Eight questions covering epochs and loss curves, choosing a model, your own drawings, saving, baselines, exporting predictions, and the method writeup.
8 questions. You need 75% to pass, and you can retake it as many times as you want.
1
What is one training epoch?
2
Training loss falls while development loss rises. What should you investigate?
3
Development performance was best at epoch 7. Which model should face the final test?
4
Your recognizer struggles with your drawings. What should you check first?
5
How can you check that saving and reloading preserved a model?
6
Your accuracy is 96%; an always-A baseline scores 97%. What is missing?
7
Your output has the right row count but renumbered IDs. Is it ready?