| # Note [Callback for nested messages] |
| # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| # nanopb's default translation for a nested, non-repeated (possibly |
| # optional) message is to include it *inline* (no indirection), with |
| # a boolean has_g/has_t field to indicate its presence or not. Why |
| # do we not like this? It's not compatible with our ownership model, |
| # where a TensorProto/GraphProto class owns the protobuf struct it |
| # is constructing. With the default translation, the protobuf struct |
| # occurs in two places: a TensorProto, AND the parent protobuf struct |
| # field. That's bad. Turning it back into a callback solves the |
| # ownership problem. |
| # |
| # Two more bonuses: at the cost of an indirection, we no longer waste fields |
| # when we aren't actually storing a graph/tensor; furthermore, circular |
| # dependencies now work! |
| |
| onnx.AttributeProto.g type:FT_CALLBACK |
| onnx.AttributeProto.t type:FT_CALLBACK |
| onnx.ModelProto.graph type:FT_CALLBACK |
| onnx.TypeProto.TensorTypeProto.shape type:FT_CALLBACK |
| onnx.TypeProto.tensor_type type:FT_CALLBACK |
| onnx.ValueInfoProto.type type:FT_CALLBACK |
| onnx.TypeProto.TensorShapeProto.Dimension no_unions:true |
| onnx.TypeProto no_unions:true |
