Containerized TensorFlow

TensorFlow with GPU support singularity container

This Singularity container supplies TensorFlow 2.3.0 optimized for use with GPU nodes. It also has opencv, numpy, pandas, seaborn, scikit-learn python libraries.

For more information on Singularity on please see: Containers

Quickstart

# Get allocation
salloc --gres=gpu:2 -N 1 -A hpcapps -t 0:10:00 -p debug
# Run singularity in srun environment
module load singularity-container
unset LD_PRELOAD
srun --gpus=2 --pty singularity shell --nv /nopt/nrel/apps/singularity/images/tensorflow_gpu_extras_2.3.0.sif

Building a custom image based on TensorFlow

In order to build a custom Singularity image based on this one, docker must be installed on your local computer. Docker documentation shows how to install docker.

1. Update Dockerfile shown below to represent the changes desired and save to working directory.
   

   FROM tensorflow/tensorflow:2.3.0-gpu-jupyter
   RUN apt-get update
   RUN DEBIAN_FRONTEND="noninteractive" apt-get -y install python3-opencv
   RUN mkdir /custom_env
   COPY requirements.txt /custom_env
   RUN pip install -r /custom_env/requirements.txt
   

2. Update requirements.txt shown below for changing the python library list and save to working directory.

   seaborn
   pandas
   numpy
   scikit-learn
   git+https://github.com/tensorflow/docs
   

3. Build new docker image

   docker build -t tensorflow-custom-tag-name .
   

4. Create Singularity image file. Note the ./images directory must be created before running this command.

   docker run -v /var/run/docker.sock:/var/run/docker.sock \
   -v $(PWD)/images:/output \
   --privileged -t --rm \
   quay.io/singularity/docker2singularity --name tensorflow_custom.sif \
   tensorflow-custom-tag-name
   

5. Transfer image file to Eagle. For this example I created a directory named /scratch/$(USER)/tensorflow on eagle

   rsync -v images/tensorflow_custom.sif eagle.hpc.nrel.gov:/scratch/$(USER)/tensorflow/