Initially, when attempting to run YOLOv5 on the Jetson Orin Nano using PyTorch 2.3.0 on JetPack 6.2 (with CUDA 12.6), an error occurred related to the lack of support for the torchvision::nms operator:
RuntimeError: operator torchvision::nms does not exist
This error occurred due to the incompatibility between the installed version of torchvision and PyTorch 2.3.0 on the ARM64 architecture, designed for Jetson.
After solving the torchvision::nms issue by installing the compatible version of torchvision (0.18.0a0), another error appeared related to the use of cuDNN while running the YOLO model:
UserWarning: Plan failed with a CuDNNError: cuDNN error: CUDNN_STATUS_EXECUTION_FAILED
This error is associated with problems when trying to perform convolution operations on the GPU through cuDNN, which can be caused by a number of reasons, including issues with CUDA drivers, cuDNN version, or execution plan incompatibility.
Steps that were taken:
Installed the compatible version of torchvision.
Installed cuDNN 8 for GPU support.
Rebooted the device to clear the GPU state.
However, the cuDNN error continued to occur, indicating issues with the convolution operation execution.