Computer Vision Dog Breed Classification with Convolutional Neural Networks, TensorFlow and Kubeflow
Talkat Kubernetes AI Day Europe 2022 · on Monday, 2022-05-16· at 12:05· Pavilion 3, Room D | Level 2 | Central Forum·
summary
A Jupyter notebook pipeline uses TensorFlow, transfer learning with ResNet-50 82% and VGG16 72%, plus a small CNN trained from scratch (much lower accuracy), to classify dog breeds from an Udacity-style dataset. Kubeflow (deployed as a service via Arrikto's environment) runs the pipeline; Kale snapshots volumes, converts annotated notebook cells into KFP steps and persists environments. The demo shows parallel training of three models, a model reveal against DNA results, and claims ~70% faster iteration when using Kale.
AI-written summary · about
chapters (12)
- 0:00Start
- 1:38Machine Learning Workflow
- 3:33Keyflow Pipelines
- 4:24Main Features
- 10:25Keyflow Central Dashboard
- 11:30Demo
- 11:50The Dog Breed Classification Problem
- 12:21Convolutional Neural Networks
- 13:27Tensorflow
- 13:51Resnet50
- 22:40Execution Graph
- 27:04Is It Possible To Predict a Mixed Breed
related talks
37mTowards Continuous Computer Vision Model Improvement with Kubeflow
53mTransforming your Jupyter Notebook to a Reproducible Machine Learning Pipeline using Kale
46mAccelerating Machine Learning DevOps with Kubeflow
32mYou Can Do Deep Learning!
01h 29mComputer Vision: An (Un?)Expected Journey, with Keras and Tensorflow
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
15mKeynote: Machine Learning on Kubernetes Made Easy With Kubeflow
10mKeynote: Machine Learning using Kubeflow and Kubernetes