Failed to parse ONNX i.e deeplabv3_resnet101.onnx semantic segmentation model on Jetson Xavier NX

Hi @dusty_nv ,

We have trained the custom semantic segmenation model referring the repo with deeplab v3_resnet101 architecture and converted the .pth model to .onnx model.

But when running the .onnx model with segnet on Jetson, we are getting error failed to parse the onnx model and not able to do inference.

However when we tried the custom trained with fcn_resnet34 architecture we were able to do inference with segnet on Jetson.

Can you let us know the cause of this error? And also, can you clarify whether there are any limitations on architecture while using Segnet? E.g.

does it only support fcn_resnet family and not deeplabv3_resnet family on Jetson?

Reference:
Jetson Xavier NX (Developer kit version):
TensorRT version - 7.1.3.0 with cuda10.2
Jetpack 4.5.1,
OpenCV 4.1.1

Please let us know if you need more information.

Hi @nishantshrivastav23, it’s only tested with FCN-Resnet, yes.

To attempt support for DeepLab-v3, you would need to dig into the error that occurs during parsing of the ONNX and if necessary, attempt to remedy the model in PyTorch so that it’s able to be imported into TensorRT.

Typically the network architectures that I select to use are able to be exported/imported via ONNX without much issue. I do recall a couple small tweaks being necessary for FCN-Resnet, but these didn’t affect the active parts of the network structure itself.

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