Allow specializing shape of predefined tensor types

pytensor/tensor/type.py

@@ -123,6 +123,18 @@ def parse_bcast_and_shape(s):
         self.name = name
         self.numpy_dtype = np.dtype(self.dtype)
 
+    def __call__(self, *args, shape=None, **kwargs):
+        if shape is not None:
+            # Check if shape is compatible with the original type
+            new_type = self.clone(shape=shape)
+            if self.is_super(new_type):
+                return new_type(*args, **kwargs)
+            else:
+                raise ValueError(
+                    f"{shape=} is incompatible with original type shape {self.shape=}"
+                )
+        return super().__call__(*args, **kwargs)
+
     def clone(
         self, dtype=None, shape=None, broadcastable=None, **kwargs
     ) -> "TensorType":

tests/tensor/test_type.py

@@ -10,6 +10,10 @@
 from pytensor.tensor.type import (
     TensorType,
     col,
+    dmatrix,
+    drow,
+    fmatrix,
+    frow,
     matrix,
     row,
     scalar,
@@ -477,3 +481,21 @@ def test_row_matrix_creator_helpers(helper):
         match = "The second dimension of a `col` must have shape 1, got 5"
     with pytest.raises(ValueError, match=match):
         helper(shape=(2, 5))
+
+
+def test_shape_of_predefined_dtype_tensor():
+    # Valid: None dimensions can be specialized
+    assert fmatrix(shape=(1, None)).type == frow
+    assert drow(shape=(1, 5)).type == dmatrix(shape=(1, 5)).type
+
+    # Invalid: Number of dimensions must match
+    with pytest.raises(ValueError):
+        fmatrix(shape=(None, None, None))
+
+    # Invalid: Fixed shapes must match
+    with pytest.raises(ValueError):
+        fmatrix(shape=(3, 5)).type(shape=(4, 5))
+
+    # Invalid: Known shapes can't be lost
+    with pytest.raises(ValueError):
+        drow(shape=(None, None))