Catching Tensor Shape Errors without Running Your Code
summary
Shows how to represent tensor shapes (e.g., H×W×B) as explicit Python type annotations called “shape types” and use a static type checker to detect shape mismatches without executing code. Covers IDE integration for viewing tensor shapes, gradual adoption in existing ML projects, support for broadcasting (NumPy, PyTorch, etc.), and the limitations of shape types.
AI-written summary · about
related talks
28mOn Tensors
11mPyTorch 2.0: Dynamic Shapes Support
24mWhat's New for Dynamic Shapes in PyTorch 2.1
08mPyTorch Front-End Features: Named Tensors and Type Promotion
35mTowards Typesafe Deep Learning in Scala
22mShape Recognition with Computer Vision
33mTeIL: a type-safe imperative Tensor Intermediate Language
04mInterview