Nvidia Jetson tensorflow version -- mismatch

[Ahrovan]

• According to the installation doc for Nvidia Jetson, will be installed tensorflow==2.3.1
• According to the installation doc for Host PC ubuntu, will be installed tensorflow==2.2.0
• After creating the model in Host and move it to Jetson, the following error is displayed:

  File "/home/ahrovan/env2/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 6843, in raise_from_not_ok_status
    six.raise_from(core._status_to_exception(e.code, message), None)
  File "<string>", line 3, in raise_from
tensorflow.python.framework.errors_impl.InvalidArgumentError: Matrix size-incompatible: In[0]: [1,67712], In[1]: [6656,100] [Op:MatMul]

• But when the model is trained in Jetson. Runs without problems.

Why are different versions of Tensorflow incompatible? And if they do not match. Why these versions not corrected in the documents?

[Ahrovan]

@DocGarbanzo Is this problem related to tensorflow version 2.3.1?
tf_2_9 Can be installed and used?

[DocGarbanzo]

@Ahrovan - can you check that you have the same image setting on your trained model as well as on your robot. I.e. IMAGE_W, IMAGE_H in your myconfig.py? The tf_2_9 branch will be merged to main soon. Please note, the jetson installation for that will be quite tricky, so please wait until we have published the instructions to the docs, because it's very unlikely you'll get it to work without them.

[Ezward]

ubuntu.yml installs tensorflow 2.2.0, see ubuntu.yml#L39 but Jetson Nano installs 2.3.1, see docs setup_jetson_nano.md#L66

In addition mac.yml installs tensorflow 2.2.0, see mac.yml#L43

Windows docs also shows to install TF 2.2.0, see setup_windows.md#L59 and setup_windows.md#L69

[Ahrovan]

pip install tensorflow==2.3.1 in PC Host env/conda solved. just need test. working on it...

[Ezward]

We have to do a little work here; I think the final fix will be when @DocGarbanzo lands the tensorflow 2.9 branch; then both raspberry pi and jetson nano will use the same version, so we can update the docs to that version.

[Ahrovan]

Please update asap. Thank you

[Ezward]

@Ahrovan you have a workaround, is that correct? I am re-opening; this is a real issue and we appreciate you bringing this up. I will close when the tf 2.9 branch lands.

[Ahrovan]

I have prepared this table. Please note that there may be minor mistakes.
One thing is very important. Tensorflow or pytorch user should choose only one. There are many problems in lower versions of JetPack.
@Ezward