Low Precision Yolo

This repo contains code for training Yolo(v1, v3, v5) architectures using different quantization. It utilizes Brevitas which is a Pytorch research library for quantization-aware training (QAT) from Xilinx. The code is developed on top of YOLOV5.

Installation

$ git clone git@github.com:sefaburakokcu/quantized-yolov5.git
$ cd yolov5
$ pip install -r requirements.txt

Inference

detect.py runs inference on a variety of sources and saving results to runs/detect.

$ python detect.py --weights /path-to-weights-folder
		    --source 0  # webcam
		            file.jpg  # image 
		            file.mp4  # video
		            path/  # directory
		            path/*.jpg  # glob
		            'https://youtu.be/NUsoVlDFqZg'  # YouTube
		            'rtsp://example.com/media.mp4'  # RTSP, RTMP, HTTP stream

Training

Run commands below to reproduce results on Widerface dataset.

$ python train.py --data widerface.yaml --cfg models/hub/yolov1-tiny-quant.yaml --weights '' --batch-size 128

Export

Sigmoid activation is used in detect layer when training whereas HardTanh is used when exporting the model for FINN.

$ python export.py --data widerface.yaml --weights runs/train/exp1/weights/best.pt --nodetect

References

• YOLOV5
• BREVITAS

Authors

• Sefa Burak OKCU
• Bestami GÜNAY