I tried to export a Relay graph to a file and import it back with mod._import(path_to_saved_graph) but it fails to parse the saved filed. I looked for a solution and I came to the conclusion that while I can dump a readable representation of the Relay graph, it is not possible to import it back to TVM with this method because of the impossibility to parse ‘?’.
I know it is possible to export the outputs of relay.build but I am trying to obtain a target agnostic graph.
So how to export a Relay graph generated from a frontend (for example from relay.frontend.from_tensorflow) and import it back to TVM to build it?
Thanks in advance for your help!