Implement ufunc_outer like add.outer for binary Elemwise operations

Author: Dhruvanshu-Joshi

pytensor/tensor/elemwise.py

@@ -1170,6 +1170,16 @@ def c_code_cache_version_apply(self, node):
         else:
             return ()
 
+    def outer(self, x, y):
+        from pytensor.tensor.basic import expand_dims
+
+        if self.scalar_op.nin not in (-1, 2):
+            raise NotImplementedError("outer is only available for binary operators")
+
+        x_ = expand_dims(x, tuple(range(-y.ndim, 0)))
+        y_ = expand_dims(y, tuple(range(x.ndim)))
+        return self(x_, y_)
+
 
 class CAReduce(COp):
     """Reduces a scalar operation along specified axes.

tests/tensor/test_elemwise.py

@@ -8,7 +8,9 @@
 
 import pytensor
 import pytensor.scalar as ps
+import pytensor.tensor as pt
 import tests.unittest_tools as utt
+from pytensor.compile.function import function
 from pytensor.compile.mode import Mode
 from pytensor.configdefaults import config
 from pytensor.graph.basic import Apply, Variable
@@ -893,6 +895,25 @@ def test_invalid_static_shape(self):
         ):
             x + y
 
+    @pytest.mark.parametrize(
+        "shape_x, shape_y, op, np_op",
+        [
+            ((3, 5), (7, 1, 3), pt.add, np.add),
+            ((2, 3), (1, 4), pt.mul, np.multiply),
+        ],
+    )
+    def test_outer(self, shape_x, shape_y, op, np_op):
+        x = tensor(dtype=np.float64, shape=shape_x)
+        y = tensor(dtype=np.float64, shape=shape_y)
+
+        z = op.outer(x, y)
+
+        f = function([x, y], z)
+        x1 = np.ones(shape_x)
+        y1 = np.ones(shape_y)
+
+        np.testing.assert_array_equal(f(x1, y1), np_op.outer(x1, y1))
+
 
 def test_not_implemented_elemwise_grad():
     # Regression test for unimplemented gradient in an Elemwise Op.