Jetson Nano 4GB Tensorflow Using GPU Help

I am using the Jetson Nano 4GB that was flashed with the following installed:

nvidia-l4t-core 32.6.1-20210726122000 (which i assume is Jetpack 4.6 or 4.6.1?)
CUDA 10.2
cudnn 8.2.1

For a real time detection project, I am implementing a Machine Learning model that is a .h5 file, and accessing the CSI camera using GStreamer enabled OpenCV. This is a very heavy code so I want to use the GPU. I have tried using Tensorflow-cpu on my code but I am literally getting a new frame every 2 seconds which is crazy slow. Im sure using the GPU would be much better than using the CPU for this case.

How can I can Tensorflow to recognize my GPU so that I can use it for my detection code? What version is compatible with CUDA 10.2 and cudnn 8.2.1 that comes installed on the Jetson Nano?

Thank you!

Dear @wfenno9814 ,
Could you check install TF from Index of /compute/redist/jp/v46/tensorflow

Are those versions of TF only compatible with python3.6?

Yes. Python 3.6 is needed

I have created a venv with python 3.6.9 and have activated it with source /venv/bin/activate. I have also installed dependencies needed like:

python3-pip pkg-config
libhdf5-serial-dev hdf5-tools libhdf5-dev zlib1g-dev zip libjpeg8-dev liblapack-dev libblas-dev gfortran

and when I run the command to install

https://developer.download.nvidia.com/compute/redist/jp/v46/tensorflow/tensorflow-2.5.0+nv21.8-cp36-cp36m-linux_aarch64.whl

I get a bunch of errors mainly saying that it failed to build the h5py wheel or something similar.

It is jetpack 4.6 release(JetPack Archive | NVIDIA Developer)
Do you have restriction to use Jetpack 4.6?

Did you check Official TensorFlow for Jetson Nano!?

I dont quite understand what you are asking here. I have Jetpack 4.6 installed on my Jetson Nano if that is what you are wondering.

@SivaRamaKrishnaNV Could you please provide the exact version of Tensorflow I need to install that is compatible with:

Jetpack 4.6
CUDA 10.2
cudnn 8.2.1
Python 3.6.9

I meant if you can upgrade jetpack release and follow TF installation steps mentioned in above shared link.

Do I need to change my Jetpack version from 4.6 to something else in order for tensorflow to access my GPU?

Please check if the discussion in Tensorflow compatibilty issue - #10 by AastaLLL helps?

Official TensorFlow for Jetson Nano! has verified steps on different jetpack release. If you don’t have any dependency on jetpack version, you may consider upgrading.

When I flashed the OS onto the Jetson Nano, I was only given the default 4.6 download option when following the steps on Nvidia’s page. I could not find a 4.6.3 version to download.

I tried reflashing again but when i turned on the Jetson, it is as if nothing happened and I didnt re-flash the card. Would the 16GB eMMC have something to do with the fact that the re-flash had no virtually no effect on the system? @SivaRamaKrishnaNV

Dear @wfenno9814 ,

Could you check using sdkmanager --archivedversions ?

Hi,

Have you followed the below topic to install TensorFlow?

The hdf5-related error should be fixed by setting ‘Cython<3’

sudo pip3 install --verbose 'protobuf<4' 'Cython<3'

Thanks.