From Experimentation to Products: The Production Machine Learning Journey
summary
Describes how Google moves experiments to production using TFX + Kubeflow pipelines and Apache Beam for distributed processing, with concrete components: ExampleGen, StatisticsGen, SchemaGen, Transform, Trainer, Evaluator, InfraValidator, Pusher and BulkInferrer. Claims model code is only ~5% of production code, demands metadata-backed pipelines, orchestration and continuous monitoring, and adds continuous training to automatically collect data, retrain and redeploy models; shows this stack scales to very large training sets and traffic.
AI-written summary · about
chapters (22)
- 0:00Intro
- 2:15Production ML
- 5:41We need MLOps
- 6:21Continuous integration, deployment and testing
- 7:29MLOps level 0: Manual Process
- 9:02Experiment
- 12:11Tales from the trenches
- 13:02TensorFlow Extended (TFX)
- 14:28TFX production components
- 16:43What is a TFX component?
- 18:20TFX orchestration
- 19:16Difference between TFX & Kubeflow pipelines
- 23:00Distributed pipeline processing: Apache Beam
- 25:28TFX standard components
- 25:53Components: ExampleGen, StatisticsGen & SchemaGen
- 28:17Components: ExampleValidator, Transform & Trainer
- 31:45Components: Tuner, Evaluator & InfraValidator
- 32:51Components: Pusher & BulkInferrer
- 33:37TFX pipeline nodes
- 34:43TRFX custom components
- 36:09Very high level architecture
- 37:03Outro
related talks
38mDistributed Processing for Machine Learning Production Pipelines
38mTensorFlow Extended (TFX): Machine Learning Pipelines and Model Understanding
42mDeep learning beyond the learning
22mFrom experiment to production - a journey of a machine learning model
35mHow to deploy machine learning models to production (frequently and safely)
08mManage MLOps and deploy Machine Learning to production with the new and improved TFX
27mHow to make ML teams 80% more productive.
31mProductionizing ML with ML Ops and Cloud AI