Putting the Fast in fast.ai - Machine Learning Models in 15 mins
summary
Fastai plus transfer learning builds an image classifier from scratch using ResNet34 pretrained on 14M images and an image databunch sourced from Google Images and cleaned automatically. The demo uses about ~800 images total across three classes, applies optional data augmentation, trains with fit_one_cycle and learning rate scheduling, and reaches about 98% accuracy (2.4% error then tuned to ~1.9%). The model exports for deployment and runs on cheap cloud GPU instances in minutes.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:45Sponsors
- 2:37Deeper Learning
- 3:24What Is Fast Ai
- 9:41Data Augmentation
- 10:42Google Image Database
- 13:47The Learning Rate
related talks
32mYou Can Do Deep Learning!
40mHow to Implement a Production Ready Deep Learning Model in 12 Weeks
38mReal-world Machine Learning with TensorFlow and Cloud AI
35mIs it Food? An Introduction to Machine Learning
02h 26mPractical Deep Learning for Data Scientists
15mMy first steps on Deep Learning with Keras
55mDeep Learning in the world of little ponies
53mDeep Learning in the world of little ponies