Module 07 Activity

Scenario

You have a pretrained backbone and a few hundred labelled examples. Decide between freezing and fine-tuning on evidence.

Task

1. Count trainable parameters under feature extraction and under fine-tuning.
2. Freeze all but the last layer, train, then verify by comparing weights before and after.
3. Record the size and memory cost of three backbone widths, and write the provenance questions you

would need answered before adopting one.

1. Build a dataset with a repeated unit. Score a random split and a grouped split, and explain the gap.
2. Set up discriminative learning rates with a scheduler, and confirm the ratio holds as both decay.
3. Run both approaches on the same split and report both against the baseline.

Deliverable

An experiment memo with both scores, both trainable-parameter counts, the learning rates used, and a verdict.

Check your work

• Only the unfrozen layer's parameters change. Verify by comparison, not by reading the code.
• The random split scores higher than the grouped split — and the lower number is the honest one.
• scheduler.step() belongs once per epoch, after the optimiser.

The verdict to write

If freezing matched fine-tuning, ship the frozen version. On small data it frequently does.

Mark complete