0.7.0: 20x-40x faster batch training, and a documentation pass - #22

Merged

andremsouza merged 9 commits into

perf/vectorised-batch-training

Jul 31, 2026

Commits on Jul 31, 2026

perf: contract the batch neighborhood by axis instead of per node
andremsouza
committed
perf: find every best-matching unit with one matrix product
andremsouza
committed
perf: add an optional numba kernel behind a fast extra
andremsouza
committed
bench: re-measure, and retire the benchmark whose subject no longer exists
andremsouza
committed
docs: give the derivations a home before the docstrings lose them
andremsouza
committed
docs: cut the narration from the docstrings and the manifest
andremsouza
committed
chore: release 0.7.0
andremsouza
committed
fix: install numba directly rather than through an extra
andremsouza
committed
docs: correct the pages that 0.7.0 made wrong, and finish the prose pass
andremsouza
committed