OnnxGraphBuilder

Helpers & Utilities · AiDotNet.Onnx

Thin facade over the vendored ONNX protobuf types that lets layer converters add nodes, initializers, inputs, and outputs to a model graph without touching the generated GraphProto directly.

One OnnxGraphBuilder is created per export. Layer converters call AddNode, AddInitializer, AddInput, AddOutput, and NextTensorName as needed; the final Build assembles a ModelProto ready to write.

This class is intentionally minimal in this initial commit — layer converters landing in subsequent commits drive the API additions they need.

Constructors

Constructor Summary
OnnxGraphBuilder(OnnxExportOptions options)

Properties

Property Summary
int OpsetVersion { get; } Opset version the resulting model targets.

Methods

Method Summary
ModelProto Build() Assembles the final ModelProto with producer metadata + opset declaration.
NodeProto AddOp(string opType, string[] inputs, string[] outputs, string name) Adds an ONNX op node with named inputs and outputs and optional attributes.
string AddFloatInitializer(string namePrefix, float[] data, int[] shape) Adds a float32 constant initializer with the given shape.
string NextTensorName(string prefix) Reserves a unique tensor name with the given prefix.
void AddFloatInput(string name, int[] shape) Declares a float32 graph input with the given (fixed or symbolic) shape.
void AddFloatOutput(string name, int[] shape) Declares a float32 graph output.
void AddInitializer(TensorProto initializer) Adds a constant initializer (e.g., a layer’s weight tensor).
void AddInput(ValueInfoProto input) Adds a graph input (a tensor the model expects from the caller).
void AddNode(NodeProto node) Adds a node to the graph.
void AddOutput(ValueInfoProto output) Adds a graph output (a tensor the model produces).
void WriteTo(Stream stream) Serializes the built ModelProto to a stream.