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 1.x is available within Singularity container

Singularity images are read only. But on can install custom python package outside the container. The packages will be installed in the HOME directory.

Say we want to install panda package for example.

We use mesoshared because the HOME directory is mounted in RW.

Load singularity

$ export LC_ALL=C
$ module load tools/singularity

Install

$ singularity exec /Softs/lumiere/tools/singularity/spool/tensorflow-gpu.img  pip install --user panda

Check installation

$ singularity exec /Softs/lumiere/tools/singularity/spool/tensorflow-gpu.img python  singularity exec /Softs/lumiere/tools/singularity/spool/tensorflow-gpu.img  python                  
Python 2.7.12 (default, Dec  4 2017, 14:50:18) 
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import panda as pd

We will use tensorflow-gpu.img container image.

Here is an example of SGE script that use tensorFlow GPU version withing singularity:

#!/bin/bash 
 
#$ -o $JOB_NAME.o$JOB_ID
 
#$ -N job_tensorflow
 
#$ -q tesla.q   ## we request tesla.q             
 
#$ -l h_vmem=48G   ## tensorflow GPU requires lot of memory
 
 
## load singularity
 
module load tools/singularity
 
## run TF with singularity
 
singularity exec --nv $SING_SPOOL/tensorflow-gpu.img python myTF_program.py

CPU version of TensorFlow is less powerful than the GPU version.

We will use Singularity container: ubuntu-16.img

#!/bin/bash 
 
#$ -o $JOB_NAME.o$JOB_ID
 
#$ -N job_tensorflowCPU
 
#$ -q all.q   ## we request all.q             
 
#$ -l h_vmem=4G   ## we request 4G
 
 
## load singularity
 
module load tools/singularity
 
## run TF within singularity
 
singularity exec $SING_SPOOL/ubuntu-16.img python myTF_program.py