AGX ORIN onnx-->TensorRT not soupport TopK

[zhangICE]

when I trying to convert onnx model to tensorRT, there is a error such as:
INVALID_NODE: Invalid Node - /rpn/TopK.
This version of TensorRT only supports input K as an initializer. Try applying constant folding on the model using Polygraphy: https://github.com/NVIDIA/TensorRT/tree/master/tools/Polygraphy/examples/cli/surgeon/02_folding_constants.

And the error still exist after I’m applying constant folding on the model using Polygraphy.

Is there anyone know how to fix it? Thanks.

The environment such as:
pytorch 1.13.0
tensorrt 8.5.3.1
JetPack 5.0.2

[AastaLLL]

Hi,

There is a newer TensorRT 8.5 (JetPack 5.1) available.
Would you mind giving it a try?

Thanks.

[zhangICE]

After I update JatPack 5.0.2 to 5.1 , the error cahnged as:
Process finished with exit code 139 (interrupted by signal 11: SIGSEGV)

Could you teach me how to fix this? Thanks.

The environment such as:
pytorch 1.13.0
tensorrt 8.5.3.1
JetPack 5.1-b147

[AastaLLL]

Hi,

Sorry for the late update.

We want to reproduce this issue internally first.
Could you share the ONNX model (before/after constant folding) with us?

Thanks.

[zhangICE]

Due to I had back to JetPack 5.0.2, I can’t share the onnx model to you based on JetPack 5.1 . But I’m pleasure to share the onnx model based on JetPack 5.0.2.
And I found that JetPack 5.0.2+tensorRT8.5 don’t support TopK(), torchvision.ops.nms(), tochvison.ops.MultiScaleRoIAlign(), torech.where() , when I convert onnx model to tensorRT.
These operators are what I have to use in my model, but I don’t know how to fix it. Would you like to help me, thanks a lot.

[AastaLLL]

Hi,

ONNX can be used cross different JePack version.
We can test it on both JetPack 5.0.2 and JetPack 5.1.

Please let us know once the model is ready.
Thanks.