TensorFlow

TensorFlow on HPC can be run in batch, interactive, or Jupyter Notebook.

TensorFlow interactive job

Start the job.

srun --job-name=tensorflow --ntasks=1 --time=1:00:00 --gres=gpu:1 --pty bash

Load the TensorFlow module.

module load tensorflow

Start your training or other commands.

python train.py options input output

When your commands end, always remember to end the interactive job.

exit

TensorFlow batch job

This is an example of running a Python script in a SLURM batch job.

learning-ml.pylearning-ml.slurm

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
session = tf.Session()
tf.print(hello)

#!/bin/bash
#SBATCH --job-name=learning
#SBATCH --ntasks=1
#SBATCH --time=00:01:00
#SBATCH --account={PI_NetID}
#SBATCH --partition=gpu
#SBATCH --gres=gpu:1
#SBATCH --output=%x-%j.out

module load tensorflow
python /scratch/g/pi_netid/learning-ml.py >> output.log  

Save these scripts to your scratch directory, and submit the job:

sbatch learning-ml.slurm

This simple job should print Hello, TensorFlow! to your output file.

TensorFlow Jupyter Notebook job

This functionality is provided by Open OnDemand!

See Jupyter on Open OnDemand for details.

TensorBoard Job

This functionality is provided by Open OnDemand!