Official TensorFlow for Jetson AGX Xavier

Yes, I totally agree with you, are you using the 4.4 DP Jetpack version?

And about the broken critical tools, I’ve seen no response in this forum, which is the Official TF forum for 20 days. NVIDIA develops and publishes things but I am amazed how terrible the help is, no response in this thread since June of 2019.

I really do not know which are the moderation tools we should ask, as these forums have proved to be no productive at all.

Finally stumbled upon an apparent breakthrough:

Don’t bother with the OpenCV as that should already be there. The rest seems to be close enough to tinker into a workable Tensorflow build. (Still compiling Tensorflow.)

The build instructions for DLIB are incomplete, you should use python3 install --set DLIB_USE_CUDA=1 to enable GPU support.

OpenCV that comes pre-installed with JetPack is compiled without GPU support.

I’m trying to build Tensorflow 2.1, but here the problems:
It supports Cuda10.1 (doesn’t exist for Jetson) and not Cuda10.2 that comes with the Jetpack 4.4

Are there any diffs, from NVIDIA for Tensorflow 2.1, available somewhere ?
My try:
modifying, the ./ for Cuda 10.2, Cudnn 8.0.0 and tensorrt 7.1.0 and remove a line 116 of ./third_party/nccl/buid_defs.bzl.tpl
Build still fails at ./stream_executor/cuda/ with ‘cudnnGetRNNDescriptor’ was not declared…

I’ve done an ota update of the jetpack4.4 from 4.3 but strangely then ENV variable JETSON_CUDNN= is not change to 8.0.0-cuda10.2.
Is there a trouble with the instalation of cudnn 8 ?

Why is no software from Nvidia even forward compatible?

Got Tensorflow 2.1 to build with Cuda 10.2.
But it needed heavy adaptation, to get it to work with Cudnn 8.0.0
-make sure the LD_LIBRARY_PATH is set to /usr/local/cuda 10.2
-modify the ./ as mentioned above
-modify ./third_party/nccl/buid_defs.bzl.tpl the line 116 containing “–bin2c-path=%s”
-modify ./third_party/gpus/ at line 225
-add ./tensorflow/core/platform/tf32_utils.h and .cc form TF-2.2
-add ./tensorflow/stream_executor/cuda/ from TF2.2
-modify ./tensorflow/stream_executor/cuda/ with elements from TF-2.2 look at the end
-modify ./tensorflow/stream_executor/cuda/ with elements from TF-2.2 look for 8000

add --config=v2 --noincompatible_do_not_split_linking_cmdline to the bazel (v29) build command

It did build and seems to work, try it on your own risk: