v0.16.0
·
273 commits
to main
since this release
SPINE v0.16.0 introduces the new opt-in HDF5 V2 format, columnar analysis, backend-neutral sparse tensors, and robust duplicate-coordinate handling.
Added
- Versioned HDF5 V2 layout: Adds an explicit opt-in format with flat variable-length payload datasets, direct ranges, selective reads, and automatic V1/V2 reader detection. V1 remains the default writer format.
- Columnar HDF5 analysis: Adds reader-controlled columnar execution, analyzer capability checks and requests, matching-friendly scalar IDs, and direct CSV export.
- HDF5 maintenance tools: Adds comparison, benchmarking, and fast V2 litification utilities under
bin/output. - Backend-neutral sparse tensors: Adds sparse tensor and batch abstractions shared across NumPy and PyTorch paths.
Changed
- Sparse duplicate handling: Preserves per-product alignment through explicit reduction maps and consistent jitter augmentation.
- Selective HDF5 performance: Avoids unnecessary long-form payload reads and reduces trusted V2 class-reconstruction overhead.
- Analysis execution policy: Makes event-wise versus columnar operation an IO-level policy with early compatibility validation.
- CLI utilities: Expands type hints, module documentation, and README coverage across
binscripts.
Fixed
- Sparse decoder cardinality: Corrects output-size handling when duplicate coordinates are removed.
- Cross-product jitter alignment: Applies one coordinate transform and one keep map consistently to all augmented sparse products.
Full changelog: v0.15.5...v0.16.0