How Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery
summary
Dendra runs PyTorch on Ray SGD and Ray Tune to scale from a laptop to dozens of GPUs in one Python script, and reports Ray Tune saves ~$16k through aggressive early-stopping. In production Ray Serve actors initialize S3 client actors co‑located with model replicas and use eager S3 fetching to keep network and GPUs saturated, enabling throughput from 10k to 100M images. Anyscale’s SDK provides programmatic cluster hot-starts from BitBucket CI and a microservice API to launch jobs automatically.
AI-written summary · about
chapters (7)
- 0:00The problem of restoration
- 2:25Dendra's technical approach
- 5:02Engineering at Dendra
- 6:46Model training with Ray Tune
- 9:07Inference and model serving
- 12:46The Anyscale advantage
- 17:22Automation and conclusion
related talks
30mScaling Machine Learning at Discord: Our Journey with Ray and Anyscale
31mFaster Model Serving with Ray and Anyscale
32mAnyscale's Ray Data: Revolutionizing Batch Inference
28mScaling Multimodal Data and Reinforcement Learning with Ray
31mManaging Large-Scale Ray Deployments: Cloud, On-Prem, and Kubernetes
35mOperating a Production-Scale Physical AI Data Platform with Ray
28mScaling AI & Machine Learning Workloads With Ray on AWS, Kubernetes, & BERT
22mRay, a Unified Distributed Framework for the Modern AI Stack