diff --git a/dali/python/nvidia/dali/data_node.py b/dali/python/nvidia/dali/data_node.py
index ebd8312dbb..3892a1ce5c 100644
--- a/dali/python/nvidia/dali/data_node.py
+++ b/dali/python/nvidia/dali/data_node.py
@@ -81,6 +81,7 @@ def gpu(self) -> DataNode:
         return self._to_backend("gpu")
 
     def cpu(self) -> DataNode:
+        self._check_gpu2cpu()
         return self._to_backend("cpu")
 
     # Note: Regardless of whether we want the cpu or gpu version
@@ -259,6 +260,32 @@ def process_index(idx, dim):
         else:
             return nvidia.dali.fn.expand_dims(sliced, axes=new_axes, new_axis_names=new_axis_names)
 
+    def shape(self, *, dtype=None, device="cpu"):
+        """Returns the run-time shapes of this DataNode as a new DataNode
+
+        Parameters
+        ----------
+        arg_dtype : DALIDataType, optional
+            If specified, the shape will be converted to this data type; defaults to INT64.
+        device : str, optional
+            The device ("cpu" or "gpu") where the result is returned; defaults to CPU.
+        """
+        from . import fn
+
+        if device == "cpu":
+            self._check_gpu2cpu()
+        return fn.shapes(self, dtype=dtype, device=device)
+
+    def _check_gpu2cpu(self):
+        if self.device == "gpu" and self.source and self.source.pipeline:
+            if not self.source.pipeline._exec_dynamic:
+                raise RuntimeError(
+                    "This pipeline doesn't support transition from GPU to CPU.\n"
+                    'To enable GPU->CPU transitions, use the experimental "dynamic" executor.\n'
+                    "Specify experimental_exec_dynamic=True in your Pipeline constructor or "
+                    "@pipeline_def."
+                )
+
 
 not_iterable(DataNode)
 
diff --git a/dali/test/python/test_pipeline.py b/dali/test/python/test_pipeline.py
index 0a6456666e..a88473a1d9 100644
--- a/dali/test/python/test_pipeline.py
+++ b/dali/test/python/test_pipeline.py
@@ -2257,18 +2257,24 @@ def pdef():
         enc, _ = fn.readers.file(file_root=jpeg_folder)
         img = fn.decoders.image(enc, device="mixed")
         peek = fn.peek_image_shape(enc)
-        return peek, fn.shapes(img, device="cpu"), fn.shapes(img.cpu())
+        shapes_of_gpu = fn.shapes(img, device="cpu")
+        shapes_of_cpu = fn.shapes(img.cpu())
+        return peek, shapes_of_gpu, shapes_of_cpu, img.shape(), img.cpu().shape()
 
     pipe = pdef()
     pipe.build()
     for i in range(10):
-        peek, shape_of_gpu, shape_of_cpu = pipe.run()
+        peek, shape_of_gpu, shape_of_cpu, shape_func_gpu, shape_func_cpu = pipe.run()
         # all results must be CPU tensor lists
         assert isinstance(peek, dali.backend_impl.TensorListCPU)
         assert isinstance(shape_of_gpu, dali.backend_impl.TensorListCPU)
         assert isinstance(shape_of_cpu, dali.backend_impl.TensorListCPU)
+        assert isinstance(shape_func_gpu, dali.backend_impl.TensorListCPU)
+        assert isinstance(shape_func_cpu, dali.backend_impl.TensorListCPU)
         check_batch(shape_of_gpu, peek, bs, 0, 0)
         check_batch(shape_of_cpu, peek, bs, 0, 0)
+        check_batch(shape_func_gpu, peek, bs, 0, 0)
+        check_batch(shape_func_cpu, peek, bs, 0, 0)
 
 
 def test_gpu2cpu_old_exec_error():
@@ -2282,11 +2288,15 @@ def test_gpu2cpu_old_exec_error():
         exec_pipelined=False,
         experimental_exec_dynamic=False,
     )
-    def pdef():
+    def pdef(to_cpu):
         gpu = fn.external_source("input", device="gpu")
-        return gpu.cpu()
+        return to_cpu(gpu)
+
+    with assert_raises(RuntimeError, glob="doesn't support transition from GPU to CPU"):
+        _ = pdef(lambda gpu: gpu.cpu())  # this will raise an error at construction time
+
+    pipe = pdef(lambda gpu: gpu._to_backend("cpu"))  # this will not raise errors until build-time
 
-    pipe = pdef()
     with assert_raises(RuntimeError, glob="doesn't support transition from GPU to CPU"):
         pipe.build()