The issue of converting onnx to trt in orinagx hardware

hi
We are currently using the Orin Agx development kit to deploy models and convert ONNX mode to TRT mode. We have no problem with YOLOv5 v7, but during the deployment of OSTRACK, we encountered a problem with model conversion because the previous model input was single input, whereas OSTRACK has two inputs. Therefore, in
Nvinfer1:: IOptimizationProfile * profile=builder ->createOptimizationProfile(); profile>setDimensions(“z”,nvinfer1::OptProfileSelector::kMIN,nvinfer1::Dims4(1,3,384,384));
profile->setDimensions(“x”,nvinfer1::OptProfileSelector::kOPT,nvinfer1::Dims4(2,3,384,384)); This section has been optimized and configured using the above method, and the model can be successfully transformed. However, I believe this is not the most correct method. How should the correct method be implemented?

And during the model conversion process, we encountered an issue with the system automatically shutting down and restarting, so we suspect it may be a problem with the model conversion. However, in the x86 system, the model conversion was successful without this issue

Hi,

Please find our example here:

You can set kMIN, kOPT, and kMAX for each input tensor separately.
Thanks.

This issue has not been resolved, and an error will still be reported. I suspect it may be due to the model being too large, as follows:
ERROR] 4: [optimizer.cpp::computeCosts::3725] Error Code 4: Internal Error (Could not find any implementation for node {ForeignNode[Reshape_62 + Transpose_63…ReduceMean_640]} due to insufficient workspace. See verbose log for requested sizes.)
ERROR] 2: [builder.cpp::buildSerializedNetwork::751] Error Code 2: Internal Error (Assertion engine != nullptr failed. )

This model has no issues with RTX4090

Hi,

Sorry for the late update.

Could you try to run your model with trtexec to see if similar errors were reported?

$ /usr/src/tensorrt/bin/trtexec --onnx=[file] --minShapes=[minShape] ...

Thanks.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.