Segfault error when exporting model from ONNX to TensorRT

I retrained the ssd resnet 50 v1 fpn 640x640 and exported to ONNX successfully. When I try to export it to TensorRT using the code here: Speeding up Deep Learning Inference Using TensorFlow, ONNX, and TensorRT | NVIDIA Developer Blog, the segfault error appears.

I tried ./trtexec --onnx=model.onnx --verbose --explicitBatch and got the same error
trt_log.txt (51.6 KB)

Using the code in the NVIDIA TENSORRT DOCUMENTATION at the section Importing from ONNX using Python gives me the same error

Developer Guide :: NVIDIA Deep Learning TensorRT Documentation

After using gdb I got a new error:

0x0000007fa957d0f8 in nvinfer1::plugin::BatchedNMSPlugin::getOutputDataType(int, nvinfer1::DataType const*, int) const ()
   from /usr/lib/aarch64-linux-gnu/


  • TensorRT Version: 7.1.3
  • GPU Type: Jetson Nano
  • Nvidia Driver Version: JetPack-4.5
  • CUDA Version: 10.2
  • CUDNN Version: 8.0.0
  • Python Version: 3.6.9
  • Baremetal or Container: Baremetal


  1. The command used to export the model to ONNX is: python -m tf2onnx.convert --saved-model saved_model --output model_resnet.onnx --fold_const --tag serve --verbose --opset 12
  2. The input dtype of the ONNX model was changed to np.float32 since INT8 is not compatible with TensorRT
  3. Layer + Register API was used to replace the NMS TensorFlow uses with BatchedNMS_TRT (I dont think this is a problem since the log I uploaded says Successfully created plugin: BatchedNMS_TRT)
  4. I have tried changing the builder.max_workspace_size() in the code from the blog to different values with no success (1 << 30), (3 << 30) (It crashes even faster)
  5. When I used onnx.checker.check_model() I got an error that no op was registered as BatchedNMS_TRT for domain_version 12. However, apparently that op is not technically a valid ONNX op so this is expected. Plus the trtexec log says Successfully created plugin: BatchedNMS_TRT)

Possible solutions

  • Using an earlier JetPack version?
  • Using CombinedNonMaxSuppression in the TensorFlow model? Since I read in a post made by @klinten that

BatchedNMSPlugin is modeled directly after the TensorFlow CombinedNMS

I dont know it this would be more ‘compatible’ with TensorRT, thus solving my problem

  • Optimizing the model a bit more before exporting to TensorRT? (if the problem occurs because of low memory)

So once again the question is:
How do I export my ONNX model to TensorRT?
Thanks in advance for your help :)


Based on the gdb log, the error occurs in the BatchedNMSPlugin.

Could you check if the layer does meet the requirement of batchedNMSPlugin?


The batchedNMSPlugin takes two inputs, boxes input and scores input.

Boxes input The boxes input are of shape [batch_size, number_boxes, number_classes, number_box_parameters] . The box location usually consists of four parameters such as [x1, y1, x2, y2] where (x1, y1) and (x2, y2) are the coordinates of any diagonal pair of box corners. For example, if your model outputs 8732 bounding boxes given one image, there are 100 candidate classes, the shape of boxes input will be [8732, 100, 4] .

More, we have an example to deploy Retinanet with TensorRT.
Could you also try the plugin layer to see if it works?


Thank you very much for your answer.

About your first question, I tried implementing the batchedNMSPlugin in a different way.
This time, the error was:

ERROR: Graph contains a cycle

log_MultiClass_nms.txt (50.6 KB)

I read that the error could be caused by a map_fn TensorFlow op, but I’m not sure if that’s the problem here.

In the meantime, I’m gonna take a look at the examples you link to :)

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.


Does the BatchedNMSPlugin work with the sample shared above?

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.