I have a different result using TensorRT10 and TensorRT8 for inference. What is the reason for this?

[hanliliy]

Windows 10 system, I use a onnx model, they convert TensorRT model using TensorRT10.12 and TensorRT 8.6，then using trtexec.exe inference same image, but the results are different, what's the reason for this?

Convert: trtexec.exe --onnx=iter.onnx --saveEngine=iter_static.engine

Inference: trtexec.exe --loadEngine=iter_static.engine --loadInputs='input':input.bin --exportOutput=trtexec-result.json

[kevinch-nv]

Different kernels may have been chosen between TRT 8 and TRT 10. How large are the differences?

[hanliliy]

Different kernels may have been chosen between TRT 8 and TRT 10. How large are the differences?
I test a seg model, There are many isolated points in TRT10，TRT8 does not have these isolated points, nor does onnx test in Python

[kevinch-nv]

Are you able to share a repro of the model and test script you're using?

[hanliliy]

Are you able to share a repro of the model and test script you're using?
Thank you for your reply! I use trtexec.exe test the Image, then convert the results to Image 512*512. the below are the files.

diff-files.zip