tadasets

tadasets generates the shapes the field tests against: a torus, a d-sphere at arbitrary dimension, a swiss roll, an infinity sign, eyeglasses. Each generator takes a noise level and can embed its output in a higher ambient dimension. It is a few hundred lines over numpy, and it is the smallest package profiled in this corpus by a wide margin.

Its importance is out of proportion to its size, because these are the point clouds whose answer you already know. A torus has one component, two independent loops, and one void. A d-sphere has one component and one d-dimensional feature and nothing between. Run a persistence implementation on them and the barcode either shows that or it does not, and if it does not you have found a bug rather than a discovery. Noise and ambient dimension are the knobs that turn a sanity check into an actual test — a barcode that is right at zero noise in three dimensions and wrong at moderate noise in twenty tells you something specific about the implementation.

For a foundry whose work is largely verification, that is a load-bearing dependency and not a convenience. It is also why the two clean-room featurizers written here — where no bit-exact oracle against the original implementation was available — lean on ground-truth geometry for the parts of correctness that can be checked absolutely.

In this corpus

tadasets is the only package here with no fixture of its own; it is a dependency, and the notes about it are about closing that dependency. It has no conda package at all, which blocked two closures at once — the scikit-tda bundle needed it transitively, and petls-pytorch declared it as a runtime dependency until our upstream contribution demoted it to a benchmark extra. tadasets-recipe supplies it, verified green, and is the one recipe in this corpus with a pull request open against a public channel; when that lands, this package stops being a gap.