ML in Production Tutorial

The idea to write up this repository is motivated by the very interesting set of blog posts from http://mlinproduction.com/ by Luigi Patruno. In particular, the ones about dockerizing a machine learning models:

Disclaimer: The core of this repository is based on this reference and I encourage everyone to read them before going into the code if you do not have much experience on this topic.

In addition to the main functionalities, I also wanted to complement it with extra features which are not discussed there:

Train the machine learning model with data stored in an AWS S3 bucket. In particular, when building the Docker image, do not copy the training data into it.
Describe how to build a Docker image passing the credentials as ARG variables and not as ENV variables. This is particular special for security reasons.
Unit Testing (needs more tests).
Set up GitHub Actions:
- Docker container GitHub Action
- Python (pytest) Action
- Codecov GitHub Action to get reports on tests coverage.
which get trigger on push and ensures reliability on the code.

I have also added a Resources section where I store useful references, interesting reading and similar approaches.

Remark: This repository structure should be seen as a minimal toy-use-case. You might want to package the files correctly, using for example Cookiecutter Data Science.

Deployment: This is out of the scope of this repository. There are many ways of doing this! See, for example, Deploying Elastic Beanstalk applications from Docker containers.

Contributing

I will keep adding functionalities to this toy model repository. If you have some other relevant resources, suggestions, comments or find bugs please create a Pull Request or drop me a line.

Setup

We are going to use a Makefile for the main functionalities.

Create virtual environment:

make venv

Activate environment and install python packages:

make install

Generate Data

This toy-model example fits a linear regression model of the form y ~ x. To generate sample data for this model you can run the command (writes it to data/train_data.csv

make generate_data

To train the model store the training data in a secure S3 bucket.

Train Model

To train and save the model (to model/ml_mod.joblib, see envs.py) run:

make train

The training phase assumes the data is stored in an S3 bucket with relative path: s3://S3BUCKET/ml_prod_tutorial/data/train_data.csv.

AWS Credentials

It is recommended to create an IAM User with the appropriate credentials. You can store the AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY and the AWS_REGION as environmental variables using the aws configure command from the aws cli. For this repository I am assuming the default profile.

Warning: never store your credentials in any script on a GitHub repository!

Build Docker image

To build the docker image simply run:

make docker_build

This image trains and saves the model in order to generate predictions. You can test it by generate predictions on sample data by running the command:

docker run docker-model inference.py

You can share this model image on Docker Hub or as a .tar file.

Flask App

The module apy.py defines a Flask application which enable us to generate prediction in real time. To spin up the app just run (we choose port = 5000):

Run docker container:

docker run -it -p 5000:5000 docker-model python api.py

To generate prediction for a given sample data point (here, x = 1.2 )you can run (locally):

curl -i -H "Content-Type: application/json" -X POST -d '{"x": 1.2}' 127.0.0.1:5000/predict

Docker container GitHub Action

In this step we set up a Docker container GitHub Action where we test the Docker image build on each push. Note that this requires training the model (this could or could not be ideal, depends on the application), which in particular requires access to the data. As stressed above, we do not want to store the AWS credentials on the repository. One option it to store them as GitHub Secrets. When could then access them using, for example, ${{secrets.AWS_ACCESS_KEY_ID}}.

juanitorduz / ml_prod_tutorial