Application Guide - TensorFlow


TensorFlow™ is an open source software library for numerical computation using data flow graphs. Nodes in the graph represent mathematical operations, while the graph edges represent the multidimensional data arrays (tensors) communicated between them.

TensorFlow was originally developed by researchers and engineers working on the Google Brain Team within Google's Machine Intelligence research organization for the purposes of conducting machine learning and deep neural networks research, but the system is general enough to be applicable in a wide variety of other domains as well.

TensorFlow is capable of taking advantage of the GPU nodes in arc-htc cluster

The guide shows how to

  • prepare a TensorFlow Anaconda virtual environment and
  • submit a test TensorFlow job.

You have two options, you can use the isntalled versions of Tensorflow.

module spider Tensorflow will show current versions:


then module load TensorFlow<version>


Or you could setup your own environment. 

Setting up your environment (Python 3)

Tensorflow is best run on the HTC systems which have GPU nodes. The following commands show how you can set up an Anaconda virtual environment for TensorFlow. Note: this method will only work on HTC nodes. 


module load Anaconda/2020.11
module load CUDA/11.1.1-GCC-10.2.0
module load cuDNN/

export CONPREFIX=$DATA/tensor-env
conda create --prefix $CONPREFIX --copy python=3.8

source activate $CONPREFIX
conda install tensorflow-gpu


Your bash script for submission for GPU nodes will be something like this:

#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --time=00:10:00
#SBATCH --job-name=tensor
#SBATCH --partition=htc
#SBATCH --gres=gpu:1

# Useful job diagnostics
echo "CUDA Devices(s) allocated: $CUDA_VISIBLE_DEVICES"
source activate $DATA/tensor-env

python -c "import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_s um(tf.random_normal([1000, 1000])))"

Please note: You cannot run TensorFlow in GPU mode directly from the login nodes. This will result in missing libcuda errors - as these systems have no GPUs available.