05 · Video Pipeline Design

Part III group exercise of Tensors for Machine Learning. A video pipeline is the path from a video file to the tensor a model actually receives:

raw file → decoded frames → preprocessed batch → model input → model output

The key idea:

Every pipeline decision chooses what information is kept, transformed, or discarded.

Live, the room works notebook 05’s core — tracing sampled frames back to the source frame each came from — then group task 05, keep the event, fit the budget: at most 16 stored frames per clip, and a brief event that must survive. Designing the whole path for two systems at once — a short-video recommender that computes one embedding per video, and a surgical-video model that labels the current phase of an operation — is the take-home version. There is no single correct answer.

Study here, then run the real thing: notebook 05 on Colab · group task 05 · workshop home

Built from the workshop’s own material by Ravi Kalia and Sebastian Laverde Chunza, CC BY 4.0.