Seminario MLOps Máster Ciencia de Datos

First Hands On

Install depencendies

• Install conda: https://conda.io/docs/user-guide/install/index.html#

• Install MLflow:

pip install mlflow

Get the Code

Fork the code from:

https://github.com/anllogui/master_ciencia_datos_21

Get the code:

git clone https://github.com/<user>/master_ciencia_datos_21.git

There are 3 main folders:

• training: code and data needed for training
• exploitation: code to deploy de model
• mlflow_server: folder to store mlflow data

We will start with the training folder:

• model: notebook for training the model
• flaskr: service with the model embedeed
• tests: service testing

Configure environment

Create environment

In the training folder type:

conda env create -f environment.yml
conda activate master_ciencia_datos_21

Train the Model

Execute Notebook

• Run Jupyter

jupyter notebook

• Go to "nb/simple_regression.ipynb".
• Execute Notebook

Review results:

• Exexute MLFlow ui:

cd nb
mlflow ui

• go to: http://127.0.0.1:5000

Second Hands On

Expose the model

The service is developed in "master_ciencia_datos_21/flaskr/linreg.py".

To run the service:

Create environment:

conda deactiate 
conda env create -f environment.yml
conda activate master_ciencia_datos_21_flask

Mac/Linux:

cd ..
export FLASK_APP=flaskr
export FLASK_ENV=development
pip install -e .
flask run

Windows:

cd ..
set "FLASK_APP=flaskr"
set "FLASK_ENV=development"
pip install -e .
flask run

To test the service:

curl -i -H "Content-Type: application/json" -X POST -d '{"yearsOfExperience":8}' http://localhost:5000/

Third Hands On

Automatize

Automatize Model Training and Versioning

papermill simple_regression.ipynb output.ipynb -p data_ver 1 -p model_ver 1

Continuous integration

Jenkins

Install Jenkins for ubuntu: https://www.jenkins.io/doc/book/installing/linux/#debianubuntu

After installing Jenkins:

sudo service jenkins start

To access to jenkins: http://localhost:8080

Install Anaconda for all users and give permissions:

sudo addgroup anaconda
sudo chgrp -R anaconda /opt/anaconda3
sudo adduser jenkins anaconda
sudo chmod 777 -R /opt/anaconda3

Automated Training in Jenkins

#!/bin/bash
echo "---- SETING ENVS ---- "

export PATH=$PATH:/opt/anaconda3/
PYENV_HOME=$WORKSPACE/venv/
export MLFLOW_TRACKING_URI="http://127.0.0.1:5000"

cd training

echo "---- GETING PROPERTIES ----"

file="./train.properties"

if [ -f "$file" ]
then
  echo "$file found."

  while IFS='=' read -r key value
  do
    key=$(echo $key | tr '.' '_')
    eval ${key}=\${value}
  done < "$file"

  echo "Model Version = " ${model_version}
  echo "Data Version  = " ${data_version}
else
  echo "$file not found."
fi

echo "---- CLEANING ENVIRONMENT ----"
if [ -d $PYENV_HOME ]; then
    echo "- Project exists: cleanning.."
    rm -Rf $PYENV_HOME 
fi
source /opt/anaconda3/etc/profile.d/conda.sh
echo "*** creating env ***"
echo $PYENV_HOME
conda env create -f environment.yml --prefix $PYENV_HOME
conda activate $PYENV_HOME
cd nb
papermill simple_regression.ipynb output.ipynb -p data_ver ${data_version} -p model_ver ${model_version}

ls -la ../../models

#push to nexus. comment if not installed
curl -v -u admin:admin -X POST 'http://localhost:8081/service/rest/v1/components?repository=maven-releases' -F "maven2.groupId=models" -F "maven2.artifactId=simple_regresion" -F "maven2.version=${data_version}.${model_version}" -F "maven2.asset1=../../models/linear_regression_model_v${model_version}.pkl" -F "maven2.asset1.extension=pkl"

Nexus

(This step is optional) Install Nexus for Ubuntu: https://medium.com/@everton.araujo1985/install-sonatype-nexus-3-on-ubuntu-20-04-lts-562f8ba20b98

Start MLFlow as a server

Create Mlflow Project: mkdir mlflow_server conda create -n mlflow_server python=3 conda activate mlflow_server pip install mlflow

mlflow server --backend-store-uri sqlite:///mydb.sqlite --default-artifact-root /Users/anllogui/dev/mlflow_server/artifacts

Train Shell

#!/bin/bash
echo "---- SETING ENVS ---- "

export PATH=$PATH:/home/anllogui/anaconda3/bin
PYENV_HOME=$WORKSPACE/venv/
export MLFLOW_TRACKING_URI="http://127.0.0.1:5000"

cd training

echo "---- GETING PROPERTIES ----"

file="./train.properties"

if [ -f "$file" ]
then
  echo "$file found."

  while IFS='=' read -r key value
  do
    key=$(echo $key | tr '.' '_')
    eval ${key}=\${value}
  done < "$file"

  echo "Model Version = " ${model_version}
  echo "Data Version  = " ${data_version}
else
  echo "$file not found."
fi

echo "---- CLEANING ENVIRONMENT ----"
if [ -d $PYENV_HOME ]; then
    echo "- Project exists: cleanning.."
    rm -Rf $PYENV_HOME 
fi
source /home/anllogui/anaconda3/etc/profile.d/conda.sh
echo "*** creating env ***"
echo $PYENV_HOME
conda env create -f environment.yml --prefix $PYENV_HOME
conda activate master_ciencia_datos_21
cd nb
papermill Simple_Regression.ipynb output.ipynb -p data_ver ${data_version} -p model_ver ${model_version}

ls -la ../models

curl -v -u admin:admin -X POST 'http://localhost:8081/service/rest/v1/components?repository=maven-releases' -F "maven2.groupId=models" -F "maven2.artifactId=simple_regresion" -F "maven2.version=${data_version}.${model_version}" -F "maven2.asset1=../models/linear_regression_model_v${model_version}.pkl" -F "maven2.asset1.extension=pkl"

Docker

• Build and run training:

cd training
docker build -t training .; docker run --name=training -v ~/../output:/output --rm --network host training

• Build and run execution:

cd exploitation
docker build -t model-exploitation .; docker run -p 127.0.0.1:5000:5000 model-exploitation

Docker compose

mlflow server:

docker-compose run --service-ports mlflow_server

training:

docker-compose run --service-ports training

all:

docker-compose up --build

• Delete old images:

docker system prune -a

• Connect to a container:

docker exec -it <container name> /bin/bash