AI-CLI

About

ai-cli is a command line script, developed to simplify the execution of machine learning projects. Docker containers are used to manage dependencies and isolate different projects from the host system.

ai-cli helps to manages these containers and provides functionality, like:

• Manage project dependecies in differnet docker images
• Separate users
• Jupyter Lab integration
• Run MLflow projects with one command
• Management of CML runners for differnet git repositories
• Conda support using Mamba
• GPU support with CUDA
• Service based domain names using nginx reverse proxy

Setup

Download the latest release and install with dpkg -i ai-cli.deb or clone the repository and execute sudo make install to install ai-cli system-wide. Each user has to execute ai-cli init before their first usage. To reach services online, the reverse proxy needs to be started with ai-cli start-proxy once, globally.

Step-By-Step Instructions

You need to follow Install, System-Wide and User Configuration for a working install. Tested on Ubuntu 20.04.

Install

1. Install dependencies: sudo apt-get install git make dpkg apache2-utils
2. Install docker-ce from docker website
3. Install docker-compose v2 from github (move the executable to /usr/local/bin/docker-compose, and sudo chmod +x /usr/local/bin/docker-compose for example)
4. Install ai-cli: git clone https://github.com/MArpogaus/ai-cli.git && cd ai-cli && sudo make install

Note: ai-cli requires docker daemon to be running.

System-Wide Configuration

Open /etc/ai-cli/config and edit:

• DVC_DATA (should exist)
• MLFLOW_DATA (must exist, should contain ${USER} for user destinction) User needs to have read/write permissions, e.g. using chown $USER:$USER PATH command
• CERTS_PATH (should contain .pem, .crt .key files for ssl)
• URLNAME (may be left as localhost if desired)
• DEFAULT_HOST (may be left as error.localhost if desired)

to reflect respective values of your system

Configure a first user like below and start reverse proxy: ai-cli start-proxy. This needs to be invoked in order to reach exposed web services like jupyter.

User Configuration

The following needs to be done for every new user in order to use ai-cli.

1. Every user that should run ai-cli needs to be added to the docker group to not require root. You might need to log-in again or even restart for the change to apply. (e.g. sudo usermod -aG docker $USER)
2. Every user needs to have a valid (global) git identity (email and name) set. (e.g. git config --global user.name "John Doe" and git config --global user.email [email protected])
3. Execute ai-cli init for every new user to setup workspace and build user containers.

ai-cli can now be used by the new user for example by starting a jupyter server with ai-cli lab. Make sure the reverse proxy is running as described above before trying to access jupyter.

Usage

ai-cli [OPTIONS] COMMAND [ARGS...]

Options

-v VOLUME            Specify additional docker VOLUMEs
-e FILE              Specify environment FILE
-g GPUS              Enable gpu support. Set specific GPUS, e.g. 0,1 for gpu 0 and 1.
-n NAME              Specify experiment NAME
-c CUDA              Build with cuda support. Specify CUDA version to use. Tags image with 'c<VERSION>' by default.
-i IMAGE             Specify docker IMAGE to be used.
-h                   Show help

Commands

init                 initialize this script for your user
start-mlflow         start mlflow server
stop-mlflow          stop mlflow server
restart-mlflow       restart mlflow server
start-proxy          start nginx reverse proxy
stop-proxy           stop nginx reverse proxy
status               show status of your environment
build                build image from dockerfile
bash                 start bash shell inside container
notebook             start jupyter notebook server
lab                  start jupyter lab server
runner               start cml runner for repo
exec PROGRAMM        execute PROGRAMM inside container
run PATH             run MLProject from PATH inside container
run-from-git URI     run MLProject from git repo at URI inside container
info                 show basic paths and env variables
version              show git commit hash this script is installed with
reset-password       Reset Password for mlflow you will setup a new one at next server start

Overview

ai-cli is based on docker images/containers. The list of images associated to a user can be seen with ai-cli status. The image ai-cli/:workspace is used by default. A different associated image can selected by specifying the -t flag. All below commands need to be executed as the user that should use the respective feature.

After a fresh install, each user needs to execute ai-cli init before starting to work. To use the mlflow dashboard, the mlflow server needs to be started using ai-cli start-mlflow. To use the jupyter lab environmen start the jupyter server using ai-cli lab. For debugging and local execution, a bash shell can be entered using ai-cli bash. Hooks can be injected as .sh-files for and after starting a container at /etc/ai-cli/hooks. An example is provided there. Conda environments is supported via Mamba by default and are mounted on startup.

Now ai-cli is ready to be used, examples are provided below.

Example commands

Assume a mlflow project at current working directory and a started mlflow server:

comand	comment
ai-cli lab	opens jupyter lab at https://[USER]-lab.[HOST]. Use token from command output to login.
ai-cli start-mlflow	opens mlflow server at https://[USER]-mlflow.[HOST]. Can be secured with password
ai-cli bash	open bash for local project in docker image.
ai-cli -e .env run .	with .env as default environment file, need MLFLOW_EXPERIMENT_NAME defined
ai-cli -e the_envir run-from-git	with additionally MLFLOW_GIT_PROJECT (git https url) defined
ai-cli run-from-git "https://<token>:<password>@<uri>#<path/to/MLProject (optional)> -v <branch or commit hash>	directly pass url, set name with -n NAME

Cuda Support

ai-cli is built to support tensorflow with cuda. To use cuda, a cuda-enabled image is necessary. To create a cuda image, use the build command and specify the desired cuda version using the -c switch. ai-cli -c 11.2 build.

For the cuda container to use the gpu, add the -g option when using any command to actually include gpus that cuda can use. This makes the specified gpu(s) available and switches the runtime to "nvidia". Note, that docker-nvidia-runtime needs to be set up correctly beforehand.

For example ai-cli -g 0 -c 11.2 bash opens a bash command line with inside the container, exposing gpu "0" with support for cuda 11.2. Note, that the images for cuda 11.2 need to be built beforhand as stated above. It can be verified that the gpu is used, using nvidia-smi inside the container. Note, that the cuda version shown by nvidia-smi does not necessarily represent the correct version because of differences between driver API and runtime API version. Also, the command is not installed by default.

CML Runner

ai-cli supports starting a CML runner locally for your CI/CD Workflow.

Here is how to use ai-cli runner GIT-ORIGIN-HTTPS ACCESS-TOKEN.

1. Create access token for your project (as owner, scope: api in gitlab) or personal access token.
2. Enable CI for your repository. (Settings->General->Visibility, project features, permissions->Repository->CI/CD->Save Changes in gitlab)
3. Configure pipeline (.gitlab-ci.yml, in gitlab). For the local runner to pickup a job, the job needs to have the tag ai-cli. For Example:

   test:
    tags:
     - ai-cli
    script:
     - echo "Hello World" >> report.md
     - cml comment create report.md

4. Copy the https clone link for your repository. Start your local runner using this link.

Now the runner is ready to pick up your jobs.

Troubleshooting

issue	help
run fails with error "too many 500 error responses"	Provide experiment name with -n flag. If name already used, try with new name. Names deleted on the mlflow-dashboard still count as used.

License

AI-CLI Simplifying AI Experiments
Copyright (C) 2022  Marcel Arpogaus, Julian Jandeleit

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with this program.  If not, see <https://www.gnu.org/licenses/>.

Acknowledgments

Parts of this work have been funded by the Federal Ministry for the Environment, Nature Conservation and Nuclear Safety due to a decision of the German Federal Parliament (AI4Grids: 67KI2012A).