Module 07 Summary

The idea this module was built around

Start frozen. Unfreeze only if the frozen version is not good enough, and lower the learning rate when you do.

What you can now do

• Choose between feature extraction and fine-tuning on data size and domain distance
• Freeze correctly, using both requires_grad and the optimiser's parameter list
• Identify the repeated unit and split on it
• Set discriminative learning rates and decay them together

The trap this module removed

A random split on small data. When four photographs show the same person, a random split puts that person on both sides, the model recognises the person rather than the property, and the inflated score survives review and collapses in production.

Figures worth remembering

Fine-tune the backbone one to two orders of magnitude below the head — typically 1e-5 against 1e-2.

Before you move on

Ask what the repeated unit is in a dataset you use. Patient, subject, session, document, site. Split on that.

Mark complete