How to install the TensorRT library with OS ubuntu v20.04 on the AGX devices

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
1

My question:

How to install the TensorRT library based on the environment of ubuntu20.04 / python3.6 of Docker container on the AGX devices?

Purpose

For some reasons, our ROS has to work in the docker container which is the Ubuntu v20.04 environment on the Jetson device, and we also wanna connect ROS to the TensorRT model for inference.

untitled

Of course, if it can work, we will use it on a lot of Jetson devices in the future.

Issue

After I did some homework, I realized there are some troubles with the installation of TensorRT library.

The host environment of AGX was already installed with Jetpack v4.4 and included the TensorRT library as well. The OS of Docker container was upgraded from Ubuntu v18.04 to Ubuntu v20.04 that the v20.04 of Ubuntu default python version is 3.8; hence, I haven’t tested the TensorRT in python3.8 before. Also, it is not possible to directly use the host library in the container because the python version is not the same.

TensorRT cannot work in python3.8 with previous lower versions. (Only the latest release Tensorrt version (i.e., 7.2.2) is able to support the python 3.8) and that one is for amd64 not for arm64.
I checked from here.

2021-02-2055
2021-02-20551658×761 142 KB

Also, although I think that I can install manually TensorRT library into the container of the jetson device, I haven’t tried this before because normally after we installed the jetpack, it would automatically install the Tensorrt. Hence, I am not sure whether it is possible to do that in the container of Jetson devices.

Of course, I also checked the NVIDIA official website that there are no packages that are able to support Ubuntu 20.04. (or for arm64)

Screenshot-10-06-08
Screenshot-10-06-08678×818 107 KB

Is there any suggestion about my issues? Or is there any further way like avoiding the aforementioned way to use TensorRT in the docker container?

I am glad to try any tests if there is any successful chance.
Thank you so much!

Sincerely,
Chieh

3

Hi,

Sorry that we don’t have a TensorRT package for 20.04 environment + Jetson.
You will need to wait for the future release.

Thanks.

4

Hi @AastaLLL,

Thanks for your reply!

Provided one information: The OS was upgraded from 18.04 to 20.04 which was based on the nvcr.io/nvidia/l4t-ml image. It is able to use the trtexec tools to convert to trt models and import the tensorrt library with python3.6 on Ubuntu20.04.

so it means that TensorRT can work on Ubuntu 18 and 20, but TensorRT cannot work on different python versions (i.e., python3.6 and 3.8)

I will keep waiting for the release.

Thank you!

5

Hi,

Thanks for sharing the Ubuntu 20.04 docker experience with us.
Another WAR is that you can implement a binding for python 3.7 based on TensorRT C++ library directly.

Thanks.