Best package for Jetson nano with Pytorch, Tensorflow, TensorRT, OpenCV etc?

Robotics & Edge Computing Jetson & Embedded Systems Jetson Nano
, , ,
1

Hello all,

Please, I would like to know if is there any image or package with Pytorch, Tensorflow, TensorRT, OpenCV etc already installed in the media?

I have tried to use the Jetson Nano installation media from NVIDIA JP461 (Ubuntu 18.04), however, it comes with no AI frameworks. Later on, I followed all the steps recommended by NVIDIA to install pytorch, and tensorflow, however on the first case (pytorch), I got lots of errors during the installation. And now tensorflow is not found by python3, even though it has already been installed.

Before that, I used SDK Manager to update, Cuda, and Developers Tools.

In another moment, I have installed Q-engineering package with Ubuntu 20.04 LTS already with pytorch, tensorflow, opencv. However, for all 3 classification images that I ran in it, I had problems with Jetson Nano RAM Memory (usually the system requires 2GB, later GPU allocates 1.4GB, and then I just have a few memory available to run my AI algorithms).

So, any help with that will be really appreciated. Thanks.

2

Hi @srevandros, I recommend trying the l4t-ml:r32.7.1-py3 container image - it comes with PyTorch, TensorFlow, TensorRT, OpenCV, JupyterLab, ect:

If you have unofficial 20.04 on your Nano, you will probably need to re-flash fresh SD card to get containers working again.

If this is with PyTorch, PyTorch consumes extra memory for all the CUDA kernels it has built (even if they are unused by your particular application), so for running inference on low-memory systems, it’s recommended just to use TensorRT standalone (you can see the jetson-inference / Hello AI World tutorial that uses TensorRT directly)

1 Like
5

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.