Yolo v7 / Yolo v5 Quantization using OpenVINO POT

This repo using OpenVINO POT tool to quantize yolo-v7 series or yolo-v5 series onnx model to IR model

requirement

pip install openvino-dev==2022.1

Usage:

1. download yolo repo

Yolo v5:
git clone https://github.com/ultralytics/yolov5

Yolo v7:
git clone https://github.com/WongKinYiu/yolov7

2. download pre trained weights (tiny version for example)

Yolo v5:
wget https://github.com/ultralytics/yolov5/releases/download/v6.2/yolov5s.pt

Yolo v7:
wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7-tiny.pt

3. place yolovX_quantize.py into Yolo root folder

4. execute train / test / val script to create labels.cache

5. (Optional) for yolo v5 replace code self.segments[i][:, 0] = 0 to self.segments[i][0][:, 0] = 0 into file <yolo_v5_repo>/utils/dataloaders.py

6. Pytorch to ONNX

Yolo v5:
python export.py --weights <path_to_model>.pt --include torchscript onnx

Yolo v7:
python export.py --weights <path_to_model>.pt --grid --simplify  --img-size 640 640

7. ONNX to OpenVINO IR model

mo.exe --input_model <path_to_model>.onnx

8. modify config (model & dataset path) in yolovX_quantize.py

9. python yolovX_quantize.py