Is there a Pip installable version of TorchVision for Jetson AGX Orin?

Installing TorchVision with:

pip install torchvision

leads to binary incompatibility of torchvision 0.14.1 with torch 1.13.0a0+936e9305.nv22.11:

Runtime error: Couldn't load custom C++ ops

Is there a compatible version of TorchVision I can install with pip? Do I have to build it from source?

Yes, you would need to build torchvision from source on your Jetson device.

Thank you. Do you have a link to reliable instructions how to build it?

Hi,

You can find the instruction in the below topic:

Thanks.

These are PyTorch installers. I already installed a newer version from https://developer.download.nvidia.cn/compute/redist/jp/v502/pytorch. What I’m looking for is torchvision, which is a PyTorch add-on.

I tried to build torchvision from source but run into this problem:

RuntimeError: 
The detected CUDA version (12.0) mismatches the version that was used to compile
PyTorch (11.4). Please make sure to use the same CUDA versions.

PyTorch (11.4) works just fine with CUDA 12.0, even though it was built with an older version. Is there any other solution than reinstalling the original OS image with older CUDA version? Can I have two CUDA versions side by side?

Any chance Nvidia can provide PyTorch wheels built with CUDA 12.0?

What is the way to decode nv22.12 sections of wheel name into CUDA version used to build it?

Hi,

Do you upgrade the CUDA package from the website below?

If yes, you will need to build PyTorch from the source.
We only provide the prebuilt for the software that included in the JetPack.

Thanks.

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