Kneron

Tutorial: Deploying YOLOv5

In this tutorial, we will take a pre-trained YOLOv5n model, convert it, and run it on the KL720 to detect objects in real-time.

Level: IntermediateTime: 15 mins

1. Download Pre-trained Model

We provide a Kneron-optimized version of YOLOv5n (Nano) that fits within the KL520/720 memory constraints.

Download yolov5n.onnx (4.2 MB)

2. Pre-processing

YOLO requires images to be resized to 640x640 and normalized (0-1).

# Python Image Prep img = cv2.imread("dog.jpg") img = cv2.resize(img, (640, 640)) img = img.astype(np.float32) / 255.0

3. Convert & Compile

Use the toolchain to compile the ONNX file.

kneron compile yolov5n.onnx --target KL720 --out yolov5n.nef

4. Post-processing

The NPU outputs raw anchor boxes. You must apply Non-Maximum Suppression (NMS) on the host CPU.

Note: Hardware NMS is available on KL720 but requires special graph configuration. For this tutorial, we use CPU NMS.
results = monitor.inference(img)
boxes = nms(results, conf_thres=0.4, iou_thres=0.5)
plot_boxes(img, boxes)

5. Live Demo

If you have a webcam connected, run the live demo script:

python examples/yolo_webcam.py
Steps