How to customize Machine Learning models the simple way
summary
Shows how to customize a pretrained image model using Model Maker and TensorFlow Hub inside Google Colab, fine-tuning a feature vector on the Flowers dataset (~3,000 images, 5 classes). A one-line create trains for 5 epochs to ~95% accuracy (validation) and exporting to TFLite yields ~92% on-device accuracy; enabling finetuning makes train accuracy jump near 99% but increases epoch time (from ~12s to ~40s). Exports include TFLite and FP16-quantized variants.
AI-written summary · about
chapters (17)
- 0:00Start
- 0:54What's Machine Learning
- 1:46What Is Transfer Learning
- 3:19Feature Extraction
- 4:13Google Collab
- 5:06Install Model Maker
- 7:08Import Packages
- 10:05Splitting Our Data in Train and Test
- 11:13Create a Model
- 13:47Evaluate Our Model
- 14:36Export Our Model
- 28:59Using Modules from Tensorflow Hub
- 30:35Tensorflow Hub
- 33:01Feature Vector
- 39:05Input Size
- 40:32Data Augmentation
- 47:25Fine Tuning
related talks
21mTransform your web apps with Machine Learning
02h 41mHello TensorFlow: 3 Workshops to Get Started with TensorFlow 2.0
12mCross platform computer vision made easy with Model Maker
46mAccelerating Machine Learning DevOps with Kubeflow
42mCreating Smarter Applications
30mTensorFlow for Mobile Heroes
01h 39mGoogle Colaboratory and Python: Deep Learning basics for CV tasks
29mPushing Machine Learning into your app When TensorFlow and MLKit rule the world