Cuda Support not avaliable for yolov8 on jetson nano?

Autonomous Machines Jetson & Embedded Systems Jetson Nano
, ,
1

I am running yolov8 on the Jetson Nano for a real-time object detection project.

In order to get yolov8 on the Jetson Nano I had to download python3.8. With this I was able to get it running.

The only problem was that the program ran incredibly slowly. So I wanted to run yolov8 on the Jetson Nano’s GPU.

After doing some research I realized that the Jetpack SDK for the Jetson Nano (4.6) was not compatible with the pytorch version that is compatible with python3.8 as seen in this PyTorch for Jetson.

So could I just ignore the fact that it needs Jetpack 5.0 and go ahead with Pytorch version 1.11.0 because it is the oldest version that supports python 3.8? Or is there a different means to get CUDA support for yolov8 on the Jetson Nano?

Any help would be appreciated!

Also, to get yolov8 running on the Jetson Nano I used this tutorial:

3

Hi,

Only JetPack 4 supports Nano.
If you want to use JetPack 5, please change to Xavier or Orin series.

We only provide the PyTorch package with the default Python version.
So for JetPack 4, it’s Python 3.6.

But you can still build the PyTorch with CUDA support for custom Python from the source.
The instructions can be found in the below topic:

Thanks.

4

Alright, just to clarify so if I go in and build PyTorch 1.11.0 in my virtual env, I could possibly run yolov8 with cuda support? It won’t screw anything up right?

Thanks!

5

Hi,

Yes, CUDA is a C++ library and can be wrapped with a custom Python version.

We checked the video you shared, they seem to run YOLOv8 on Nano with CUDA support successfully.
Thanks.

6

All right, thank you! I’ll let you know once I do it!

7

UPDATE: when I ran this command -
pip3 install numpy torch-1.14.0-cp36-cp36m-linux_aarch64.whl

I received this error:
ERROR: torch-1.14.0-cp36-cp36m-linux_aarch64.whl is not a supported wheel on this platform.

8

Hi,

You will need to build Pytorch from the source.
The package only supports Python 3.6 but it seems you need Python 3.8?

Thanks

9

Thank you for the clarification! I just realized I was trying to install from pip!

So to install from source would I basically follow what this says?

Also, yeah yolov8 didn’t support python 3.6 so I created a virtual environment to install python3.8 on the Jetson nano.

Sorry if this is a wired question lol I got the Jetson nano a few years ago but I am just starting image recognition on the board now.

10

Hi,

No, the link also tries to install a prebuilt package (pip3 install torch-xxx.whl).
To build it from the source, please find the instructions below:

> Build from Source
...
$ python3.8 setup.py bdist_wheel

Thanks