Part 7 · Keep Learning

Build it yourself (transfer)

The repo now contains a tutorial at tutorials/masked-diffusion/: a numpy skeleton of everything Maths II derived — the forward masking process, the reveal-probability sampler step, the NELBO weight, the weighted MLM loss, and a DDIM step — with the implementations replaced by TODOs and a test suite that knows the right answers (including statistical checks that your sampler preserves the marginals):

cd tutorials/masked-diffusion
python3 -m pytest -q        # red until your implementations are right

When the suite is green, you have not read masked diffusion — you have written it. Stuck on one function? solutions/ has a reference implementation; the honest move is to peek at one function, not the file.