NVIDIA-ISAAC-ROS / isaac_ros_common

Common utilities, packages, scripts, Dockerfiles, and testing infrastructure for Isaac ROS packages.
https://developer.nvidia.com/isaac-ros-gems
Other
191 stars 137 forks source link

torchvision compatibility issue... #30

Closed bblumberg closed 2 years ago

bblumberg commented 2 years ago

To get torch and torchvision to work (This probably should be documented some place):

  1. Binaries from the pytorch website don't use cuda on the agx orin.
  2. Follow the instructions to download pytorch wheel from nvidia: https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html). For what it is worth, I used torch-1.12.0a0+84d1cb9.nv22.4-cp38-cp38-linux_aarch64.whl and not the absolute latest because there was a compatibility issue with yolov5.
  3. You must download torchvision and build from source.
  4. For what it is worth, if you use the realsense container, and want to use the python interface, it is pyrealsense2.pyrealsense2, as in: import pyrealsense2.pyrealsense2 as rs
hemalshahNV commented 2 years ago

Thanks for working this out! We'll explore updating our Isaac ROS Common images to build torchvision from source directly.

jaiveersinghNV commented 2 years ago

The CUDA-enabled version of torch is available in the Isaac ROS Dev base images released in v0.11.0, but torchvision was not rebuilt from source.