Yolov5s onnx export but with required output format

[sezer-muhammed]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hi,
I need to run yolov5s on jetson nano, I know jetson can run directly pth file but I want to do it with onnx file.

I convert yolov5s model into onnx file without any problem, and I can run it on jetson nano.

But problem is output type

.

As you can see in image, yolov5 has a output with size 114k85
but example model that compatible with jetson inference library from dusty_nv has an output with scores and boxes.

This causes problem because I guess they coded in this way, so I need to turn yolov5 into this format.

Normally I would change models output layer but yolov5s is a bit complicated for me.

So how can I achieve this?

thanks

Additional

No response

[d57montes]

The yolov5 model output still needs to be post-processed: this includes non-max-suppression among other things. You can see #708 that talks about this. I believe they link some other issues that discusses this even more.
However, you can get get combined non-max-suppression in your ONNX model by following the workflow described in #5938.

[sezer-muhammed]

Thanks a lot!