Tensorflow: “No protocol specified"

post by tschmah on Jul 11, 2022

I have Ubuntu 20.04 with a Lambda stack that was provided at time of purchase of my workstation and updated today using “sudo apt-get update && sudo apt-get dist-upgrade”

I created a virtual environment for testing:

python3 -m venv --system-site-packages testenv
source testenv/bin/activate

Within this environment, I am able to use tensorflow successfully via a jupyter notebook, however I get an error message “No protocol specified” if I start a python session from the command line and then import tensorflow:

>>> import tensorflow as tf
>>> No protocol specified

What does this error mean? Do I need to set up an environment variable?

post by markd on Jul 21, 2022

I am guessing it is failing the same way without venv also.

From searches it shows references to issues with DISPLAY.

$ python -c ‘import tensorflow as tf ; print("\nTensorflow version: ",tf. **version**, "\nTensorflow file: ",tf. **file**)

Tensorflow version: 2.8.0

Tensorflow file: /usr/lib/python3/dist-packages/tensorflow/ init.py

I have not been able to repeat this.

I have seen a similar issue once with two processes running on the same machine with Tensorboard. As it was using a invalid port (one reserved for X11)

TensorBoard on the default port of 6006 * (6000 - 6007 are reserved for X11)

If this is a Tensorboard in the job you can change Tensorboard to use port 9000.

It would get stuck trying to start orted (part of MPI).

post by tschmah on Jul 24, 2022

Thank you for your thoughts, markd. You are correct that it fails the same way without venv also:

$ python -c ‘import tensorflow as tf’
No protocol specified

$ python --version
Python 3.8.10

I agree that the problem seems to be something to do with displays. I’m not using TensorBoard.

post by markd on Jul 31, 2022

Another way to find out what is happening is run your code with:

$ strace python -c ‘import tensorflow as tf’ >& tensorflow-strace.txt

You can send that output file here or contact ‘support@lambdal.com’ with that file and nvidia-bug-report.log.gz from ‘sudo nvidia-bug-report.sh’

The two main commands would be if you are ssh’ing into the machine:

1. Enable X11 forwarding (it can be disabled at the client or server level also).

   $ ssh -X machine
   

   You can check the server setting:

   $ grep X /etc/ssh/sshd_config
   

2. You can play with: a. ‘xhost +’ will give access to anybody to connect to your display (okay to test, but I would not keep it that way). I would recommend xhost +local:[hostname or ip]

   (NOTE: To remove xhost setting: ‘xhost -’)

   $ xhost +<mymachine or IP>
   

   example: xhost +10.0.0.2

b. Set the DISPLAY

$ export DISPLAY=:0.0

c. xauth:

$ xauth merge xauth_extract

or remote host via ssh:

$ xauth extract - $DISPLAY \| ssh user@<machine or IP> xauth merge -