YOLO to ONNX Conversion

Converting a YOLO Model to ONNX

LogicalDOC performs object detection using the ONNX Runtime Java API. Therefore, a trained YOLO model must be converted from the native PyTorch (`.pt`) format to the Open Neural Network Exchange (ONNX) format before it can be imported into LogicalDOC.

ONNX (Open Neural Network Exchange) is an open standard for representing machine learning models. It enables models trained in one framework, such as PyTorch, to be executed efficiently in different programming languages and runtime environments.

The conversion process does not retrain or modify the model. Instead, it exports the trained network, including its learned weights and computational graph, into a portable format that can be executed by ONNX Runtime.

Why ONNX?

Ultralytics trains YOLO models using PyTorch. However, LogicalDOC is implemented in Java and performs inference using the ONNX Runtime library.

Using ONNX provides several advantages:

• No dependency on the Python runtime during inference.
• Native Java support through the ONNX Runtime API.
• Faster startup and lower memory consumption than embedding a Python interpreter.
• Cross-platform model portability.
• Support for hardware acceleration when available.

Export the Model

The following script converts the trained `best.pt` model into the ONNX format:

from ultralytics import YOLO

model = YOLO("runs/detect/target/weights/best.pt")

model.export(
format="onnx",
simplify=True,
nms=True
)

Where:

• `format="onnx"` exports the model in the ONNX format.
• `simplify=True` simplifies the exported computation graph to improve compatibility and potentially reduce inference overhead.
• `nms=True` embeds the Non-Maximum Suppression (NMS) step into the exported model, allowing the ONNX Runtime to return the final detections directly.

After the conversion completes, the generated model (`best.onnx`) can be imported into LogicalDOC for inference.