1. Question 1 — How to view the training history?

• ❌ Use model.fit

• ❌ Pass ‘history=true’

• ❌ Download model

• ✅ Create a variable ‘history’ and assign it to model.fit

✔ Example:

---

2. Question 2 — 150×150 image → 3×3 convolution → output size

Default Conv2D uses valid padding:

Output = 150 − 3 + 1 = 148×148

• ❌ 150×150

• ✅ 148×148

• ❌ 450×450

• ❌ 153×153

---

3. Question 3 — What does image_dataset_from_directory do?

• ❌ Load images

• ❌ Pick size

• ❌ Auto-labeling

• ✅ All of the above

---

4. Question 4 — Training accuracy ~1.0, validation accuracy ~0.75 by epoch 2

This indicates overfitting, so training longer didn’t help.

Correct choice:

• ❌ Overfit validation

• ✅ There was no point training after 2 epochs, as we overfit to the training data

• ❌ Bigger training set = better train accuracy

• ❌ Bigger validation = better train accuracy

---

5. Question 5 — 150×150 with 2×2 pooling

You answered correctly:

• ❌ 149×149

• ❌ 148×148

• ❌ 300×300

• ✅ 75×75

---

6. Question 6 — API to inspect convolution layer effects

• ❌ model.convolutions

• ❌ model.pools

• ❌ model.images

• ✅ model.layers API

Layers can be inspected like:

---

7. Question 7 — Why validation accuracy > training accuracy?

• ❌ Equally valuable

• ❌ No relation

• ✅ Validation accuracy uses unseen data → better indicator of real-world performance

• ❌ Validation smaller so not useful

---

8. Question 8 — Why overfitting is more likely with small datasets?

• ❌ Validation resembles training

• ❌ Not enough data to activate neurons

• ❌ Faster training -> missed features

• ✅ Because smaller datasets don’t contain enough feature variety

Explanation:
Less diversity → model memorizes noise → overfits.

---

🧾 Summary Table

---