Module 06 Activity

Scenario

Train something - and establish first what it has to beat.

What you build

A fine-tuned model with matching preprocessing, a two-phase learning rate, and results reported against three baselines.

Steps

1. Visualise the first-layer kernels of a pretrained model and describe what three of them respond to.
2. State how far your domain is from the source domain, and how many layers you expect to retrain because of it.
3. Train the head with the backbone frozen, then unfreeze at a much lower rate. Report both rates and both scores.
4. Take the preprocessing from the checkpoint's own transform. Assert one known input produces a known output.
5. Compute the majority-class baseline and a trivial pixel-statistic baseline before training.
6. Report the model against all three references, on the metric that matters.

Evidence to hand in

• Three first-layer kernels described.
• The domain-distance assessment and its implication.
• Both training phases with rates and scores.
• The preprocessing assertion.
• Both baselines, computed before training.
• The three-way comparison on the metric that matters.

Review checklist

• Preprocessing comes from the checkpoint, not from memory.
• The fine-tuning learning rate is much lower than the head rate.
• Both baselines were computed before training, not after.
• The comparison uses the metric that matters, not just accuracy.
• If the model does not beat the baselines, that is reported.

Mark complete