TensorRT conversion error for TAO RetinaNet model on Jetson Xavier NX

Hi, during model conversion from .onnx to tensorrt .plan, we are getting a conversion error for the TAO RetinaNet model on Jetson Xavier NX.

Conversion is done like this:
trtexec --onnx=model.onnx --saveEngine=model.plan

Log here tensorrt_conversion_log.txt (103.1 KB)

We are using the AAEON BOXER-8251AI (AAEON BOXER-8251AI | AI@Edge Fanless Embedded Box PC with NVIDIA Xavier NX | Microsoft Azure Certified - AAEON). Driver and tensorrt version were installed as below. The model was previously also run on other devices including a jetson orin, where we had no compilation issues. Can you help us with this? Thanks


TensorRT Version : tensorrt/now, nvidia-tensorrt/now 5.0.2-b231
GPU Type : NVIDIA® Jetson Xavier™ NX
Nvidia Driver Version : nvidia-jetpack/now 5.0.2-b231
CUDA Version : cuda-11-4/now 11.4.14-1, nvidia-cuda/now 5.0.2-b231
CUDNN Version : libcudnn8/now, nvidia-cudnn8/now 5.0.2-b231
Operating System + Version : Ubuntu 20.04.6 LTS
Python Version (if applicable) :
TensorFlow Version (if applicable) : not installed
PyTorch Version (if applicable) : not installed
Baremetal or Container (if container which image + tag) : baremetal

Dear @harryhirsch,
Could you check ONNX opset. Please see TensorRT Parsing ONNX Model Error - #9 by philminhnguyen helps.

I am getting this error when checking the model.

ValidationError Traceback (most recent call last)
/tmp/ipykernel_225078/4085541278.py in
4 # model = onnx.load(filename)
----> 5 onnx.checker.check_model(“model.onnx”)

/opt/conda/lib/python3.7/site-packages/onnx/checker.py in check_model(model, full_check)
123 # If model is a path instead of ModelProto
124 if isinstance(model, str):
→ 125 C.check_model_path(model, full_check)
126 else:
127 protobuf_string = (

ValidationError: Field ‘shape’ of ‘type’ is required but missing.

I am also not able to load the model in onnxruntime with TRT execution provider as described here Error when running retinanet model in onnxruntime with tensorrt execution accelerator

However, I am able to convert the model to tensorrt on multiple different platforms including an A5000 GPU, Jetson Orin, so I think it must be something to do with the Jetson Xavier NX.

Dear @harryhirsch,
Could you share the ONNX model?

Hi, I send the model to you in a chat since it’s trained on our data.

Dear @harryhirsch ,

It looks like issue with Jetpack version. I could build TRT model using trtexec on jetpack 5.1.2. Could you check with latest Jetpack release for Xavier NX?