STalign#
squidpy’s experimental.tl reimplements
STalign’s affine + LDDMM alignment in JAX. This section
is the evidence that the port is the same algorithm, in two parts that answer different questions.
squidpy API equivalence — the public API,
squidpy.experimental.tl, is covered by sixteen of STalign’s seventeen analyses rewritten against it and executed on a GPU, one page per upstream notebook. These show the results are reasonable on real data, through the public surface only.Intentional divergences — the one place the port deliberately departs from upstream’s algorithm, and the reading behind it.
The numerics are gated separately, by a test suite rather than by a page:
test_stalign_reference.py
in squidpy asserts every stage of the port against values upstream STalign itself computed, since a figure can
look right while the algorithm underneath has drifted. The reference bundle
page says what those values are and why they are stored rather than recomputed.
Upstream STalign is pinned throughout to
b2068ed,
and the squidpy fork to
aa2b414.
Every page names the commit behind its own numbers.