Improve static shape of shaped xtensor Ops

pytensor/tensor/basic.py

@@ -29,13 +29,13 @@ from pytensor.graph.fg import FunctionGraph, Output
 from pytensor.graph.op import Op
 from pytensor.graph.replace import _vectorize_node
 from pytensor.graph.rewriting.db import EquilibriumDB
-from pytensor.graph.type import HasShape
+from pytensor.graph.type import HasDataType, HasShape
 from pytensor.link.c.op import COp
 from pytensor.link.c.params_type import ParamsType
 from pytensor.printing import Printer, min_informative_str, pprint, set_precedence
 from pytensor.raise_op import CheckAndRaise
 from pytensor.scalar import int32
-from pytensor.scalar.basic import ScalarConstant, ScalarType, ScalarVariable
+from pytensor.scalar.basic import ScalarConstant, ScalarVariable
 from pytensor.tensor import (
     _as_tensor_variable,
     _get_vector_length,
@@ -292,13 +292,8 @@ def _get_underlying_scalar_constant_value(
     max_recur : int
         The maximum number of recursion.
 
-    Notes
-    -----
-        There may be another function similar to this one in the code,
-        but I'm not sure where it is.
-
     """
-    from pytensor.compile.ops import DeepCopyOp, OutputGuard
+    from pytensor.compile.ops import DeepCopyOp, OutputGuard, TypeCastingOp
     from pytensor.sparse import CSM
     from pytensor.tensor.subtensor import Subtensor
 
@@ -319,13 +314,20 @@ def _get_underlying_scalar_constant_value(
                 raise NotScalarConstantError()
 
         if isinstance(v, Constant):
-            if isinstance(v.type, TensorType) and v.unique_value is not None:
-                return v.unique_value
+            v_type = v.type
+            if isinstance(v_type, HasShape) and isinstance(v_type, HasDataType):
+                if v_type.ndim == 0:
+                    return np.array(v.data, dtype=v.type.dtype)
 
-            elif isinstance(v.type, ScalarType):
-                return v.data
+                elif (not any(s is None for s in v_type.shape)) and (
+                    np.prod(v_type.shape) == 1
+                ):
+                    return np.array(v.data, dtype=v_type.dtype).squeeze()
 
-            elif isinstance(v.type, NoneTypeT):
+                elif isinstance(v_type, TensorType) and v.unique_value is not None:
+                    return np.array(v.unique_value)
+
+            elif isinstance(v_type, NoneTypeT):
                 return None
 
             raise NotScalarConstantError()
@@ -333,9 +335,9 @@ def _get_underlying_scalar_constant_value(
         if not only_process_constants and getattr(v, "owner", None) and max_recur > 0:
             op = v.owner.op
             max_recur -= 1
-            if isinstance(op, Alloc | DimShuffle | OutputGuard | DeepCopyOp):
-                # OutputGuard is only used in debugmode but we
-                # keep it here to avoid problems with old pickles
+            if isinstance(
+                op, Alloc | DimShuffle | TypeCastingOp | DeepCopyOp | OutputGuard
+            ):
                 v = v.owner.inputs[0]
                 continue
             elif isinstance(op, Shape_i):
@@ -343,7 +345,6 @@ def _get_underlying_scalar_constant_value(
                 inp = v.owner.inputs[0]
                 if isinstance(inp, Constant):
                     return np.asarray(np.shape(inp.data)[i])
-                # The shape of a broadcastable dimension is 1
                 if isinstance(inp.type, HasShape) and inp.type.shape[i] is not None:
                     return np.asarray(inp.type.shape[i])
 
@@ -600,7 +601,10 @@ def get_scalar_constant_value(
     If 'v' is not a scalar, it raises a NotScalarConstantError.
 
     """
-    if isinstance(v, TensorVariable | np.ndarray):
+    if isinstance(v, Variable) and isinstance(v.type, HasShape):
+        if v.type.ndim != 0:
+            raise NotScalarConstantError("Input ndim != 0")
+    elif isinstance(v, np.ndarray):
         if v.ndim != 0:
             raise NotScalarConstantError("Input ndim != 0")
     return get_underlying_scalar_constant_value(

pytensor/tensor/subtensor.py

@@ -676,7 +676,7 @@ def get_constant_idx(
     >>> b.owner.op.idx_list
     (0, slice(1, 2, None))
     >>> get_constant_idx(b.owner.op.idx_list, b.owner.inputs, allow_partial=True)
-    [v, slice(np.int64(1), np.int64(3), None)]
+    [v, slice(1, 3, None)]
     >>> get_constant_idx(b.owner.op.idx_list, b.owner.inputs)
     Traceback (most recent call last):
     pytensor.tensor.exceptions.NotScalarConstantError
@@ -696,7 +696,7 @@ def get_constant_idx(
                     val,
                     only_process_constants=only_process_constants,
                     elemwise=elemwise,
-                )
+                ).item()
             except NotScalarConstantError:
                 if allow_partial:
                     return val

pytensor/xtensor/rewriting/shape.py

@@ -119,7 +119,6 @@ def lower_expand_dims(fgraph, node):
 
     # Convert inputs to tensors
     x_tensor = tensor_from_xtensor(x)
-    size_tensor = tensor_from_xtensor(size)
 
     # Get the new dimension name and position
     new_axis = 0  # Always insert at front
@@ -130,7 +129,7 @@ def lower_expand_dims(fgraph, node):
         result_tensor = expand_dims(x_tensor, new_axis)
     else:
         # Otherwise broadcast to the requested size
-        result_tensor = broadcast_to(x_tensor, (size_tensor, *x_tensor.shape))
+        result_tensor = broadcast_to(x_tensor, (size, *x_tensor.shape))
 
     # Preserve static shape information
     result_tensor = specify_shape(result_tensor, out.type.shape)

pytensor/xtensor/shape.py

@@ -123,7 +123,10 @@ class UnStack(XOp):
             raise ValueError(
                 f"Number of unstacked lengths {len(unstacked_length)} must match number of unstacked dims {len(self.unstacked_dims)}"
             )
-        unstacked_lengths = [as_tensor(length, ndim=0) for length in unstacked_length]
+        unstacked_lengths = [
+            as_tensor(length, allow_xtensor_conversion=True)
+            for length in unstacked_length
+        ]
         if not all(length.dtype in discrete_dtypes for length in unstacked_lengths):
             raise TypeError("Unstacked lengths must be discrete dtypes.")
 
@@ -441,7 +444,7 @@ class ExpandDims(XOp):
         if self.dim in x.type.dims:
             raise ValueError(f"Dimension {self.dim} already exists in {x.type.dims}")
 
-        size = as_xtensor(size, dims=())
+        size = as_tensor(size, allow_xtensor_conversion=True)
         if not (size.dtype in integer_dtypes and size.ndim == 0):
             raise ValueError(f"size should be an integer scalar, got {size.type}")
         try:

pytensor/xtensor/vectorization.py

@@ -16,6 +16,7 @@ from pytensor.graph.type import HasShape
 from pytensor.scalar import discrete_dtypes
 from pytensor.tensor import (
     TensorVariable,
+    as_tensor,
     broadcast_shape,
     broadcast_to,
     tensor,
@@ -232,7 +233,7 @@ class XRV(XOp, RNGConsumerOp):
             )
 
         extra_dim_lengths = [
-            as_xtensor(dim_length).values
+            as_tensor(dim_length, allow_xtensor_conversion=True)
             for dim_length in extra_dim_lengths_and_params[: len(self.extra_dims)]
         ]
         if not all(

tests/tensor/test_basic.py

@@ -3504,12 +3504,11 @@ class TestGetUnderlyingScalarConstantValue:
         assert get_underlying_scalar_constant_value(s) == c.data
 
     def test_copy(self):
-        # Make sure we do not return a writeable internal storage of a constant,
+        # Make sure we do not return the internal storage of a constant,
         # so we cannot change the value of a constant by mistake.
         c = constant(3)
         d = get_scalar_constant_value(c)
-        with pytest.raises(ValueError, match="output array is read-only"):
-            d += 1
+        d += 1
         e = get_scalar_constant_value(c)
         assert e == 3, (c, d, e)
 

tests/xtensor/test_random.py

@@ -132,6 +132,14 @@ def test_dtype():
     assert x.type.dtype == "float32"
 
 
+def test_static_shape():
+    x = xtensor("x", dims=("a", "b"), shape=(1, None))
+    y = xtensor("y", dims=("c", "d"), shape=(2, None))
+    out = normal(x, 1, extra_dims=y.sizes)
+    assert out.type.dims == ("c", "d", "a", "b")
+    assert out.type.shape == (2, None, 1, None)
+
+
 def test_normal():
     rng = random_generator_type("rng")
     c_size = tensor("c_size", shape=(), dtype=int)

tests/xtensor/test_shape.py

@@ -25,7 +25,7 @@ from pytensor.xtensor.shape import (
     unstack,
     zeros_like,
 )
-from pytensor.xtensor.type import as_xtensor, xtensor
+from pytensor.xtensor.type import XTensorType, as_xtensor, xtensor
 from pytensor.xtensor.vectorization import vectorize_graph
 from tests.xtensor.util import (
     check_vectorization,
@@ -369,16 +369,22 @@ def test_expand_dims():
 
     # Implicit size 1
     y = x.expand_dims("country")
+    assert y.type.dims == ("country", "city", "year")
+    assert y.type.shape == (1, 2, 2)
     fn = xr_function([x], y)
     xr_assert_allclose(fn(x_test), x_test.expand_dims("country"))
 
     # Test with multiple dimensions
     y = x.expand_dims(["country", "state"])
+    assert y.type.dims == ("country", "state", "city", "year")
+    assert y.type.shape == (1, 1, 2, 2)
     fn = xr_function([x], y)
     xr_assert_allclose(fn(x_test), x_test.expand_dims(["country", "state"]))
 
     # Test with a dict of name-size pairs
     y = x.expand_dims({"country": 2, "state": 3})
+    assert y.type.dims == ("country", "state", "city", "year")
+    assert y.type.shape == (2, 3, 2, 2)
     fn = xr_function([x], y)
     xr_assert_allclose(fn(x_test), x_test.expand_dims({"country": 2, "state": 3}))
 
@@ -390,6 +396,8 @@ def test_expand_dims():
     # Test with a dict of name-coord array pairs
     with pytest.warns(UserWarning, match="only its length is used"):
         y = x.expand_dims({"country": np.array([1, 2]), "state": np.array([3, 4, 5])})
+    assert y.type.dims == ("country", "state", "city", "year")
+    assert y.type.shape == (2, 3, 2, 2)
     fn = xr_function([x], y)
     xr_assert_allclose(
         fn(x_test),
@@ -399,12 +407,16 @@ def test_expand_dims():
     # Symbolic size 1
     size_sym_1 = scalar("size_sym_1", dtype="int64")
     y = x.expand_dims({"country": size_sym_1})
+    assert y.type.dims == ("country", "city", "year")
+    assert y.type.shape == (None, 2, 2)
     fn = xr_function([x, size_sym_1], y)
     xr_assert_allclose(fn(x_test, 1), x_test.expand_dims({"country": 1}))
 
     # Test with symbolic sizes in dict
     size_sym_2 = scalar("size_sym_2", dtype="int64")
     y = x.expand_dims({"country": size_sym_1, "state": size_sym_2})
+    assert y.type.dims == ("country", "state", "city", "year")
+    assert y.type.shape == (None, None, 2, 2)
     fn = xr_function([x, size_sym_1, size_sym_2], y)
     xr_assert_allclose(fn(x_test, 2, 3), x_test.expand_dims({"country": 2, "state": 3}))
 
@@ -415,16 +427,24 @@ def test_expand_dims():
 
     # Test with axis parameter
     y = x.expand_dims("country", axis=1)
+    assert y.type == XTensorType(
+        dtype=x.dtype, dims=("city", "country", "year"), shape=(2, 1, 2)
+    )
     fn = xr_function([x], y)
     xr_assert_allclose(fn(x_test), x_test.expand_dims("country", axis=1))
 
     # Test with negative axis parameter
     y = x.expand_dims("country", axis=-1)
+    assert y.type == XTensorType(
+        dtype=x.dtype, dims=("city", "year", "country"), shape=(2, 2, 1)
+    )
     fn = xr_function([x], y)
     xr_assert_allclose(fn(x_test), x_test.expand_dims("country", axis=-1))
 
     # Add two new dims with axis parameters
     y = x.expand_dims(["country", "state"], axis=[1, 2])
+    assert y.type.dims == ("city", "country", "state", "year")
+    assert y.type.shape == (2, 1, 1, 2)
     fn = xr_function([x], y)
     xr_assert_allclose(
         fn(x_test), x_test.expand_dims(["country", "state"], axis=[1, 2])