Is your feature request related to a problem? Please describe.
NGC does not host tao_deploy container for aarch64. It is not clear how to build such a docker image by yourself as documentation is repeated in various places (tao documentation, this github, nvidia forums). This makes it difficult to use tao_deploy on Jetson platforms. Building this container, by following instructions here and instead linking tao_deploy container in other sources would make simple and clear how to run this container on jetson.
-
Instructions in this repo recommend starting from l4t tensorrt container, running several commands that can produce a working, tao_deploy container with buildx.
-
The documentation on NVIDIA site incorrectly omits some of the necessary steps: https://docs.nvidia.com/tao/tao-toolkit/text/tao_deploy/tao_deploy_installation.html#installing-tao-deploy-through-wheel - only asks to do pip install nvidia-tao-deploy which fails on mpi4py and then on pycocotools-fix.
-
NVIDIA forum https://forums.developer.nvidia.com/t/how-to-properly-install-tao-deploy-bash-tao-deploy-command-not-found/246408 thread goes on to recommend running TensorRT container with Tao documentation which is incorrect, then recomends just running the regular Tao container on x86 and using output on Jetson.
Describe the solution you'd like
Pre-build and publish container for aarch64 to ngc.
Describe alternatives you've considered
Add a Dockerfile.aarch64 in this repo to make it easy to actually build a container.
Additional context
Is your feature request related to a problem? Please describe.
NGC does not host tao_deploy container for aarch64. It is not clear how to build such a docker image by yourself as documentation is repeated in various places (tao documentation, this github, nvidia forums). This makes it difficult to use tao_deploy on Jetson platforms. Building this container, by following instructions here and instead linking tao_deploy container in other sources would make simple and clear how to run this container on jetson.
Instructions in this repo recommend starting from l4t tensorrt container, running several commands that can produce a working, tao_deploy container with buildx.
The documentation on NVIDIA site incorrectly omits some of the necessary steps: https://docs.nvidia.com/tao/tao-toolkit/text/tao_deploy/tao_deploy_installation.html#installing-tao-deploy-through-wheel - only asks to do
pip install nvidia-tao-deploywhich fails on mpi4py and then on pycocotools-fix.NVIDIA forum https://forums.developer.nvidia.com/t/how-to-properly-install-tao-deploy-bash-tao-deploy-command-not-found/246408 thread goes on to recommend running TensorRT container with Tao documentation which is incorrect, then recomends just running the regular Tao container on x86 and using output on Jetson.
Describe the solution you'd like
Pre-build and publish container for aarch64 to ngc.
Describe alternatives you've considered
Add a Dockerfile.aarch64 in this repo to make it easy to actually build a container.
Additional context