Hi,
Here is a similar issue that fixed by using opset-11 onnx format:
https://github.com/NVIDIA/TensorRT/issues/386#issuecomment-604183816
May I know which opset version do you use?
If not opset-11, would you mind to give it a try?
You can check if this issue is fixed by running our trtexec binary.
$ /usr/src/tensorrt/bintrtexec --explicitBatch --onnx=[name].onnx
Thanks.