Module 12 Summary
What you have built
A framed problem, a leak-free split, two baselines and a network, per-epoch evidence, a deliberate threshold, and a card that says where the model must not be used.
What the capstone proved you can do
- Fix the decision and the error costs before touching the data
- Identify the repeated unit and split on it
- Report a network against both a trivial and a classical baseline
- Choose a threshold from the cost asymmetry rather than accepting 0.5
- Write limitations and next steps that a reviewer could act on
The three limitations every honest version states
The split is only as good as the repeated unit you identified. A single seed means results move with initialisation. And synthetic or small data limits every claim that follows.
What actually transfers
The PyTorch API is in the documentation. What is not is the habit underneath it: state the shapes, verify a gradient by hand once, seed the run, compare against baselines, log every epoch, split on the repeated unit, check calibration before trusting a threshold, and write down where the model must not be used.
That habit is what makes a model something a colleague can rely on. Take it to the next problem, and start by asking what decision it is meant to support.