Build a Production ML System with only Python on Free Serverless Services

The video is a lecture by Professor Jim Dowling, an associate professor at the KTH Royal Institute of Technology. He discusses the process of building a production machine learning system using only Python. The lecture is based on a free course he ran called “Serverless ML“. Here are the key points from the video: The Goal: The aim is to build a production machine learning system end-to-end using only Python. The system will be a serverless machine learning system that runs on a generous free tier of various services. The System: The system consists of three main components: state (features, labels, models), compute (using Hopsworks for state and Modal for compute), and a user interface (using Streamlit or Hugging Face). Data Scientists’ Role: Dowling argues that the role of data scientists is evolving. It’s no longer enough to just train models; data scientists should be involved in the entire process of building machine learning systems. This includes working with static data, batch data, and real-time data. Building the System: The system is built by decomposing the problem into three independent parts: a feature engineering pipeline, a model training program, and an inference program. Each of these pipelines can be worked on independently and can even be handled by different teams. Minimal Viable Prediction Service: Dowling suggests that data scientists should aim to build a minimal viable prediction service as quickly as possible and then iteratively improve it. This service should start with a problem that needs to be solved and should be presented to stakeholders as a living model in its intended ecosystem. Feature Pipelines: Dowling demonstrates how to write feature pipelines using Python and pandas. He also discusses the importance of validating data before it’s written to the feature store. Training and Inference Pipelines: The training pipeline reads features and labels, trains a model, and writes the model back to the platform. The inference pipeline downloads the model and the features to make predictions. Air Quality Prediction System: The lecture includes a demonstration of an air quality prediction system that Dowling and his students built. The system collects data from various locations, predicts air quality for the next week, and compares air quality between different locations. Use of Tools: Dowling uses tools like Hopsworks, Modal, Streamlit, and Hugging Face to build the system. He also uses GPT-4 to accelerate the development of feature pipelines. Future of Data Science: Dowling concludes by discussing the future of data science. He believes that as tools and technologies continue to evolve, it will become easier for data scientists to build complex machine learning systems. PUBLICATION PERMISSIONS: PyData Organizers provided Coding Tech with the permission to republish PyData talks on this channel. CREDITS: PyData YouTube channel: @PyDataTV