v0.6.0

0.6.0 (2026-06-16)

Features

• add icd preset for saliency-guided training (#303) (ba7ba4d)
• analyze progress output and CV cached-pred reuse (e7d6364)
• bound the XAI cache and stop unbounded hash-keyed growth (1be85ac)
• dashboard port auto-fallback, train --dry-run, config warning echo (309d1eb)
• warn once when branch search projects a very long runtime (#318) (0231591)

Bug Fixes

• apply augmentations in list order and pass async sample indices (76ea8d9)
• invalidate XAI cache when a different model shares the cache dir (#317) (202e258)

Performance Improvements

• lazy-import cv2 so import bnnr no longer pays the OpenCV cost (85f0689)

v0.5.7

0.5.7 (2026-06-12)

Bug Fixes

• select the best baseline epoch, not the last (#301) (95b9d8d)

v0.5.6

0.5.6 (2026-06-12)

Bug Fixes

• make analyze sequential for shuffled val loaders (#299) (f204f29)

v0.5.5

0.5.5 (2026-06-12)

Bug Fixes

• forbid unknown config keys and enforce epoch bounds (#297) (985dcc8)

v0.5.4

0.5.4 (2026-06-11)

Bug Fixes

• load checkpoints with weights_only=True first (#298) (233dc36)

v0.5.3

0.5.3 (2026-06-11)

Bug Fixes

• implement real local_mean fill in DetectionICD (#295) (8b72635), closes #255

v0.5.2

0.5.2 (2026-06-11)

Bug Fixes

• raise torch floor to 2.10.0 for security fixes (#293) (76b52bc)

v0.5.1

0.5.1 (2026-06-10)

Bug Fixes

• keep canonical augmentation names instead of indexed aliases (#289) (8e285f9), closes #258

v0.5.0

0.5.0 (2026-06-10)

Features

• implement real Opti-CAM and split out GradCAMExplainer (#287) (385aadd)

v0.4.16

0.4.16 (2026-06-10)

Bug Fixes

• make ICD/AICD CPU-bound to remove apply_tensor recursion (#285) (69435d6)