[Q] What is the excact output ConstantOfShape with empty input shape

The definition of ConstantOfShape contains a notable special case:

input: 1D tensor. The shape of the expected output tensor. If empty tensor is given, the output would be a scalar

However the test case for this is missing. I also couldn’t find any other notion how a “scalar” tensor is defined.
So what are the dimensions of the output tensor in this case? To put it in code: Is the following supposed to pass:

graph = self._make_graph([],
        [make_node("Constant", [], ['shape'],
                   value=make_tensor('shape', TensorProto.INT64, (0,), [])), # Empty tensor?
         make_node("ConstantOfShape", ['shape'], ['y'], value=make_tensor('value', TensorProto.INT32, (1, ), (2, )))],
        [])
self._assert_inferred(graph,
        [make_tensor_value_info('shape', TensorProto.INT64, (0,)),
make_tensor_value_info('y', TensorProto.INT32, [])]) # scalar tensor
# Runtime output would be the scalar value "2" in a tensor of an empty shape(?)

[Flamefire]

While I agree with the definition of @yuslepukhin (see my example test code) this is nowhere documented. So as documentation about “scalar tensor” and a test case for this special case is missing I wouldn’t close this until that is done.

[stale[bot]]

Is this still relevant? If so, what is blocking it? Is there anything you can do to help move it forward?

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.