YOLOX-l ONNX is generating valid prediction coordinates and garbage detection score in directML environment, however in CPU environment both prediction coordinates and detection scores are valid.

[aditikohli]

The custom YOLOX_l model is converted into ONNX version using file ( tools/export_onnx.py) and the model is integrated in C# API.

When tested YOLOX_l ONNX model :

1. With CPU environment - Both bounding box and coordinate detection scores are valid.
   2**. With DirectMl environment - bounding box coordinates are valid but detection scores are all 0.**

System Configuration:
windows 11
RTX 5000 ADA Mobile Generation (Ada Lovelace series)
Support DirectX 12 (FL 12.1)
C# API

When tested the same YOLOX_l ONNX model with different GPU
windows 11
RTX A 2000 (Ampere Series)
Support DirectX 12 (FL 12.1)
C# API
Both in CPU and directML environment valid bounding box and coordinate detection scores are generated.

Is this issue related to YOLOX ONNX version compatibility to specific GPUs?

Kindly YOLOX Team, guide to resolve the issue.