L21 | English F3 Student Learning Log

Persona: I'm a Form 3 (Grade 9) student and I only read English. I know basic Python (variables, for loops, lists) and school algebra up to quadratics. I learn from the official pages, the videos with captions, and the book. Kaggle notebook steps are written as a practice log that follows the official notebooks.

What I understood (in my own words)

It's hard to judge generated images by eye, so we use scores like FID.
Lower FID is better: generated images are closer to real ones.
DDIM can skip steps and generate with far fewer of them, so it's much faster.
With many experiments you can't remember everything, so W&B records them for you.

Where I got stuck, and how I got unstuck

Where I got stuck	How I solved it
The maths of FID	I remember the idea: compare the average and spread of features in two piles of images; the smaller the difference, the better.
Why DDIM can skip steps	The lesson says it uses a different derivation that doesn't need random noise at every step. I skipped the details for now.

My answers to the exercises

Q1. Two models have FID 5 and 40. Which one's images look more real?

The model with FID 5. FID measures the distance between generated and real feature distributions; lower means closer to real.

Q2. DDPM uses 1000 steps and DDIM 50 steps for similar quality. How many times faster is DDIM?

1000 ÷ 50 = 20 times faster.

Q3. Why do you need a tracking tool like W&B when running many experiments?

You need to record every setting you changed, the scores and the images, or you'll soon forget which setting was best. W&B logs and charts it all automatically, making comparison and sharing easy.

My Zettelkasten card

Card｜'Good' needs an objective score; 'fast' needs a step count

It's hard to judge generative models by eye, so researchers use scores like FID to turn 'how close to real' into a number. Speed is about how many steps it takes. DDIM matches DDPM's 1000-step quality in 50 steps, which is 20 times faster.

Our swim team used to pick relay members on the coach's gut feeling, and people felt it was unfair. Then the coach added timed trials, turn times and stroke counts. With numbers, everyone accepted the choices more and knew what to work on. Numbers don't capture everything, though; race nerves and tactics are hard to measure. FID is the same: a low score doesn't guarantee every image looks good.

My next step: log my 50 m times for five practice sessions in a spreadsheet and chart them to see if I'm improving.

🔗 Related notes to grow next: → Sport science: 'Turn feelings into measurable metrics' → Research: 'Experiment logs are your best memory' → Efficiency: 'Same quality in fewer steps wins'

Back to lesson note：L21_DDIM_202611041526