L15 | 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)

Padding adds a border of zeros so the image doesn't shrink after convolution.
Stride 2 means the kernel jumps two pixels at a time, halving width and height.
An autoencoder is like zipping an image into a tiny file and unzipping it to see if it comes back.
When experiments are slow, building a good tool first saves time later.

Where I got stuck, and how I got unstuck

Where I got stuck	How I solved it
The output size formula	I checked it on a small case: 5×5 input, 3×3 kernel, no padding, stride 1 gives 3×3, and (5 − 3) / 1 + 1 = 3.
Decorator syntax looked strange	I think of it as wrapping one function inside another to add extra behaviour.

My answers to the exercises

Q1. A 28×28 image, 3×3 kernel, padding 1, stride 2: what is the output size?

(28 + 2×1 − 3) / 2 + 1 = 27 / 2 + 1. Rounding 27 / 2 down gives 13, so 13 + 1 = 14. Output is 14×14.

Q2. If an autoencoder's goal is 'output equals input', what's the point of it?

The small middle representation is forced to keep only the most important information, so autoencoders can compress data and remove noise. Stable Diffusion's VAE uses this idea to turn images into small latents.

Q3. Why did Jeremy stop to build a Learner when the autoencoder experiments weren't going well?

Every change meant rewriting lots of training code, which was slow and error-prone. With a Learner, changing the model or settings takes one or two lines, so experiments go much faster.

My Zettelkasten card

Card｜When things aren't working, stop and sharpen your tools

Jeremy's autoencoder kept failing. Instead of pushing on, he stopped and built a Learner to wrap up the repeated training code. With a better tool, each experiment needed only a line or two of changes, and he got there faster.

For weeks my ollies on my skateboard weren't getting higher no matter how much I practised. A friend checked my board and found the trucks were far too loose, so my pop was wasting energy. After tightening them, the same technique went noticeably higher. I'd assumed it was my technique; it was my equipment. I also remind myself not to polish tools forever instead of doing the work. Jeremy built the Learner after a few experiments showed him the bottleneck, not on day one.

My next step: turn three bits of Python I keep retyping into functions and save them as my own helper file.

🔗 Related notes to grow next: → Work habits: 'If you've done it three times, make it a tool' → Sport: 'Equipment problems often look like technique problems' → Tech: 'Compression means keeping only what matters'

Back to lesson note：L15_自編碼器_202611041526