Repository navigation
v0.4.0-manuscript
·
19 commits
to main
since this release
Based on v0.3.0.
Backends
- cuSpatial does not scale to Atera. Use
cuspafor faster, scalable point-in-polygon joins. - cuSpatial's quadtree is unstable. Replace it with
fastquadtree, built on downsampled data, for tiling. - Replace
torch_scatter.scatter_maxwith nativetorch.scatter_reduce. - Require CUDA 13 and Python 3.13. Drop CUDA 12.
Scale
- Add CSR indexes (
index_to_ptrandquery_ptrinsegger.data.csr) that the int32 limit doesn't cap. - Map indices with
searchsorted. It is faster thantorch.isinon sorted data and has no int32 limit. - Chunk
points_in_polygonsto stay under cuspa's int32 pair limit. - Build prediction tiles from neighboring tiles instead of materializing the whole graph.
- Cut memory: store the count matrix and node coordinates once, use one
posinstead ofgeometry, drop the boundaryGeoDataFrameafter graph setup, and lower the host-memory peak when partitioning large graphs. - Take the quadtree root extent from all points, use a square root box with a minimum tile size, and clip leaves to the data extent.
Readers and CLI
- Add a MERSCOPE reader.
- Add
--segmentation-mode {nucleus,cell}. - Fix the spatialdata fields. Export treats a missing
convergedcolumn as True.
Fixes
gene_missing_strategy="remove"drops missing genes before normalizing.- The raw-count check on
gene_corr_referencewas inverted. It now accepts raw counts.