MLOps

Introduction to MLOps

MLOps (Machine Learning Operations): is a practice that aims to bring together the development and operation of machine learning models in a collaborative and efficient manner. It involves applying agile software development and DevOps practices to the development, deployment, and management of machine learning models in a production environment. MLOps aims to improve the speed, quality, and reliability of machine learning model development and deployment, while also making it easier to manage the end-to-end lifecycle of machine learning models.

ML Solution can be bad than traditional solution

• introduce developing complexity to the problem

• introduce data dependency

• introduce additional cost

• might not be reliable as traditional solution

• hard to analyze the effect of ML models on user behavior

• is it ethical

• developing dataset mismatch production dataset

• metrics doesn't reflect what we need

Stacks

Stack:

• Gradio: Javascript interface for model inference. Other choices are Streamlit, PyScript.

• Gantry: data analysis for ML user feedback and active learning based on them

• Google Colab: jupyter notebook IDE (you can use util?? to exam libraries)

• AWS Lambda: serverless

• Docker and Kubernetes: automatic deploy, virtualization. Non-experts can use AWS EKS or Google Kubernetes Engine

• Pytorch Lightning: highlevel training engine wrapper around Pytorch. Other might choose HuggingFace, Keras, Flax, JAX

- Weights&Bias: experiment analysis, artifact tracking, and versioning. Other might choose MLFlow, Tensorboard (but no tracking and versioning)