feat: add option for decode output in export_onnx.py

[wep21]

No description provided.

[wep21]

@FateScript Could you review this?

[FateScript]

Thx for contribution. Could you plz tell us why it's needed?

[wep21]

@FateScript TensoRT now supports ScatterND and NMS plugin, so the decode part generated from export_onnx.py can be directly converted to tensorrt engine file.
https://docs.nvidia.com/deeplearning/tensorrt/container-release-notes/rel_21-07.html

[FateScript]

@wep21 Did you try export nms and run it suceessfully(with the correct output)?

[wep21]

@FateScript Yes, I have tested it with my ros2 implementation. https://github.com/tier4/tensorrt_yolox
Is it preferable to share some python codes to reproduce this?

[FateScript]

@FateScript Yes, I have tested it with my ros2 implementation. https://github.com/tier4/tensorrt_yolox Is it preferable to share some python codes to reproduce this?

Thx, I could have a try myself.

[FateScript]

LGTM.

[jbohnslav]

Hello @wep21, can you describe how you got the ONNX graph above? NMS happens in yolox.utils.boxes.py::postprocess. When I add this function to my forward pass, it doesn't use the EfficientNMS_TRT node.

Did you use a ONNX graph surgeon?

[wep21]

@jbohnslav Yes, I created a conversion script with onnx graph surgeon.
https://github.com/wep21/yolox_onnx_modifier