---

name: burn-onnx description: This skill should be used when the user asks about "ONNX import", "PyTorch to Burn", "safetensors", "model conversion", "burn-import", "unsupported operator", "opset version", or importing external models into Burn. version: 0.1.0

Burn Model Import

Knowledge for importing ONNX, PyTorch, and Safetensors models into Burn.

ONNX Import Overview

Burn uses burn-import crate for ONNX model conversion at build time.

Setup

Add to Cargo.toml:

Copy[dependencies]
burn = { version = "0.16", features = ["wgpu"] }

[build-dependencies]
burn-import = "0.16"

Build Script

Create build.rs:

Copyuse burn_import::onnx::ModelGen;

fn main() {
    ModelGen::new()
        .input("models/my_model.onnx")
        .out_dir("model/")
        .run_from_script();
}

Generated Code

The build generates a Rust module with:

• Model struct matching ONNX architecture
• Weights embedded or loadable
• Type-safe forward method

Copymod model {
    include!(concat!(env!("OUT_DIR"), "/model/my_model.rs"));
}

use model::Model;

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

Operator Coverage

burn-import supports common operators. Check coverage before import:

• Fully supported: Conv, Linear, ReLU, BatchNorm, MaxPool, etc.
• Partial: Some transformer ops, dynamic shapes
• Unsupported: Custom ops, some recent opset additions

Troubleshooting

Unsupported Operator

If import fails with unsupported op:

1. Check if newer burn-import version supports it
2. Consider opset conversion (downgrade to supported version)
3. Implement manually and splice into model

Shape Mismatches

ONNX uses dynamic shapes, Burn uses static:

1. Check input shapes match expected
2. Use reshape to convert dynamic to static
3. Verify batch dimension handling

PyTorch Import

For PyTorch models without ONNX:

1. Export from PyTorch: torch.onnx.export(model, dummy_input, "model.onnx")
2. Use ONNX import flow above

Or manual weight loading:

1. Save PyTorch weights as safetensors
2. Define matching Burn architecture
3. Load weights with key remapping

Additional Resources

Consult references/topic-map-onnx.md for:

• Complete operator coverage list
• Opset version compatibility
• Advanced import options
• Weight remapping patterns