This post links to the workaround using a dummy tensor and this post gives an example implementation. Is this not working for you?