Adding ONNX info + fix for export

[adujardin]

This PR adds ONNX exporting instructions and also includes a fix to export the yolov4 models (by default it's the same issue as ultralytics/yolov3#1169, the solution posted by @zjd1988 is fixing it).

All yolov4 variants ONNX models tested are working using TensorRT 7.0+ (not tested other frameworks).

[AlexeyAB]

@adujardin Hi,
Did you try to convert from generated ONNX-file to the CoreML-model?

[adujardin]

Hi @AlexeyAB, no I never tried CoreML. I guess it should be compatible as it's fairly close to the format of ultralytics/yolov3 but I'm not familiar with it. Using ONNX opset_version 11 might be better but I had an issue with TensorRT if opset_version was different from 10.

Also not sure if it's the export process or the models themself but the accuracies are lower than vanilla darknet yolov4 model (it's better than yolov3 though, exported in onnx with the same process and inference implementation)

[yoobright]

yolov4-tiny convert not works