Implementation of Yolo using Facebook's Detectron2 Framework.
With added quantization support following the work in
Chen, Peng, et al. "Aqd: Towards accurate quantized object detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.
This repo implements YoloV5 within Facebook's Detectron2 framework. Currently, only YoloV5m has been fully tested. Support is included for YoloV4-tiny. Support will be extended for other Yolo versions.
This repo also enables quantization and quantization-aware-training using the framework provided in
Chen, Peng, et al. "Aqd: Towards accurate quantized object detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.
The quantization framework is implemented in QTool: A low-bit quantization toolbox for deep neural networks in computer vision. Use the quantization branch to train and test quantized models.
detectron2/projects
directory for consistency.
git clone https://github.com/ShechemKS/Yolo_Detectron2.git
That's it! It will just work.
To train the model run
python train_net.py --config-file configs/yolov5-Full.yaml
You may include any of the usual Detectron2 config options.
To use the model for inference, run
python inference_net.py --config-file configs/yolov5-Full.yaml --inputs path/to/image-dir/ --output path/to/save-dir/