Target: run CIFAR10-style transfer, report accuracy + confusion matrix, and ablate frozen vs unfrozen training.
1. Frozen backbone first
Freeze the pretrained backbone, train only the head. This is the honest baseline: fast, stable, and it isolates what transfer actually contributes.
2. Unfreeze ablation
Then unfreeze the last block with a 10x-smaller LR and compare. Report both accuracies plus confusion matrices — the ablation must show where the extra capacity helped (which classes moved) not just a headline delta.
3. Trap: test-leak via augmentation
Augmentation pipelines fit on (or sampling from) test statistics leak. Fit all normalization/augmentation statistics on train only; the test transform is deterministic.
Recap
Frozen-head baseline, unfreeze ablation with per-class deltas, and train-only augmentation statistics.