Simple singularity recipe for Keras (Python 3) & Flux (Julia) (personal use) Base image: nvidia/cuda + julia at Dockerhub

$ sudo singularity build /opt/singularity/flux-jskim.sif flux-julia.def

$ sudo singularity build /opt/singularity/keras-jskim.sif keras-py3.def

$ sudo singularity build /opt/singularity/torch-jskim.sif torch-py3.def

$ julia

• press ]

(v1.4) pkg> up

• Install python via pyenv and pyenv-virtualenv

$ pyenv install 3.8.2
$ pyenv virtualenv 3.8.2 env
$ pyenv local env
(env) $ pip install package_name

• If you have changed sif file name, change corresponding name in keras and flux script

$ mkdir -p ~/usr/bin
$ chmod +x keras
$ chmod +x flux
$ cp keras ~/usr/bin
$ cp flux ~/usr/bin

or you can make symbolic link (ln -s src dest to your bin directory)

• Assume ~/usr/bin has been added to your PATH

• assume we have some_case_name and save code results to ${HOME}/data/some_case_name

$ flux -b some_case_name my_code.jl

$ torch -b some_case_name my_code.py

$ keras -b some_case_name my_code.py

$ flux -s

$ torch -s

$ keras -s

1. Use envionments in Pkg because host and guest system share ~/.julia so there would be build issue

   • Link : https://julialang.github.io/Pkg.jl/v1/environments/#Creating-your-own-projects-1

2. Shell script may not be work in specific order.

• run following command

    flux -b "name" -n

• connect local machine via SSH tunneling using following command, then open web browser by this address, localhost:8157`.

ssh "remoteId"@"remoteIP" -p "remoteSSHPort" -NL 8157:localhost:8889 "remoteId"@"remoteIP"

1. Better docs
2. Better usage (-h option)
3. Better command line arguments handling in shell script