Transfer Learning & Fine-Tuning: Adapt Pre-trained Models

About this course

“This course contains the use of artificial intelligence.” Why adapt a pre-trained model? Training a deep network from scratch takes a lot of labelled data, compute and engineering time. Transfer learning lets you start from a pre-trained model and adapt it to your task with far less of each. The hard part is knowing how to adapt it: freeze it, tune part of it, or tune all of it, and how to tell whether it worked. In this course you will Choose between feature extraction, partial fine-tuning and full fine-tuning, based on how much data you have and how similar your task is.

Isolate a pre-trained backbone, replace its head and unfreeze layers in stages with conservative learning rates.

Use parameter-efficient methods to cut memory and storage costs and to limit catastrophic forgetting.

Spot data leakage, label imbalance and negative transfer before you release a model.

Plan artifact versioning, runtime serving and long-term data drift monitoring. Practice and resources Two follow-along labs measure the ideas on a small network you can train in seconds: when transfer learning helps and when it hurts (Section 3), and four mistakes to catch by measuring them (Section 5). Notebooks are downloadable so you can run them on your own computer.

A one-page cheat sheet for every section, plus a glossary of key terms. Who this is for Machine learning engineers, data scientists, technical architects and engineering managers who adapt existing models instead of training from scratch. Basic familiarity with neural networks and some Python will help you get the most from the labs.

Read the full description on Udemy on Udemy (opens in a new tab)

Prices, discounts and availability are set by Udemy. We may earn a commission when you purchase through links on this site.