I found a lot of threads of issues with PX2 and Tensorflow and Docker, however I couldnt get a comprehensive impression of what is the current status. https://devtalk.nvidia.com/default/topic/1049100/general/tensorflow-installation-on-drive-px2-/
Could you please clarify:
If I use a tensorflow-gpu docker from the official tensorhub, will I be able to use the gpus on the nvidia?
Is there any other custom dockerfile that enables the gpus support inside a docker?
If I cant use docker, is there a prebuilt version of tensorflow with python3 for the px2?
Does the prebuilt version with python 2.7 support the gpus natively?
The ideal solution would be the use of a docker similar to the way to use it on a regular pc with nvdia-docker for the tensorflow needs. Does this possibility exist?
You will need to use nvidia-docker to have a GPU version TensorFlow.
However, nvidia-docker doesn’t support PX2.
A possible solution is to install TensorFlow directly on the PX2 without using docker.
This will require you to build it from source for the aarch64 support.
You can check this topic for the building steps. Or install this prebuilt file directly.
$ sudo apt-get install libhdf5-serial-dev hdf5-tools
$ sudo pip install tensorflow-1.13.1-cp27-cp27mu-linux_aarch64.whl
$ sudo pip install numpy
Thank you for the support.
The prebuilt version did not install on our PX2, I posted this in the thread you linked. But if docker is not an option we will have another look at it.
Is it planned to integrate the docker support on the nvidia in future releases? This sounded like there could be a possibility: https://devtalk.nvidia.com/default/topic/1022239/general/does-drivepx2-support-nvidia-docker-/ Unfortunately the link in the forum does no longer work.
We don’t have a concrete schedule for nvidia-docker support on PX2 currently.