cxmscb / YOLOv8-Openvino

A repository that helps to convert the YOLOv8 detection model to OpenVINO format via onnx and make it more optimized with int8 quantization.

YOLOv8 Openvino Conversion+Quantization

Installation

conda create --name yolorepo python==3.8.5
conda activate yolorepo
python -m pip install --upgrade pip
pip install -r requirements.txt -f https://download.pytorch.org/whl/cpu

Inference (Converting & Quantization)

Add the YOLOv8 model you want to convert to the src/Models path. Then add the necessary parameters to the terminal command below and run it.

python main.py --modelPath src/Models/best.pt --imagePath src/Assets/testImage.jpg --datasetPath src/datasets/val/images --imageSize 480

• modelPath: Path of the pretrained yolo model.
• imagePath: Path of the image that will be used to compare the outputs.
• datasetPath: Path of the dataset that will be used for calibration during quantization.
• imageSize: Image size that the model trained.

In the Output.png image you can see the results of Torch, Openvino and Quantized Openvino models respectively.

Test (Quantized Model)

You can try the quantized model in the test.py file so that you can try it on a single image.

My Results

Speed Test

# yolov8n
Vino Mean FPS:           107.43413924442955
Quantized Vino Mean FPS: 258.2177983756975

Size Test

# yolov8n
Openvino size: 12.1 MB
Quantized size: 3.1 MB

Image Test