Module 07 Knowledge Check
5 questions. Pass mark 4 out of 5. Answer every question before checking the answer key below, then retry after reading the feedback.
1. You have a few hundred labelled images. Which approach is the safer starting point?
- A. Fine-tune the whole backbone
- B. Freeze the backbone and train a new head
- C. Train from scratch
- D. Use a larger backbone
2. You set requires_grad = False on the backbone but still pass model.parameters() to the optimiser. What is the risk?
- A. Nothing, it is equivalent
- B. Frozen tensors can still be updated by momentum or weight decay
- C. Training will error
- D. Gradients will be doubled
3. Forty photographs of ten people. A random split scores higher than a split by person. Which number is honest?
- A. The random split
- B. The grouped split
- C. Their average
- D. Neither
4. When fine-tuning, how should the backbone's learning rate compare to the head's?
- A. The same
- B. One to two orders of magnitude lower
- C. Higher
- D. It does not matter
5. Where should scheduler.step() be called?
- A. Once per batch
- B. Once per epoch, after the optimiser
- C. Before the optimiser
- D. Once per run
---
Answer Key and Explanations
Check these only after attempting every question.
1. B — Freeze the backbone and train a new head
Feature extraction needs little data and cannot damage the pretrained weights. Unfreeze only if it is not good enough.
2. B — Frozen tensors can still be updated by momentum or weight decay
Freezing has two halves. Verify by comparing weights before and after, not by reading the code.
3. B — The grouped split
A random split puts the same person on both sides, so the model recognises the person rather than the property. The lower number is the real one.
4. B — One to two orders of magnitude lower
The head is random and needs large updates; the backbone holds pretrained structure and needs small ones. A single rate serves one of them badly.
5. B — Once per epoch, after the optimiser
Calling it per batch decays the rate as many times as there are batches, collapsing it within the first epoch.
Practical Check
Apply this module to your own work: complete the module activity for *Transfer Learning and Fine-Tuning for Vision*, then write one sentence naming what your result shows and one naming what it does not.
Strong Answer Pattern
A strong answer names the task, the evidence used, the check performed, and the remaining limitation. It avoids "proved", "guaranteed", or "always" unless the evidence genuinely supports it.