tensorflow

Tensorflow

Tensorflow is a large, complex toolkit with a lot of dependancies. We therefore recommend installing with Conda.

Installing with Conda

To install Conda:

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh

bash Miniconda3-latest-Linux-x86_64.sh -b -p $HOME/miniconda3
eval "$(${HOME}/miniconda3/bin/conda shell.bash hook)"
conda init

In order to be make Conda available automatically when you log into the cluster you will also need to add the following to your ~/.bash_profile

if [ -e ${HOME}/.bashrc ]
then
    source ${HOME}/.bashrc
fi

Here is some information on the difference between bashrc and bash_profile

After making these changes log out and log back in.

Installing tensorflow and other packages

Once Conda has been set up install tensorflow with

eval "$(${HOME}/miniconda3/bin/conda shell.bash hook)"
conda create -n tfgpu python=3.11.11
conda activate tfgpu
conda install -y pip
pip install --upgrade pip
python3 -m pip install 'tensorflow[and-cuda]'

After activating the tfgpu environment you can install any additional packages you may need, for example

conda install scipy 

It's worth reading through the Conda users guide. Some useful commands are

• conda list lists all installed packages
• conda search finds available packages that match the provided name, for example conda search torch will find all avaialable versions of torch, pytorch etc
• conda update updates packages

Running the tensorflow example

This directory contains a simple example tensorflow.py that reports on the available devices and performs a simple tensor calculation. To run it on a GPU on the cluster

condor_submit tensorflow_demo.sub

Note that this submit file includes

Requirements = CUDADriverVersion >= 12.0

OrangeGrid includes many different kinds of GPUs, rather than specifying a specific model number it is better to specify the minimum parameters that the job needs in order to run. In this case, recent versions of Tensorflow require a recent version of CUDA.

After submitting you can check on the progress with

condor_q netid

or monitor it with

watch -n 5 condor_q netid

In both cases replace netid with your SU Net ID.

When it completes you can check the output with

cat output/tensorflow_demo.out

The wrapper script

Note that tensorflow_demo.sub does not call tensorflow_demo.py directly. This is because the job needs to be set up so that it will run inside the Conda environment, which is not enabled by default. The submit files therefor calls a wrapper script, which sets up the environment and then runs the tensorflow code. For most simple applications you should be able to modify tensorflow_wrapper.sh without modifying the submit file. Note also that the submit file requires

---

Please email any questions or comments about this document to Research Computing at researchcomputing@syr.edu.