Tensorflow container image

For JP 4.6.1, can we use the general image TensorFlow | NVIDIA NGC which has the latest tag 22.03-tf2-py3. The image didn’t explicitly say Jetson but it is multi-arch and has arm64 support. I tried and it seems to run on AGX board, and has cuda 11.6, tensorrt 8.2.3, cudnn 8.3.3, TF 2.8.0, etc.

Or do we have to use the L4T version for Jetson: NVIDIA L4T TensorFlow | NVIDIA NGC which has latest tag r32.7.1-tf2.7-py3, which has older version of cuda, tensorrt, cudnn, TF , etc, compared to the other tensorflow image.

Hi,

Please use the L4 T-based image.
The CUDA, cuDNN, and TensorRT version is synchronized to the package included in the JetPack.

Since you are using Xavier, there is a newer JetPack 5.0 DP that has CUDA 11.
Please note that r32.7.1-tf2.7-py3 is for r32.7.1 OS which is JetPack 5.0 DP JetPack 4.6.1.
For JetPack 5.0DP, please use the r32.6.1-tf2.5-py3 r34.1 container for compatibility.

Thanks.

r32.7.1 is for JP4.6.1.

I don’t see r34.1.0 (JP 5.0 DP) yet at NVIDIA L4T TensorFlow | NVIDIA NGC

Hi,

Sorry for the incorrect information.
You can find the container for JP5.0DP now:

Thanks.

Thanks for the update! Unfortunately I can’t use the container for JP 5.0DP since my agx runs JP 4.6.1 as mentioned in the first post. JP 5.0 DP container can only run on a host that also runs JP 5.0 DP, based on this post

Hi,

If you want a newer CUDA package, you will need to upgrade the JetPack to 5.0DP.
If not, please use the r32.7.1-tf2.7-py3 with CUDA 10.2 on a JetPack 4.6.1 environment.

Thanks.

As mentioned here, JP 5.0 only supports DevKit for now, it won’t boot custom boards, so we will stick with JP 4.6.1.

Hi,

For JetPack 4.6.1, the latest TensorFlow container will be r32.7.1-tf2.7-py3.

Thanks.