PyTorch at Uber
summary
Uber ATG describes a single, joint perception–prediction neural network that consumes multi‑sensor time series and HD maps and outputs multiple future trajectories to capture uncertainty. Production workflow starts in Python, uses PyTorch features like TorchScript ignore annotation to keep dev-only logic, moves hot paths to custom C++ ops, leverages JIT fusion and stateful modules for latency, and debugs by serializing models (model binaries unzip as human‑readable zip) and dumping intermediate activations.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:17THE SELF-DRIVING PROBLEM
- 0:33END TO END APPROACH
- 0:56TRADITIONAL STACK
- 1:54UBER ATG VISION
- 2:10OUR RESEARCH AND PRODUCTION MODELS
- 2:22PERCEPTION & PREDICTION NETWORK
- 3:14TAKING A COMPLEX MODEL TO PRODUCTION
- 5:02COMPLEX COMPUTATION GRAPH
- 5:20DATA STRUCTURES
- 7:05CONTROL FLOW
- 7:16CUSTOM
- 8:09USER DEFINED CLASSES
- 8:41FUSION
- 9:47PRODUCTION DEPLOYMENT WITHOUT COMPROMISING FLEXIBILI
- 10:07PRODUCTION-READY TOOLING
- 11:14SERIALIZATION
- 12:18HUMAN-READABLE FROZEN GRAPHS
related talks
31mDeep learning for self-driving vehicle technology
35mUber's Journey in Building Multi-Modal Foundation Models
17mKeynote: Machine Learning & A.I. At Uber
26mUber’s Intelligent Insights Assistant
18mAI Platform Q&A
22mIncreasing the Impact of AI Through Better Software
33mScaling PyTorch Training with Ray Data on Uber's Michelangelo AI Platform
33mUber's GenAI Leap: Batch Predictions Using Ray and vLLM