wrong confidence score (negative confidence score) on Jetson Nano inference about tensorrt-for-yolo-series HOT 3 CLOSED

I use the following command got different result on different result:

python3 trt.py -e yolov8n.trt  -i src/1.jpg -o yolov8n-1.jpg --end2end

Inference result on GTX 1060:

print(final_boxes, final_scores, final_cls_inds)
[[377.    57.75 907.   708.25]] [0.77539062] [0.]

Inference result on Jetson Nano (Tegra X1)

print(final_boxes, final_scores, final_cls_inds)
[[ 377.   58.  907.  708.]] [-0.22753906] [ 0.]

The bounding box coordinates appear to align quite closely; however, the confidence score appears to be suffering from a conversion error. Do I need to do additional settings during the ONNX to TensorRT transformation phase?

the error casue by nms plugin, recommend use fp32.

python export.py -o yolov8s.onnx -e yolov8s.trt --end2end --v8 -p fp32

from tensorrt-for-yolo-series.

#93 refer this

from tensorrt-for-yolo-series.

thanks for ur reply sir!

from tensorrt-for-yolo-series.