Syllabus Lesson 192 of 246 · Module 17: Noise in, pictures out (diffusion)
Module 17: Noise in, pictures out (diffusion)

Gate: a new shape (AI off)

Why it matters Every recipe in this module was tuned on the spiral. A recipe you can only run on the data it was tuned for is a recipe you do not yet understand.

Gate rules: AI tutor off. The new data: two rings, an inner one with 40% of the points and an outer one with 60%. With the spiral's settings (difflab.train('rings')), the samples land on the rings only about half the time.

Change the training settings so the model draws the rings. You may use lr, steps (up to 5000), hidden (up to 128), batch, kind and time. Each training run takes a few seconds.

Train with seed 7 as the starter does: the check replays the start of your run to confirm the settings, then judges your model on fresh samples and passes it with at least 0.75 on shape and 0.95 coverage.

Words and symbols in this lesson
eps
the bell-curve noise that was added; the network learns to predict it
on shape, coverage
the 2D judge: share of samples near the real shape, and share of the shape with a sample near it
Your turn

Set settings to a dict for difflab.train, using any of lr, steps (at most 5000), hidden (at most 128), batch (at most 512), kind and time. Keep the starter's training line (seed 7): the check judges that model, then samples 1500 points with a new seed. To pass: coverage at least 0.95 and on shape at least 0.75.

AI off: no tutor, hints or worked solution on this one. Your own thinking only.

Spotted a problem in this lesson? Report it · Ask about this lesson

This exercise opens soon

Finish the lesson before it and this one unlocks. The whole course is free, no account needed, and you can read this lesson right now.