Introduction to containerized CUDA-based machine learning workflow

This repository contains the basic ideas for creating a containerized CUDA machine learning workflow with pytorch and docker. The project was developed under Python 3.11 and Linux Ubuntu 22.04 lts.

Installation / Prerequirements

Installation of docker. Check your version with:

docker -v

Installation of NVIDIA GPU drivers. Check your GPUs:

nvidia-smi

Set flag to using new buildsystem of docker (perhaps obsolete in the future).

echo \
'{
  "features": {
    "buildkit": true
  }
}' | sudo tee /etc/docker/daemon.json > /dev/null

Installation of NVIDIA Container Toolkit

Enable NVIDIA runtime

sudo nvidia-ctk runtime configure
sudo systemctl restart docker.service

Usage

Build container

sudo docker build -t introcontcudaml:latest .

Ensure the presence of the local file structure for the mount of the container

wd_IntroContCudaML/
├── config
│   └── myconfig.json
├── input
└── output
    └── model.pth (here exports the container the output)

Have a look into the docs folder for an example config file.

Run container

sudo docker run --gpus all -v ~/Projects/wd_IntroContCudaML:/IntroContCudaML/data introcontcudaml -c /IntroContCudaML/data/config/myconfig.json

option	description
--gpus all	access NVIDIA GPU resources
-v ~/path/to/wd:/IntroContCudaML/data	mounts the working directory into the container at data path
-c /IntroContCudaML/data/config/myconfig.json	custom flag for specifying the config file

see also docs for docker container run

For apptainer usage (assuming that apptainer pull was used to pull the image and convert it to a sif)

apptainer exec --nv --bind ~/path/to/wd:/IntroContCudaML/data/ --pwd /IntroContCudaML introcontcudaml_latest.sif python3 src/main.py -c data/config/myconfig.json