burn-onnx

name	burn-onnx
description	Burn ONNX support - Import ONNX models into Burn for inference on any backend. Supports compile-time code generation and runtime loading.

burn-onnx

Overview

burn-onnx enables importing ONNX models into Burn, allowing them to run on any Burn backend (CPU, GPU, WebAssembly).

Key Features

• Compile-time Code Generation: Convert ONNX to Rust code
• Runtime Loading: Load ONNX models at runtime
• Multi-backend Support: Run on any Burn backend
• Automatic Optimization: Benefit from Burn's optimizations

Usage

Compile-time Code Generation (Recommended)

In build.rs:

use burn_onnx::OnnxCodegen;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    OnnxCodegen::new()
        .input("model.onnx")
        .output("src/model.rs")
        .generate()?;
    
    Ok(())
}

In your code:

use burn_ndarray::NdArrayBackend;

mod model;
use model::Model;

let model: Model<NdArrayBackend> = Model::default();
let output = model.forward(input);

Runtime Loading

use burn_onnx::OnnxModel;
use burn_ndarray::NdArrayBackend;

let model = OnnxModel::<NdArrayBackend>::new("model.onnx");
let output = model.forward(input);

Multiple Models

// In build.rs
OnnxCodegen::new()
    .input("encoder.onnx")
    .output("src/encoder.rs")
    .generate()?;

OnnxCodegen::new()
    .input("decoder.onnx")
    .output("src/decoder.rs")
    .generate()?;

burn.toml Configuration

[model_exports]
default_backend = "ndarray"
default_device = "Cpu"

[ndarray]

Best Practices

1. Prefer compile-time generation for better performance
2. Use runtime loading for dynamic model loading
3. Check operator support before importing
4. Optimize models before conversion

Operator Support

burn-onnx supports common ONNX operators:

• Arithmetic: Add, Sub, Mul, Div
• Linear: MatMul, Gemm
• Activation: Relu, Sigmoid, Tanh, Softmax
• Normalization: BatchNormalization, LayerNormalization
• Convolution: Conv, ConvTranspose

Performance Comparison

Mode	Performance	Flexibility
Compile-time	High	Low
Runtime	Medium	High

When to Use

• Importing pre-trained models
• Migrating from PyTorch/TensorFlow
• Cross-platform deployment
• WebAssembly inference

Related Crates

• burn: Core framework
• burn-ndarray: CPU inference
• burn-wgpu: GPU inference