onnx-extractor

// Model summary
println!("{}", model);

// Get individual metadata items
let opset = model.default_opset_version(); // Option<i64>
let metadata = model.metadata(); // &HashMap<String, String>
let ir_ver = model.ir_version(); // i64

// Graph access
let graph = model.graph();

// Tensor access
let tensor = graph.tensors().get("input_name"); // Option<&Tensor>
let tensor_names = graph.tensors().keys(); // Iterator<Item = &String>
let inputs = graph.get_input_tensors(); // Iterator<Item = &Tensor>
let outputs = graph.get_output_tensors(); // Iterator<Item = &Tensor>
let weights = graph.get_weight_tensors(); // Iterator<Item = &Tensor>

// Operation access
let operation = graph.get_operation("op_name"); // Option<&Operation>
let conv_ops = graph.get_operations_by_type("Conv"); // Iterator<Item = &Operation>
let op_types = graph.operation_types(); // HashSet<&str>
let op_counts = graph.count_operations_by_type(); // HashMap<&str, usize>

// Topological order
let topo_order = graph.topological_order()?; // Vec<&Operation>

// Extracting a tensor (moves out of graph so data can outlive it)
// (Note: `model` must be declared as `mut model`)
let owned_tensor = model.graph_mut().tensors_mut().remove("weight"); // Option<Tensor>

Tensor Functions

use onnx_extractor::TensorDataRef;

let tensor = model.graph().tensors().get("weight").unwrap();

// Name, shape and data type
println!("Name: {:?}", tensor.name());
println!("Shape: {:?}", tensor.shape());
println!("Data type: {:?}", tensor.data_type());

// Check if tensor contains data (does not trigger external file loading or mmap)
let has_data = tensor.has_data(); // bool

// Borrow tensor data
let data_ref = tensor.data()?; // TensorDataRef<'_>
println!("Data size: {} bytes", data_ref.len());

// Get data as contiguous byte slice
// Returns Some(&[u8]) for Raw and Numeric variants; returns None for Strings as they are not contiguous in memory
if let Some(bytes) = data_ref.as_slice() {
    println!("Byte slice length: {}", bytes.len());
}

// Access string elements for String tensors
if let Some(strings) = data_ref.strings() {
    for s in strings {
        println!("String element: {:?}", s);
    }
}

// Access Raw data
if let TensorDataRef::Raw(bytes) = &data_ref {
    println!("Bytes length: {}", bytes.len());
}

// Consume tensor and get owned data
// This allows the data to outlive the model itself
// (Note: `model` must be declared as `mut model`)
if let Some(owned_tensor) = model.graph_mut().tensors_mut().remove("weight") {
    let owned_data = owned_tensor.into_data()?; // TensorData
}

Tensor Data Enums

The data() method returns a TensorDataRef which borrows from the model:

pub enum TensorDataRef<'a> {
    /// Contiguous buffer (mmap-backed or loaded)
    Raw(Bytes),
    /// String tensor elements (shared Bytes)
    Strings(&'a [Bytes]),
    /// Typed numeric data
    F32(&'a [f32]),
    F64(&'a [f64]),
    I32(&'a [i32]),
    I64(&'a [i64]),
    U64(&'a [u64]),
}

The into_data() method returns an owned TensorData (using Bytes or Vec storage from the protobuf):

pub enum TensorData {
    /// Contiguous buffer (mmap-backed or loaded)
    Raw(Bytes),
    /// String tensor elements (shared Bytes)
    Strings(Vec<Bytes>),
    /// Typed numeric data
    F32(Vec<f32>),
    F64(Vec<f64>),
    I32(Vec<i32>),
    I64(Vec<i64>),
    U64(Vec<u64>),
}

Operation Functions

let op = model.graph().get_operation("conv1").unwrap();

// Basic info
println!("Type: {}", op.op_type());
println!("Inputs: {:?}", op.inputs());
println!("Outputs: {:?}", op.outputs());

// Attribute access
let attributes = op.attributes();
if let Some(attr) = attributes.get("kernel_shape") {
    let kernel_size = attr.as_ints(); // Option<&[i64]>
}

let stride = attributes.get("stride").and_then(|a| a.as_int()); // Option<i64>
let activation = attributes.get("activation").and_then(|a| a.as_string()); // Option<&Bytes>
// Or for validated UTF-8:
let activation_str = attributes.get("activation").and_then(|a| a.as_string_validated().ok()); // Option<&str>

// Subgraph access (Control Flow subgraphs)
if let Some(subgraph) = attributes.get("then_branch").and_then(|a| a.as_graph()) {
    println!("Subgraph name: {:?}", subgraph.graph_name());
}

Data Types

Access the DataType enum for type checking:

use onnx_extractor::DataType;

let tensor = model.graph().tensors().get("input").unwrap();
match tensor.data_type() {
    DataType::Float => println!("32-bit float"),
    DataType::Double => println!("64-bit float"),
    DataType::Int32 => println!("32-bit int"),
    _ => println!("Other type"),
}

// Type properties
let size = tensor.data_type().size_in_bytes(); // Option<usize>
let is_float = tensor.data_type().is_float();
let is_int = tensor.data_type().is_integer();

External Data Support

ONNX models can store large tensor data in external files. This crate supports lazy loading of external data with automatic caching:

// Load model with external data files
let model = Model::load_from_file("large_model.onnx")?;

// External data files (e.g., "large_model.onnx.data") are automatically discovered
// from the model directory and loaded lazily when tensor data is accessed

let tensor = model.graph().tensors().get("large_weight").unwrap();

// Data is loaded from external file on first access and cached for subsequent use
let data = tensor.data()?;
println!("Loaded {} bytes from external file", data.len());

// Multiple tensors can share the same external file efficiently
// The file is only loaded once and cached

External Data Features

• Lazy Loading: External files are only loaded when tensor data is accessed
• Shared Caching: Multiple tensors sharing the same external file benefit from caching
• Offset & Length: Supports reading specific ranges from large external files

About the protobuf (onnx.proto)

This crate generates Rust types from the ONNX protobuf at build time using prost-build.

Platform Notes

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

License

MIT