YOLOv4 on tensorRT INVALID_ARGUMENT: Cannot find binding of given name: nmsed_classes

Hello,

I have an issue while trying to run my code based on yolov4.engine that i generated from my onnx file, and i get this error:

[E] [TRT] INVALID_ARGUMENT: Cannot find binding of given name: num_detections

[E] [TRT] INVALID_ARGUMENT: Cannot find binding of given name: nmsed_boxes

[E] [TRT] INVALID_ARGUMENT: Cannot find binding of given name: nmsed_scores

[E] [TRT] INVALID_ARGUMENT: Cannot find binding of given name: nmsed_classes

I am running this on jetson AGX Xavier Dev Kit and the code works fine on my host machine, i got the error only at the moment when i run it on my jetson card.

Does anyone has an idea i am really stuck

Thank you.

Hi,

A common issue is from the different TensorRT version.
Could you share the detailed TensorRT version used in your host and Xavier with us?

Thanks.

Hi! thank you for replying
on my host machine i have :
Tensorrt 7.2.3-1+cuda11.1

on my jetson AGX Xavier:
Tensorrt 7.1.3-1+cuda10.2 it was installed with last jetpack 4.5.1 with sdk manager

i even tried to create a new project with qt and i wrote the same code, i added the files yolov4 needs to run, and it compiles fine but i have the error when i execute the code.

Thanks.

does anyone have an idea about how to solve this problem because i am stuck and i can’t run yolov4 on my jetson dev kit

Thank you

Hi,

Do you have the source that we can check on?

It seems that the engine doesn’t build with the expected output layer (num_detections, …).
So TensorRT cannot find the corresponding binding (input/output layer).

Please also noted that TensorRT doesn’t support portability.
You need to create an engine file from the .cfg and .weight on the Xavier directly.

Thanks.

Hi,

Thanks for your reply, i was generating the engine directly on the jetson AGX but i was getting the same problem, i solved this by using another project that runs yolov4 with C++ and it worked so i guess the problem was due the old project itself.

Thanks for the reply and for your effort.
Have a good day

Thanks for the feedback.
Good to know it works now.