sanjeevphd/demlopment-example

DeMLopment - putting ML in Software Development. Very much a work in progress, so check back often! :)

An example of a full stack, production grade, machine learning engineering.

Overview

Let me break down the one-liner above.

• full stack - all the way from problem understanding, data mgmt., model development, to deployment, monitoring and retraining
• production grade - it has to add value, in a business sense and can grow and scale with the business
• machine learning engineering - employ best practices rooted in software engineering principles
• example - demonstrate with a concrete use case, which might be trivial but captures the essence of process

Approach/Methodology

• start at the end
• test-driven development
• agile mindset - small changes, rapid iterations, steady progress
• cloud-ready local development
• automate the automatable

Note

• This is an active project and very much a work in progress, so I expect things to change often.
• The documents are raw, unedited, and weakly formatted. Please excuse any typos, etc.

Usage

There are several offerings here.

The Docker Way

This is the no hassle offering. The only requirement is Docker.

docker run -it demlopment:latest

Note

Update to the Docker command

The Docker Dashboard should reveal the container orchestration behind the scene that reflects the three major domains that come together to make the magic happen.

1. An inference endpoint that uses Torchserve for model managment, metrics, deployment, A/B testing, etc.
2. A development environment for model development. Everything from data loading, splitting, training, validation, hyperparameter tuning, experiment tracking to testing and model selection
3. A data management environment for ingesting, storing and transforming data

Any or all of the three domains can be cloud-native or local/on-premisis depending on the use case. At reasonable scales, it is "reasonable" to assume that all domains are cloud native. My view is that as a data scientist, ML engineer, DL scientist, most time is spent in the model domain and a local development makes perfect sense, while outsourcing the data management and model deployment to the cloud. This allows for faster iterations and even potentially avoids some yak shaving.

Local Installation

• Clone the GitHub repo.

  git clone <repo_url>

• Install the dependencies and the package in a local virtual environment.

  make venv_pip

  Note

  Update requirements to include pytest.

• Run tests

  python -m pytest tests

Documentation

Build docs locally by running: make html (requires Sphinx. The finished HTML will be located under _build/docs/index.html.