Skip to content

v0.4.0-manuscript

Choose a tag to compare

@Tobiaspk Tobiaspk released this 30 Oct 16:10
· 19 commits to main since this release

Based on v0.3.0.

Backends

  • cuSpatial does not scale to Atera. Use cuspa for 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_max with native torch.scatter_reduce.
  • Require CUDA 13 and Python 3.13. Drop CUDA 12.

Scale

  • Add CSR indexes (index_to_ptr and query_ptr in segger.data.csr) that the int32 limit doesn't cap.
  • Map indices with searchsorted. It is faster than torch.isin on sorted data and has no int32 limit.
  • Chunk points_in_polygons to 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 pos instead of geometry, drop the boundary GeoDataFrame after 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 converged column as True.

Fixes

  • gene_missing_strategy="remove" drops missing genes before normalizing.
  • The raw-count check on gene_corr_reference was inverted. It now accepts raw counts.