Crate onnx_extractor 

onnx-extractor

A minimal ONNX model loader for extracting weights, tensors, operations, and graph structure.

This crate provides a simple interface to load ONNX models and extract:

• Tensor weights & raw data (zero-copy slices, owned buffers, external files)
• Tensor shapes, dimensions, and data types
• Operations and attributes (nodes, inputs/outputs, subgraphs)
• Graph topology and execution order

Zero-Copy Design

Tensor::data() returns a TensorDataRef which borrows tensor data without copying:

• Raw variant uses shared ownership (Bytes)
• Numeric variants borrow directly (&[T])
• Strings variant borrows elements (&[Bytes])

Tensor::into_data() returns owned TensorData without copying:

• Raw variant returns shared ownership (Bytes)
• Numeric variants return owned vectors (Vec<T>)
• Strings variant returns owned vectors of shared elements (Vec<Bytes>)

Endianness: Raw tensor data uses little-endian byte order as defined in the ONNX specification. Typed fields use native host representation.

Quick Start

use onnx_extractor::Model;

let model = Model::load_from_file("model.onnx")?;
println!("{}", model);

// Access tensor shape and data
if let Some(tensor) = model.graph().tensors().get("weight_1") {
    println!("Weight shape: {:?}", tensor.shape());
    let data = tensor.data()?;
    if let Some(bytes) = data.as_slice() {
        println!("Data size: {} bytes", bytes.len());
    }
}

Re-exports

pub use attribute_value::AttributeValue;

pub use data_type::DataType;

pub use error::Error;

pub use graph::Graph;

pub use model::Model;

pub use operation::Operation;

pub use tensor::Tensor;

pub use tensor::TensorData;

pub use tensor::TensorDataRef;

Modules

attribute_value

data_type

error

graph

model

operation

tensor

Structs

Bytes

A cheaply cloneable and sliceable chunk of contiguous memory.