This requires a binary proto, which can be generated with multiple libraries (e.g.
https://crates.io/crates/protobuf,
https://crates.io/crates/prost). The Rust API currently requires a serialized protobuf because it's not yet clear which proto library would be best for the tensorflow crate and its users, so we don't force users to use a specific library.
On that topic, do people have a preference for which proto library to use? One feature which the tensorflow create could make use of but (as far as I'm aware) none of the main libraries support well is parsing text protos.