Dokai

Collection of Docker images for ML/DL and video processing projects.

Overview of cpu.* images

Several types of images are presented:

• core: Python (3.11.9)
• ffmpeg: FFmpeg (6.1)
• base: Python ML and CV packages listed here
• pytorch: PyTorch (2.1.2), torchvision (0.16.2), torchaudio (2.1.2) and torch based libraries listed here

Overview of gpu.* images

Several types of images are presented:

• core: CUDA (12.2.2), cuDNN (8.9.6), Python (3.11.9)
• ffmpeg: FFmpeg (6.1) with NVENC/NVDEC support
• base: Python ML and CV packages listed here
• pytorch: TensorRT (9.2.0), PyTorch (2.1.2), torchvision (0.16.2), torchaudio (2.1.2) and torch based libraries listed here
• video: VideoProcessingFramework (2.0.0) and Tensor Stream (0.5.0) for full HW acceleration of video decoding

Supported NVIDIA architectures: Pascal (sm_60, sm_61), Volta (sm_70), Turing (sm_75), Ampere (sm_80, sm_86), Ada Lovelace (sm_89), Hopper (sm_90).
You can identify the architecture of your GPU here.

Overview of gpu.*.opt images

There are also optimized GPU images based on nvidia-runtime images, which have the same packages and libraries installed as stated above. To install them add .opt suffix at the end of the docker image name, e.g. dokai:gpu.pytorch.opt.

<details><summary>Difference between `nvidia-devel` and `nvidia-runtime` images</summary> <p> Nvidia's CUDA images come in three flavors and are available through the NVIDIA public hub repository.

• base: starting from CUDA 9.0, contains the bare minimum (libcudart) to deploy a pre-built CUDA application. Use this image if you want to select which CUDA packages you want to install manually.
• runtime: extends the base image by adding all the shared libraries from the CUDA toolkit. Use this image if you have a pre-built application using multiple CUDA libraries.
• devel: extends the runtime image by adding the compiler toolchain, the debugging tools, the headers, and the static libraries. Use this image to compile a CUDA application from sources.

Runtime and devel are the ones that are mostly used and the differences between them are the image sizes (~3Gb), presence of compilers, and debugging tools.

</p> </details>

Overview of *.rootless images

Initially, all Dokai images were root-based, meaning that they had root privileges inside of the docker container. According to the Docker best-practices it is better to run docker containers without sudo rights with explicitly set user. Because of that *.rootless images were added to duplicate all existing images, but without sudo privileges. The name of the user is set to dokai and it has 1000 set as UID/GID.

Sooner or later these images will fully replace root-based ones as they are more secure.

Examples

Pull an image from the GitHub Container registry

docker pull ghcr.io/osai-ai/dokai:24.06-gpu.pytorch

Docker Hub mirror

docker pull osaiai/dokai:24.06-gpu.pytorch

Check available GPUs inside the container

docker run --rm \
    --gpus=all \
    ghcr.io/osai-ai/dokai:24.06-gpu.pytorch \
    nvidia-smi

In case of *.rootless images there may permission problems happen within docker container:

• If the source code is copied to the container with the command COPY, make sure to change the owner of the copied files and folders with additional flag, for example:

FROM dokai:gpu.video.opt.rootless

WORKDIR /home/$UNAME/workdir
COPY --chown=$UID:$GID ./ ./

...

• If the source code is mounted as a volume with the flag -v make sure to mount it to the correct user home directory path, and --user flag is passed, for example like this:

docker run --rm -it \
    --user=$(id -u):$(id -g) \
	-v $(shell pwd):/home/dokai/workdir \
	ghcr.io/osai-ai/dokai:24.06-gpu.pytorch.opt.rootless \
	bash

Or make sure that the mounted volume has enough permission for the dokai:1000:1000 user to work.

The result of above commands strongly depends on the local setup, although it should work in most cases.

Package versions

CPU images

<details><summary>dokai:24.06-cpu.core</summary> <p>

ghcr.io/osai-ai/dokai:24.06-cpu.core

Image based on python:3.11.9-slim-bookworm which includes:

• Debian Bookworm Operational System;
• Python 3.11.9;
• etc..

On top of that packages are installed, here's a brief description and purpose of each:

Utilities:

• nano, vim, tmux are text editors and terminal multiplexers used for editing files and managing terminal sessions.
• unzip is a utility used to decompress ZIP archives, commonly used for extracting files from compressed archives.
• git is a distributed version control system used for tracking changes in source code during software development.
• wget, curl are command-line tools used for downloading files from the internet, commonly used in scripts and automation tasks.
• htop is interactive system monitoring tool used for displaying system resource usage.
• sysstat collects and reports system performance statistics, including CPU, memory, and disk usage.

</p> </details> <details><summary>dokai:24.06-cpu.ffmpeg</summary> <p>

ghcr.io/osai-ai/dokai:24.06-cpu.ffmpeg

Image based on dokai:24.06-cpu.core:

Additionally, installed:

• FFmpeg (n6.1)

</p> </details> <details><summary>dokai:24.06-cpu.base</summary> <p>

ghcr.io/osai-ai/dokai:24.06-cpu.base

Image based on dokai:24.06-cpu.ffmpeg:

Additionally, installed:

• Python ML and CV packages: requirements.txt

</p> </details> <details><summary>dokai:24.06-cpu.pytorch</summary> <p>

ghcr.io/osai-ai/dokai:24.06-cpu.pytorch

Image based on dokai:24.06-cpu.base:

Additionally, installed:

• Pytorch, torch-audio, torch-vision;
• Pytorch-related packages: requirements.txt

</p> </details>

GPU images built from source

<details><summary>dokai:24.06-gpu.core</summary> <p>

ghcr.io/osai-ai/dokai:24.06-gpu.core

Image based on nvidia/cuda:12.2.2-cudnn8-devel-ubuntu22.04 which includes:

• Ubuntu 22.04 Operational System;
• CUDA (12.2.2);
• cuDNN (8.9.6);
• NVCC;
• etc..

On top of that packages are installed, here's a brief description and purpose of each:

Python (3.11.9):

• python3.11 is a programming language used for scripting and software development.
• python3.11-dev contains header files and development tools for building Python extensions.
• python3.11-distutils contains utilities for distributing and installing Python packages.
• python3.11-tk contains Tkinter, a GUI toolkit for Python.

Builders, configurators and compilers:

• build-essential package contains essential tools required for building software on Ubuntu, including compilers (gcc, g++, etc.) and build-related tools.
• Yasm and nasm are assemblers used for compiling assembly language code into machine-readable format, commonly used in building software.
• Ninja is a small build system that is fast and efficient, used as an alternative to GNU Make in some projects.
• cmake is a cross-platform build system generator used to control the software compilation process using simple platform-independent configuration files.
• pkgconf is a package configuration system used to help locate and configure libraries needed for building software.
• gfortran is a GNU Fortran compiler used for compiling Fortran code.
• autoconf and libtool are tools for generating configure scripts and managing software builds on various platforms.

Utilities:

• nano, vim, tmux are text editors and terminal multiplexers used for editing files and managing terminal sessions.
• unzip is a utility used to decompress ZIP archives, commonly used for extracting files from compressed archives.
• git is a distributed version control system used for tracking changes in source code during software development.
• wget, curl are command-line tools used for downloading files from the internet, commonly used in scripts and automation tasks.
• htop and nvtop are interactive system monitoring tools used for displaying system resource usage.
• sysstat collects and reports system performance statistics, including CPU, memory, and disk usage.

Libraries:

• libsm6, libxext6, libxrender1, libgl1-mesa-glx libraries for X Window System used for rendering graphics and managing graphical applications.
• libtcmalloc-minimal4 is a memory allocation library, providing optimized memory management functions.
• libx264-dev is a library for encoding H.264 video streams.
• libsndfile1 is a library for reading and writing audio files.
• libssl-dev is a development package for OpenSSL, providing cryptographic functions.
• libpng-dev and libjpeg-dev are development libraries for handling PNG and JPEG image formats, respectively.
• libmp3lame-dev is a development library for encoding and decoding MP3 audio files.
• liblapack-dev is a library for numerical linear algebra functions.
• libopenblas-dev is an optimized BLAS (Basic Linear Algebra Subprograms) library.