2]).to(ctx)
@@ -40,6 +40,15 @@ def test_mul_scalar(v_scalar):
     assert torch.allclose(A1.val * v_scalar, A2.val, rtol=1e-4, atol=1e-4)
     assert A1.shape == A2.shape
 
+    # A / v
+    A2 = A1 / v_scalar
+    assert torch.allclose(A1.val / v_scalar, A2.val, rtol=1e-4, atol=1e-4)
+    assert A1.shape == A2.shape
+
+    # v / A
+    with pytest.raises(TypeError):
+        v_scalar / A1
+
 
 @pytest.mark.parametrize("val_shape", [(3,), (3, 2)])
 def test_pow(val_shape):