I tested Yolov4 that implemented with darknet native and it worked well.
I hope you have a plan to deploy a project with Scaled-Yolov4 darknet net too, or @linghu8812, can you give some suggestions for me.
I tried to convert Yolov4-CSP with your export_onnx.py, then convert to TensorRT by trtexec, and load TRT model by this function. Humzzz, boxes seem predicted incorrectly (i.e width and height in the result is larger than actual).
trtexec, and load TRT model by this function. Humzzz, boxes seem predicted incorrectly (i.e width and height in the result is larger than actual).