Docker Image for ML with PyTorch

Deprecation

ml-pytorch is no longer being devleoped and will be removed. Please switch to docker-base-ml.

Use cases

You should use this image if you

need to run PyTorch on an Nvidia CUDA 11.8.0 GPU device
want to use fish shell
don't want to re-install a virtualenv everytime
want to run large ML training job on a remote server via SSH
want to deploy a base image to a cloud computing provider (eg: runpod.io)

You shouldn't use this image

if you want to run PyTorch on Apple Silicon
in an otherwise unsecured environment (because it adds a sudo-privileged user with plaintext password)
if you want to use bash shell
if you want to run jupyter notebook
if you need a pinned python virtualenv

Features

`Dockerfile`

This image has the following installed

git, rsync, ssh, and other Linux utilities
fish is set to default for the user neo (see below)
python3 and pip
python3.9 and pip
updated conda base environment
Virtualfish is available as vf
A vf environment called pytorch which contains
- numpy, pandas, matplotlib, seaborn
- scikit-learn, scipy
- torch, torchtext
- transformers, datasets, sentencepiece
- wandb, wbutils
- ipython, jupyter
- pytest, coverage, and more

Use pytorch virtualenv as follows

# You should be in the (default) fish shell (not bash)
vf activate pytorch
ipython
# Run your python code
# Exit ipython
vf deactivate

User

This image creates a sudo-privileged user for you. You should be able to do everything (including installing packages using apt-get) as this user without having to become root.

Key	Value
Username	`neo`
Password	`agentsmith`

Get the image

From Docker Hub

docker pull ankurio/docker-base-ml:latest
docker run -it ankurio/docker-base-ml:latest /usr/bin/fish

Clone from GitHub Packages

# You need to login even though the image is public
docker login docker.pkg.github.com
Username: # type your GitHub username
Password: # type your GitHub password (not token)
# Login Succeeded  <- Success!

# Now, you can pull the image
docker pull ghcr.io/ankur-gupta/docker-base-ml:latest
docker run -it ghcr.io/ankur-gupta/docker-base-ml /usr/bin/fish

Build it yourself

Alternatively, you can clone the repository and build it yourself. It will take ~10 minutes to build the first time.

# Clone the repository
git clone git@github.com:ankur-gupta/docker-base-ml.git

# From $REPO_ROOT
cd $REPO_ROOT
docker build . -t docker-base-ml
docker run -it docker-base-ml /usr/bin/fish

Usage

# This command assumes that you pulled the image from DockerHub. Modify if you got the image
# in other ways.
docker run -it ankurio/docker-base-ml:latest /usr/bin/fish

# Execute these within the docker container
# Use the installed `pytorch` virtualenv
neo@5e50bd3d637b ~> vf activate pytorch
neo@5e50bd3d637b ~> (pytorch) ipython
# Python 3.10.12 (main, Jun 11 2023, 05:26:28) [GCC 11.4.0]
# Type 'copyright', 'credits' or 'license' for more information
# IPython 8.14.0 -- An enhanced Interactive Python. Type '?' for help.
In [1]: import torch
In [2]: torch.zeros(3)
Out[2]: tensor([0., 0., 0.])
# Run your pytorch code here
# ...
# ...

# Exit ipython
exit  # or press Control+D

# Deactivate virtualenv
vf deactivate

# Exit the docker container
exit  # or press Control+D

Docker tips

Debugging build errors

A RUN echo "print something" line in Dockerfile will be useful if you use "plain progress"

docker build . -t docker-base-ml --progress=plain

Clean all containers and images

This does not delete the docker cache. From this StackOverflow post.

# Delete all containers - do this first!
docker rm $(docker ps -a -q)

# Delete all images
docker rmi $(docker images -q)

No space left on device

If you encounter this type of error,

ERROR: failed to solve: failed to register layer: write /usr/local/cuda-11.8/targets/sbsa-linux/lib/libcublasLt_static.a: no space left on device

it means that you may need to free disk space. You can do it using these commands.

docker system prune -a  # this will delete the cache
# WARNING! This will remove:
#   - all stopped containers
#   - all networks not used by at least one container
#   - all images without at least one container associated to them
#   - all build cache

# Are you sure you want to continue? [y/N] y
# Deleted build cache objects:
# kukqetqe618pmdgf5vz1npyn6
# ...
# gdculirrh0mr1g2x2qa7llunk
# Total reclaimed space: 21.19GB

docker volume prune
# WARNING! This will remove anonymous local volumes not used by at least one container.
# Are you sure you want to continue? [y/N] y
# Total reclaimed space: 0B

docker network prune
# WARNING! This will remove all custom networks not used by at least one container.
# Are you sure you want to continue? [y/N] y

You can check docker's space usage like this

docker system df
# TYPE            TOTAL     ACTIVE    SIZE      RECLAIMABLE
# Images          2         1         7.192GB   7.192GB (100%)
# Containers      1         0         1.621kB   1.621kB (100%)
# Local Volumes   0         0         0B        0B
# Build Cache     88        3         11.83kB   10.54kB