6.0 Study Guide and Task Sheet: Deep Learning and Transformer Basics

The main study route is now in Chapter 6 entry. Use this page only as a quick checklist while you practice.

One-Line Mental Model

batch data→model forward→loss→backward gradients→optimizer step→curves

If the code feels long, find these six steps first.

Expected Final Output

By the end of Chapter 6, your visible output should be a small evidence folder, not only finished reading notes:

deep_learning_evidence/
  shape_trace.txt
  training_log.csv
  loss_curve.png
  best_checkpoint_note.md
  attention_note.md
  failure_sample_note.md

If this folder is missing, the chapter is not finished yet, even if every page has been read.

Practice Checklist

Check	Evidence
I can explain forward, loss, backward, optimizer	training-loop note
I can run a minimal PyTorch script	train.py
I can print tensor shapes through a model	shape trace
I can compare training and validation curves	curve image or CSV
I can explain what Attention changes	attention note
I can finish the evidence-pack workshop	deep_learning_workshop_run/

Check reasoning and explanation

Use this checklist as a self-review rubric:

1. A training-loop note should explain forward pass, loss, backward pass, and optimizer step without copying code line by line.
2. A valid PyTorch script should run from a clean folder and print at least one shape, loss, or metric that proves it executed.
3. A useful shape trace should include batch size, channel/feature dimensions, and the point where tensors enter the classifier or loss.
4. A curve artifact should support a diagnosis: improving, underfitting, overfitting, unstable, or unclear.
5. An attention note should explain what Q/K/V and masking change compared with earlier sequence models.
6. A finished evidence pack should be rerunnable and should contain enough artifacts for someone else to understand the result.

Evidence Rubric

Artifact	It should answer
Training-loop note	What happens in forward, loss, backward, and optimizer step?
Shape trace	How do tensor shapes change through the model?
Curve image or CSV	Is the model underfitting, overfitting, or improving steadily?
Attention note	What information does attention add, and what remains hard?
Failure sample note	Which sample fails, and what does that tell you about data, model, or labels?

Evidence to Keep

Before leaving Chapter 6, keep one compact evidence pack:

Shape Trace

one model with printed tensor shapes

Training Log

train and validation loss over time

Best Checkpoint

how the best model was selected

Attention Note

Q/K/V, mask, and next-token bridge

Failure Sample

one wrong or weak prediction with next action

Project Folder

runnable evidence pack or README

Ready To Continue

Continue to Chapter 7 when you can train one small model, save the training log, inspect failure cases, and explain why the model improved or failed.