Implementation:Roboflow Rf detr Onnx Simplify

Knowledge Sources	RF-DETR
Domains	Deployment, Model_Optimization
Last Updated	2026-02-08 15:00 GMT

Overview

Concrete tool for simplifying exported ONNX models using graph optimization and the onnxsim library.

Description

onnx_simplify() applies a two-stage optimization: first using the custom OnnxOptimizer for RF-DETR-specific graph passes, then using onnxsim.simplify() for general ONNX simplification with numerical validation. The simplified model is saved with a .sim.onnx suffix.

Usage

Called automatically when model.export(simplify=True) is used, or can be called standalone on an existing ONNX file.

Code Reference

Source Location

• Repository: rf-detr
• File: rfdetr/deploy/export.py
• Lines: L89-114 (onnx_simplify)
• File: rfdetr/deploy/_onnx/optimizer.py
• Lines: L27-39 (OnnxOptimizer.__init__)

Signature

def onnx_simplify(
    onnx_dir: str,
    input_names: List[str],
    input_tensors: List[torch.Tensor],
    force: bool = False,
) -> str:
    """
    Simplify an ONNX model file.

    Args:
        onnx_dir: Path to input ONNX file
        input_names: Model input names
        input_tensors: Sample input tensors for validation
        force: Overwrite existing simplified file

    Returns:
        Path to simplified ONNX file (.sim.onnx)
    """

Import

from rfdetr.deploy.export import onnx_simplify

I/O Contract

Inputs

Name	Type	Required	Description
onnx_dir	str	Yes	Path to input ONNX file
input_names	List[str]	Yes	Model input names (e.g. ["input"])
input_tensors	List[Tensor]	Yes	Sample inputs for numerical validation
force	bool	No	Overwrite existing simplified file (default: False)

Outputs

Name	Type	Description
simplified file	str	Path to simplified ONNX file ({name}.sim.onnx)

Usage Examples

Via Export API

from rfdetr import RFDETRBase

model = RFDETRBase()
model.export(simplify=True)
# Produces both inference_model.onnx and inference_model.sim.onnx

Related Pages

Implements Principle

• Principle:Roboflow_Rf_detr_ONNX_Model_Simplification

Requires Environment

• Environment:Roboflow_Rf_detr_ONNX_Export_Environment