What to hand in

A short table, a short paragraph and a started folder of photographs.

1. The equal-budget comparison — with your own numbers

Run both commands from step 5 and fill in the table with your numbers:

model	parameters	train	val	seconds
MLP, --hidden 25
CNN, (16, 32)

In our table the difference was +3.1 points. What are the direction and size of your difference? If it is under 1–2 points, write what that means.

2. A wider network

In our run CNN(channels=(32, 64), dropout=0.3) had 50,186 parameters and reached 89.3 % on validation (15 epochs) — but it took 117 seconds while the CPU was busy.

Do not change train.py: write a separate small script (e.g. wide.py), reusing what shuffled.py does. Run it and answer:

• How many extra points did roughly 2.4 times the parameters buy?
• How many times slower did it train?
• Which is better value for money — a wider CNN or a narrower one trained for more epochs? (Pick one hypothesis and test it.)

3. One paragraph

4–6 sentences: when is it worth choosing a CNN, and when an MLP? Use your own numbers from parts 1 and 2, and mention the shuffled-pixels experiment. There must be at least one sentence saying what these numbers do not prove.

4. Start collecting your own photographs

In lesson 9 you work with your own categories from lesson 1's homework. Now:

1. Create a photos/ folder with one subfolder per category (photos/category_a/, photos/category_b/ …).
2. Start adding photographs — at least 2 categories, at least ~60 pictures in each (lesson 9 shows why that many, and why 100 is better).
3. Set aside separately about 20 more pictures per category that you never touch while training. That will be your final test.

For now, only collect. Do nothing else with them.