Skip to content

v0.12.0

Choose a tag to compare

View all tags
@francois-drielsma francois-drielsma released this 11 May 06:02
· 428 commits to main since this release
v0.12.0
96a6f8c

Added

  • Staged HDF5 caching: Add staged cache readers and writers that support one cache file per source file, per-stage completeness tracking, provenance validation, and staged cache reuse across sequential workflows (#128).
  • Mixed dataset loading: Add MixedDataset plus staged/flat HDF5Dataset support so live LArCV inputs can be aligned with cached HDF5 products in training and inference jobs (#128).
  • Generic HDF5 parsers: Add cached-tensor, cached-index, and cached-object parsers for HDF5-backed SPINE products, including feature ablation and cluster-tensor reconstruction support (#128).
  • Data augmentation: Add rotation and pixel-jitter augmentation plus broader augmentation test coverage and geometry-aware worker initialization (#127).
  • Validation tooling: Update bin/output_check_valid.py to prefer staged-cache completeness and provenance metadata when available while preserving legacy fallback checks.

Changed

  • I/O package structure: Reorganize spine.io around top-level readers, writers, parsers, datasets, augmentation, collation, overlay, and sampling utilities, replacing the older core/torch split (#128).
  • Writers: Extend HDF5, staged HDF5, and CSV writers with cleaner prefix/suffix/directory handling and driver-facing staged-writer integration (#128).
  • Documentation: Refresh the spine.io API docs to reflect the new staged-cache and dataset structure, and harden docs builds against missing optional ML dependencies.
  • Testing: Expand spine.io, augmentation, bin-script, and staged-cache regression coverage substantially; restore full spine.io coverage after the refactor (#127, #128).

Fixed

  • GrapPA caching path: Preserve the standard GrapPA path while supporting cached cluster/edge/feature inputs, and fix related geometric-feature and Numba indexing/reshape issues (#128).
  • Cluster-label adaptation: Switch full-chain cluster label adaptation to use orig_index provenance instead of dense ghost masks, enabling cached deghosted workflows with evolving segmentation predictions (#128).
  • Optional imports: Make optional-dependency proxies and docs builds robust when PyTorch and other heavy dependencies are absent or mocked.
  • Stage writer stability: Isolate per-stage schema state correctly so staged cache writes do not leak product definitions across stages (#128).

Notes

  • Caching workflow maturity: This release includes the core staged-caching infrastructure needed for sequential training and inference workflows. The end-to-end full-chain caching-enabled training workflow has not yet been exhaustively validated across every stage and may still require additional integration debugging.

Full Changelog: v0.11.1...v0.12.0

Assets 2
Loading