Using tensorflow and Jupyter on HiPerGator

Starting a tensorflow session using the hipergator’s GPU

You can open an interactive session on the hipergator using the dedicated gpu purchased. To do so, follow these steps:

srun -p gpu --gres=gpu:tesla:1 --time=01:00:00 --pty -u bash -i

IMPORTANT! the --pty -u bash -I has to be the last command called, or you’ll eventually end up with an error soon after logging in

inside the interactive session you can then load modules such as Kiras and tensorflow

module load gcc/5.2.0 openmpi keras

module load tensorflow

To launch a python session directly in your terminal:

launch_tensorflow python

If instead you want to run a script, type:

launch_tensorflow python

Be careful, UFRC documentation may be a little misleading. Using srun -p hpg2-gpu --gres=gpu:1 --pty -u bash -i --time=01:00:00 --mem=2gb' would cause an error, because –timeand–memhave been called after–pty -u bash -I`.

That said, two useful pages are:

If you need some guidance, ask Sergio (or purchase) “Deep learning with Python”, from Francois Chollet (

Enjoy your Neural Networks!

Starting a tensorflow session and a jupyter notebook

This script will start a jupyter notebook on the hipergator. The details of how to connect will change each time, so instructions for connecting will be saved to jupyter.log.

The instructions will include a command to type from your local machine, which will look like:

ssh -NL 8080:XXXX.ufhpc:XXXX

This says, “connect port 8080 on my computer to a specific port on a specific hipergator machine.” Thus, visiting http://localhost:8080 in your browser will connect you to the jupyter notebook.

#SBATCH --job-name=jupyter
#SBATCH --output=jupyter.log
#SBATCH --nodes=1
#SBATCH --ntasks=16
#SBATCH --mem=2gb
#SBATCH --time=12:00:00
#SBATCH --qos=ewhite

ml tensorflow
port=$(shuf -i 20000-30000 -n 1)
echo -e "\nStarting Jupyter Notebook on port ${port} on the $(hostname) server."
echo -e "\nSSH tunnel command: ssh -NL 8080:$(hostname):${port} ${USER}"
echo -e "\nLocal URI: http://localhost:8080"
export LD_LIBRARY_PATH=/usr/local/cuda/extras/CUPTI/lib64:$LD_LIBRARY_PATH
launch_tensorflow jupyter-notebook --no-browser --port=${port} --ip='*'

jupyter notebook list

For convenience, you can run the following script to start the hipergator job and automatically print out the log file with the instructions for connecting.


rm jupyter.log
touch jupyter.log
sbatch jupyter.job

tail -f jupyter.log


To launch tensorboard on the server, either start a job with the following commands or type them into a development session:

ml tensorflow

launch_tensorflow tensorboard --logdir=XXXX/

Tensorboard will print something like the following in your terminal: http://dev2.ufhpc:6006

From a local terminal, type ssh -NL 8000:URL_AFTER_HTTP, where URL_AFTER_HTTP will look something like dev2.ufhpc:6006, depending on what was printed above.

Then send your browser to http://localhost:8000/