Intro

Forked from branch https://github.com/ayermolo/darkflow.git to support YoloV4 pb and onnx file generation from weights. Take a look of this master branch of darknet for YoloV4 Updates. More variants of YoloV4 weights

Dependencies

Python3.6.1, tensorflow 1.14, numpy, opencv 3, tf2onnx 1.6 version branch 1000b66

Getting started

• Get yolov4.cfg - 245 MB: yolov4.weights (Google-drive mirror yolov4.weights ) paper Yolo v4 just change width= and height= parameters in yolov4.cfg file and use the same yolov4.weights file for all cases.

• Modify darkflow\darkflow\utils\loader.py

  • Change line 121 self.offset = 16 to self.offset = 20

• Run python flow --model path of yolov4.cfg file --load path of yolov4 weights --savepb

  • python36 flow --model cfg/yolov4.cfg --load ..\weights\yolov4.weights --savepb

• Run python -m tf2onnx.convert --input Frozen pb file --output Onnx file name --inputs Input node --outputs Output nodes --verbose --opset Opset number

  • python36 -m tf2onnx.convert --input yolov4.pb --output yolo_v4.onnx --inputs input:0 --outputs BiasAdd_93:0,BiasAdd_101:0,BiasAdd_109:0 --verbose --opset 10

• Updated ONNX file to have a custom Op as last node for YoloV4 post processing. For Reference see update_onnx_model.py. For more details on Operations used, check https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md