Deep learning beyond the learning
summary
Covers the full deep learning pipeline with concrete tooling and numbers: training hundreds to thousands of models via hyperparameter sweeps, storing model binaries on HDFS/GCS plus metadata in MongoDB/ArangoDB, and using TensorFlow Serving version policies for zero‑downtime and canary rollouts. Recommends Mesos/DC/OS or Kubernetes for GPU/TPU multi-tenancy and shows a working integration—DC/OS package install tensorflow—while noting a demo cluster spin‑up attempt encounters a URL failure. Also highlights TensorFlow Lite shrinking models to ~300KB–1MB.
AI-written summary · about
chapters (16)
- 0:00Intro
- 2:26Deep Learning: The Promise
- 5:00Deep Leaming: The Challenges
- 7:04Deep Learning: The Challenges
- 8:19Training Challenges
- 9:46Input Data Management
- 13:00Alternatives
- 20:55Resource Isolation and Allocation
- 25:29MESOS ARCHITECTURE
- 26:34Model Management
- 27:55TensorFlow Hub
- 29:17Model Serving
- 30:32TensorFlow Lite
- 32:25Rendezvous Architecture
- 33:39Monitoring
- 37:12Platforms
related talks
42mDeep learning beyond the learning
49mDistributed Deep Learning
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
54mDeep learning on AWS with NVIDIA: From training to deployment
33mDeploying and Scaling Spark ML and Tensorflow AI Models
42mDeploying ML at Scale with Kubernetes and TFX
27mTips and tricks for distributed large model training
01h 18mPutting Tensorflow Models into Production