Struct OnnxModel 

pub struct OnnxModel {
    model: InferenceModel,
}

A parsed ONNX model, ready to run at any spatial size.

The graph is kept in tract’s shape-inferring form and specialized to the concrete input shapes on each run call, so one instance serves volumes of different dimensions (fully-convolutional nets).

Fields

model: InferenceModel

Implementations

impl OnnxModel

pub fn load(bytes: &[u8]) -> Result<Self, OnnxError>

Parse an ONNX model from its serialized bytes.

pub fn run(&self, inputs: &[Tensor]) -> Result<Vec<Tensor>, OnnxError>

Run the model with inputs bound to graph inputs in order; returns every graph output as an f32 Tensor.

pub fn run_single(&self, input: &Tensor) -> Result<Tensor, OnnxError>

Convenience for single-input / single-output nets.

pub fn plan_for(&self, input_shapes: &[&[usize]]) -> Result<OnnxPlan, OnnxError>

Compile a reusable execution plan specialized to fixed input shapes.

run re-clones and re-optimizes the whole graph on every call — fine for one-shot whole-volume inference, but wasteful when running many equal-shaped tensors (e.g. every patch of a tiled inversion). Build one OnnxPlan for the patch shape and reuse it: tract’s optimizer (into_optimized — constant-folding conv weights, operator fusion, plan building) then runs exactly once instead of per patch.