INNAPIGraphBuilder
Interfaces ยท AiDotNet.Deployment.Mobile.Android
Translates a model file (TFLite / ONNX / etc.) into an NNAPI operation
graph by adding operands and operations to a freshly-created
ANeuralNetworksModel handle. Plug an implementation into
NNAPIBackend via its constructor so LoadModel
can actually compile a real NNAPI graph instead of falling back to the
managed CPU executor.
How It Works
Each NNAPI backend instance owns its own _model IntPtr. The
builder is invoked once per LoadModel
call, after the backend has called ANeuralNetworksModel_create
but before ANeuralNetworksModel_finish. The builder is
responsible for:
Splitting graph construction out of the backend keeps the backend model-format-agnostic โ TFLite, ONNX, and bespoke IRs all plug in via the same interface.
Methods
| Method | Summary |
|---|---|
bool BuildGraph(IntPtr modelHandle, byte[] modelBytes) |
Populates the supplied modelHandle with operands and operations decoded from modelBytes. |