Is there a build of Tensorflow 2.x for Jetson NX somewhere?

AI & Data Science Deep Learning (Training & Inference) Frameworks
1

I want to use Tensorflow on my Jetson Xavier NX Development Kit, but I can’t find a container for that architecture. Is there one somewhere? Or do I have to figure out how to make one on my own?

2

Here are the official TF wheels for Xavier AGX: Official TensorFlow for Jetson AGX Xavier. They should work on Xavier NX as well.

3

This doesn’t work. First, it installs an incompatible version of keras-preprocessor. after fixing that, importing tensorflow into python3 fails with a bunch of missing stuff:
$ python3
Python 3.6.9 (default, Apr 18 2020, 01:56:04)
[GCC 8.4.0] on linux
Type “help”, “copyright”, “credits” or “license” for more information.

import tensorflow as tf
2020-06-15 18:12:36.042040: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library ‘libcudart.so.10.2’; dlerror: libcudart.so.10.2: cannot open shared object file: No such file or directory
2020-06-15 18:12:36.042124: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.
2020-06-15 18:12:40.232186: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library ‘libnvinfer.so.7’; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2020-06-15 18:12:40.232470: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library ‘libnvinfer_plugin.so.7’; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2020-06-15 18:12:40.232534: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:30] Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

Help, please.

4

Looks like TF can not find correct CUDA libraries.
What is your JetPack version? (echo $JETSON_JETPACK)
What is your CUDA version? (echo $JETSON_CUDA)

In /usr/local you should see (among others) following folders:

 ...
 cuda -> cuda-10.2/
 cuda-10.0/
 cuda-10.2/
 ...
5

Both of those environment variables are null. That seemed completely wrong, so I downloaded the current SD card image of JetPack 4.4 and started again. On fresh boot, both variables are still null. How is this possible? This page says the SD image is the way to install at least JetPack onto a Jetson Nano/Xavier NX. That appears to be wrong. I must admit that the website is intensely confusing about what goes on a host and what goes on the target. I had become convinced that I could actually develop on the Jetson itself. If this is not true, you guys should make that abundantly clear.

6

Working from this page: JetPack SDK | NVIDIA Developer

7

For reference here are my results:

user@xavier-nx:~$ echo $JETSON_JETPACK
4.4 DP
user@xavier-nx:~$ echo $JETSON_CUDA
10.2.89
user@xavier-nx:~$ ll /usr/local
total 56
drwxr-xr-x 12 root root  4096 Mai 21 08:08 ./
drwxr-xr-x 12 root root  4096 Apr 30 05:16 ../
drwxr-xr-x  2 root root  4096 Mai 21 23:59 bin/
lrwxrwxrwx  1 root root     9 Apr 30 05:09 cuda -> cuda-10.2/
drwxr-xr-x 12 root root  4096 Apr 30 05:08 cuda-10.2/
drwxr-xr-x  2 root root  4096 Apr 26  2018 etc/
drwxr-xr-x  2 root root  4096 Apr 26  2018 games/
drwxr-xr-x  3 root root  4096 Mai 21 23:59 include/
drwxr-xr-x  2 root root  4096 Mai 21 08:09 jetson_stats/
drwxr-xr-x  6 root root 12288 Mai 21 23:59 lib/
lrwxrwxrwx  1 root root     9 Apr 26  2018 man -> share/man/
drwxr-xr-x  2 root root  4096 Apr 26  2018 sbin/
drwxr-xr-x  9 root root  4096 Mai 21 23:59 share/
drwxr-xr-x  2 root root  4096 Apr 26  2018 src/

Did you already update your installation?
sudo apt-get update
sudo apt-get upgrade

Can you rule out that you accidently downloaded the SD-card image for Jetson Nano?

8

Using Nano version on a Nano. Upgrading doesn’t change things. /usu/local has cuda → cuda-10.0.

9

Title says “… Tensorflow 2.x for Jetson NX …”, now you are mentioning Nano - which is it now?
Jetson NX sd-card image is Jetpack 4.4DP by default, while the images for Nano are still Jetpack 4.3 (to my knowledge).

You must choose the Tensorflow release for your Jetpack version…

10

Yes, thought the platforms were more similar. Anyway, Jetson NX on a JetsonXavier NX is no better. Env variables are still null. /use/local/cud’s → cuda-10.2, no jetson_stats directory.

11

Still looks like there is a problem with your L4T/Jetpack installation.

You probably want to open a new subject in the Jetson Embedded System subforum.