tf-gpu-0.1.def

Bootstrap: docker 
From: tensorflow/tensorflow:1.11.0-gpu-py3

%post
    mkdir -p /global/home/users
    mkdir -p /global/scratch
    # /bin/nvidia-smi is a binary on the GPU nodes. Singularity with --nv seems to mount /bin/nvidia-smi and fails unless the file already exists in the container. 
    touch /bin/nvidia-smi

%runscript
    python "$@"

%startscript
    # Not working because can't figure out how to write to a directory accessible to user and instance.start seems to execute without reference to the user
    # if [ "${SLURM_JOB_ID}" != "" ]; then jupyter notebook --no-browser --ip=${SLURMD_NODENAME} 2>> slurm-${SLURM_JOB_ID}.out; else env && jupyter notebook --no-browser --ip="localhost" 2> jupyter-tf-gpu.out; fi