Towards Continuous Computer Vision Model Improvement with Kubeflow
Talkat KubeCon + CloudNativeCon North America 2019 · on Tuesday, 2019-11-19· at 15:20· Room 31ABC - San Diego Convention Center, Upper Level·
summary
Snapchat uses Kubeflow Pipelines to close the loop: seed images from Scan and logs query a visual Our Story image index with nearest-neighbor search, crop high-activation regions, send candidates for human labeling, merge labels into TFRecords, retrain the classifier and deploy if evaluation passes. Pipelines handle cross-cloud transfers and Dataflow jobs; current tooling gaps lead to hacks like an artifact-caching decorator based on image-tag UUIDs and a shared kfp schema package for component type guarantees.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:29Agenda
- 1:52Snapchat by the Numbers
- 3:40Examples of Community Lenses
- 5:11Scan Backend Simplified
- 7:07So this is simple then? Deep Learning
- 8:33Your model is only as good as your data
- 13:23Demo: Querying Similar Images
- 15:28Backup - Demo Failed
- 17:45Requirements for Pipeline Orchestrators
- 20:14Pipeline Orchestrator Comparisons (Late 2018)
- 22:44Pipelines Overview
- 26:01From our talk proposal...
- 27:41Artifact Caching
- 32:01Typing in KFP Components
related talks
22mMLOps at Snapchat: Continuous Machine Learning with Kubeflow & Spinnaker
29mComputer Vision Dog Breed Classification with Convolutional Neural Networks, TensorFlow and Kubeflow
27mKubeflow for enabling AI Powered Drug Discovery and development...
46mAccelerating Machine Learning DevOps with Kubeflow
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
34mManaging Machine Learning in Production with Kubeflow and DevOps
31mKubeflow 2026 State of the Union: Orchestrating the Full ML Lifecycle
54mScaling Machine Learning Pipelines in Cloud