The report
The report is the most important file of this lesson. The code may be excellent, but without a report nobody knows what it does or how far it can be trusted.
Your job
Create report.md (on the right) and fill it in. The template is bilingual: keep the language you write in and delete the other. Every <...> is a place for you to fill.
Two sections usually take the longest.
Section 2: the accuracy story
In this course the accuracy number did not move along a single road. Here is the story of our stand-in — the same one your report has to tell with your numbers:
| stage | accuracy | measured on |
|---|---|---|
| random guessing (1 / 4 classes) | 25 % | – |
| CNN trained from scratch, no augmentation (lesson 9) | 91.5 % | new images |
| ResNet-18, new head only (10) | 91.2 % | new images |
ResNet-18, layer4 fine-tuned (10) |
94.0 % | new images |
| capstone (no augmentation, early stopping) | 93.75 % | new images |
And the sentence we would write: "A small CNN trained from scratch on four clear classes already reaches ~92 %; frozen ImageNet features alone do not help (91.2 %); fine-tuning the last block gives +2.5 points over the scratch CNN; and the final configuration stays in the same place (93.75 % against 94.0 %) — the difference is one image out of 400."
Notice: it is not a steadily rising curve. There is one small jump, and the last two numbers stay in the same place. Your story may look different, especially if your photos are natural coloured ones — there it can climb more (in our side experiment with CIFAR-10, on validation: 52.5 % → 65.0 % → 79.4 %). Write down what was.
Section 7: what did not work
This section is not a punishment. It shows that you know where your knowledge ends. An example from ours:
Augmentation (mirror flip, ±10° rotation, 10 % shift) together with early stopping lowered the test from 93.75 % to 89.0 %; coat recall fell 0.88 → 0.72. We did not isolate the cause: we suspect that geometric transforms on 28×28 thumbnails blown up to 224 wash out the difference between a coat and a pullover, but we did not check it.
Notice the word "suspect". If you know the cause, you checked it; if not, write exactly that.
Before you hand in
Read the report against the rubric of step 6. For every row you must answer "yes" or write why not.