Modern, typed Python for (multimodal) ML
summary
Argues that Python type hints are now practical in ML and shows concrete tooling: mypy and beartype for checks, Pydantic+FastAPI for APIs, jaxtyping for tensor shapes, and DocArray to represent, send and store multimodal data. Builds a podcast demo that loads audio/text URLs, produces 512‑dim embeddings, indexes them with HNSWlib and exposes endpoints via FastAPI (also supports gRPC/protobuf through Jina). Claims broad compatibility across PyTorch/NumPy/JAX and shows a working demo with a few hiccups. DocArray integrates with FastAPI and HNSWlib and jaxtyping covers tensor shapes across frameworks.
AI-written summary · about
chapters (10)
- 0:00Introduction to the topic
- 1:22Duck typing in Python
- 5:10Understanding type hints
- 7:15History of Python typing
- 11:10Modern typing ecosystem
- 17:14ML specific typing challenges
- 26:24Representing data with DocArray
- 40:49Sending data with DocArray
- 44:20Storing data in vector DBs
- 51:02Conclusion and Q&A
related talks
39mModern typed python: dive into a mature ecosystem from web dev to machine learning
38mMigrating existing codebases to using type annotations
57mWhy You Should Care About Types
47mThe State of the Art for Probabilistic Programming
49mKeynote 4: Here Come The Robots - Python and Machine Learning
35mCloud-Native MLOps Platform for Python- and ML Engineers
02h 24mMagical NumPy with JAX
29mBuilding Machine Learning demos with Python