Tensorflow 2.1.0 for JetPack 4.4 was incorrectly build with V1.x API setting

Autonomous Machines Jetson & Embedded Systems Jetson Xavier NX
1

The current version of Tensorflow 2.1.0 Python for JetPack 4,4, the only one that is available for the Jetson Xavier NX was compiled incorrectly. This means that the interface of Tensorflow classes/functions is not conform the one defined for version 2.1.0.

This means that code written for Tensorflow 2.1.0 in python won’t execute correctly as some methods/functions are not available or have different parameters.

There is a workaround for this when you are using Tensorflow 2.1.0 using only your own code by using the the trick mentioned here:

But this will not work when you are using Python packages that use Tensorflow 2.1.0 directly (or you get into a whole mess of editing managed packages.

When will there be an corrected version available for JetPack 4.4 or even a newer version (2.1.1). As it is I’m blocked with testing our Python TF 2.0 application on the Jetson Xavier NX!

3

Hi,

Sorry for the inconvenience.

We have passed this issue to our internal team and now is under investigation.
Will let you know if any progress.

Thanks.

4

As you can see from the linked forum post in my post, this problem was found already more than two months ago (May 14 or maybe even earlier), so it’s a bit strange that there hasn’t been any action on this point sooner. Strangely enough I see your name in a posting on that forum thread, so you should have been aware of this problem.

But if this leads to the problem being resolved I and everybody else who was impacted by this incorrect build issue will be very happy.

5

Here is an easy way to demonstrate the problem, as ‘eager execution’ is the default in tf 2

import os
import tensorflow as tf
import cProfile
tf.executing_eagerly()
False

6

Hi, pieterjan.kuyten

Sorry for the inconvenience.
This issue won’t occur in our next release which targets for this week.
We understand that this brings lots of problem and particular for XavierNX.

Each of our official release need to go to the standard QA testing.
So it really takes some time to make it public accessible.
And we cannot share too much information here before the release is ready for public.

I will update here if the link is published.

Thanks.

7

Thanks for the update. I’ll be waiting with high hopes, because that would finally allow us to get further with our tests on Jetson Xavier NX

8

Hi,

The package is released.

$ sudo pip3 install --pre --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v44 tensorflow==2.2.0+nv20.6

Thanks and sorry for the inconvenience.