wangsssky / YOLO-Nano

An unofficial pytorch implementation of YOLO Nano

Introduction

• YOLO Nano: a Highly Compact You Only Look Once Convolutional Neural Network for Object Detection. paper.
• This repo is based on liux0614.

Project Structure

root/
  datasets/
    coco/
      images/
        train/
        val/
      annotation/
        instances_train2017.json

Installation

git clone https://github.com/wangsssky/YOLO-Nano.git
pip3 install -r requirements.txt

COCO

To use COCO dataset loader, pycocotools should be installed via the following command.

pip install "git+https://github.com/philferriere/cocoapi.git#egg=pycocotools&subdirectory=PythonAPI"

To train on COCO dataset:

python3 main.py --dataset_path datasets/coco/images --annotation_path datasets/coco/annotations 
                --dataset coco --conf_thresh=0.8 --gpu

Convert to onnx

• cd deploy, run convert2onnx.py
• run python -m onnxsim yolo-nano.onnx simplified.onnx, you may install onnx-simplifier first.
• try run_onnx.py, test it by onnx runtime