v0.4.6

Changelog

[0.4.6] — 2026-05-28

Added

• Sample analyze HTML report for preview without install.
• CIFAR-10 benchmark suite (benchmarks/): compares training without BNNR, torchvision RandAugment, and BNNR branch search with ICD/AICD — includes validation metrics and OptiCAM attention overlays on shared validation images.

Fixed

• LR scheduler for baseline-only runs (max_iterations=0): CosineAnnealingLR no longer uses T_max=0 in CIFAR/STL pipelines.
• Version sync: package, dashboard API, and PyPI metadata aligned to 0.4.6.

[0.4.5] — 2026-05-22

Security

• CodeQL wave 2: additional path hardening and CI workflow permissions (see PRs #42–#45).

[0.4.4] — 2026-05-22

Security

• CodeQL path hardening: consolidated fix for Uncontrolled data used in path expression alerts in dashboard backend, exporter, and events loader (path_security helpers with ensure_child, validate_run_id, resolve_events_jsonl).

[0.4.3] — 2026-05-22

Fixed

• bnnr demo post-run messages: Your report and XAI paths print immediately after training (before the dashboard wait loop), using the real artifact location reports/<run>/artifacts/xai/.

[0.4.2] — 2026-05-22

Fixed

• quick_run tests on Python 3.10+: mock patches now target the bnnr.quick_run submodule directly (the public bnnr.quick_run name is the function re-exported from bnnr.__init__).

[0.4.1] — 2026-05-22

Added

• CLI bnnr demo: zero-flag onboarding run on CIFAR-10 with a small demo CNN, preset demo (saliency-guided ICD + ChurchNoise), live dashboard, and post-run paths to the JSON report and xai/ heatmaps.
• Preset demo: model-aware augmentation set for the demo command (build_demo_augmentations / get_preset("demo", model=..., target_layers=...)).
• default_demo_config() for CLI demo defaults (m_epochs=1, max_iterations=1, XAI on).
• docs/quickstart_api.md: one-page quick_run reference.

Changed

• quick_run: uses default_train_config() when config is omitted; infers target_layers from the last Conv2d; dashboard=True starts the live dashboard before training (non-blocking after run()).
• Docs: README, getting started, CLI, API reference, and golden path promote bnnr demo / quick_run as the primary quickstarts.

[0.4.0] — 2026-05-22

Changed

• Training refactor (6.1 / 6.2): split monolithic training code into bnnr.training submodules and reduced top-level bnnr re-exports; public import paths and behavior remain backward compatible.

[0.3.1] — 2026-05-21

Added

• CLI bnnr quickstart: interactive wizard for a zero-config demo run (built-in defaults, sample limits 128/64, optional live dashboard).
• Zero-config bnnr train: --config is optional; built-in quickstart defaults when omitted (default_train_config()).
• Community infra: CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, issue/PR templates, label sync config.

Changed

• PyPI readme: dedicated README.pypi.md with absolute asset URLs (logo, dashboard screenshots, demo poster) so images render on pypi.org; synced with week-1 audit README.
• GitHub README: problem-first structure, dashboard screenshots, ICD/AICD panels, inline demo video with audio (1280p).
• Classifiers: Development Status 4 - Beta (was Alpha on PyPI 0.3.0 metadata).

[0.3.0] — 2026-05-15

Added

• Analyze: standalone model analysis feature (bnnr analyze) — metrics, XAI diagnostics, data quality checks, failure pattern detection, structured recommendations, and HTML report generation.
• Analysis module (src/bnnr/analysis/): schema, findings engine, recommendations engine, class diagnostics, calibration (ECE), cross-validation, and HTML report renderer.
• STL-10 pipeline: built-in STL-10 dataset support with VGG-style and quick CNN architectures.
• CLI: bnnr analyze command with --model, --data, --output, --task, --cv-folds, --xai-samples options.
• Evaluation module (src/bnnr/evaluation.py): prediction collection and evaluation utilities for analyze.

Changed

• Pipelines: supported datasets expanded to include stl10 alongside existing mnist, fashion_mnist, cifar10, imagefolder, coco_mini, yolo.
• Exports: analyze_model and AnalysisReport added to public API (bnnr.__init__).

[0.2.9] — 2026-04-12

Added

• CI: notebooks-smoke job and scripts/validate_user_notebooks.py — validates catalog notebooks (nbformat, kernelspec, syntax of code cells after stripping magics) without executing training.

Changed

• Reporting: atomic write for report.json; load_report validates JSON shape and required keys, with clearer ValueError messages and ValidationError wrapping for config.
• Dashboard: start_dashboard returns None when optional dashboard dependencies are missing instead of a misleading localhost URL; CLI and examples handle the degraded case.
• Pipelines: unknown augmentation preset logs a UserWarning in addition to the existing logger warning when falling back to auto.
• Training: custom metric failures are always logged with logger.warning(..., exc_info=True) instead of only when verbose is enabled.
• Dashboard API: invalid JSONL lines in the paginated events endpoint are logged with logger.warning.
• Docs / CLI: docs/cli.md aligned with supported datasets and presets; related doc updates for start_dashboard return value.
• CI / dev: comment on lint vs pytest matrix; pre-commit pins updated (ruff, mypy) closer to dev tooling; build-package waits on notebooks-smoke.

[0.2.8] — 2026-04-10

Changed

• Tests vs examples: Removed YOLO smoke_*.py scripts from examples/detection/ (they duplicated CI tests). Training compatibility is covered only by tests/test_ultralytics_detection_train_integration.py. The optional XAI overlay helper lives under tests/integration/smoke_yolo_xai_overlay.py.
• CI / YOLO26: On GITHUB_ACTIONS, the yolo26_cpu_adapter fixture fails the job if ultralytics is older than 8.3 or yolo26n.pt cannot be loaded, so YOLO26 + BNNR is always exercised on CI (locally, missing weights still skip). Tests test_yolo26_* are marked @pytest.mark.yolo26.

Added

• test_yolo26_multiple_train_steps_finite — eight consecutive train_step calls on yolo26n.pt.

[0.2.7] — 2026-04-10

Fixed

• Ultralytics YOLO26 / E2ELoss (UltralyticsDetectionAdapter): before loss(), the task model and last (detection) module are forced into train mode so the head returns a one2many dict with boxes (inference mode previously led to KeyError / wrong layout). Fused heads (cv2 / cv3 cleared after fuse()) are detected early with loss_yolo_fused_head; KeyError / TypeError from the criterion are caught as loss_yolo_pred_format_error with logs including the installed ultralytics version and head state.
• Detection XAI / probes: predict_detection_dicts restores the prior training flag after eval + predict, so interleaved inference cannot leave the YOLO task stuck in eval() for the next training step.

Added

• Tests: optional yolo26n.pt integration step (skipped when Ultralytics is older than 8.3 or weights are unavailable; tightened in 0.2.8 to fail on CI).

Changed

• Dev dependency: ultralytics>=8.3.0,<9 (YOLO26 / current loss paths).

[0.2.6] — 2026-04-10

Fixed

• Detection XAI (Ultralytics): generate_detection_saliency now scores every eligible Conv2d map by isotropy of the channel-mean saliency (downsampled), then re-runs a forward pass to capture the best layer. Taking only the chronologically last H,W>1 conv could still yield strong vertical/horizontal “barcode” structure after upsampling; the score min(column-wise std, row-wise std) prefers maps that actually vary in both axes.

[0.2.5] — 2026-04-10

Fixed

• CI: detection XAI regression test no longer triggers Ruff F841 (use column-wise std assert for 2D saliency vs vertical-stripe artifacts).

[0.2.4] — 2026-04-10

Fixed

• Detection XAI (Ultralytics): generate_detection_saliency no longer uses the final Conv2d when its feature map has height or width 1 (common in YOLO heads). Resizing such maps to the image size produced vertical/horizontal “barcode” heatmaps; the code now keeps the last 4D conv feature map with both spatial sizes > 1, with a safe fallback to target_layers hooks.

[0.2.3] — 2026-04-10

Fixed

• Ultralytics YOLO (UltralyticsDetectionAdapter): detection images stay in [0, 1] end-to-end (uint8 round-trip in BNNRTrainer always /255); predict passes float tensors in [0, 1] (Ultralytics does not divide torch tensors by 255). train_step uses fp32 loss, plain backward (no GradScaler on this path), grad clipping, cls clamped to [0, nc-1], degenerate bbox filtering, IndexError from assigner caught.
• Imports: lazy sklearn in calculate_metrics and lazy NMF in CRAFT so import bnnr works when Colab’s numpy/scipy/sklearn stack is temporarily broken until pip install -U aligns versions.

[0.2.2] — 2026-04-10

Fixed

• Detection training: _average_metrics no longer assumes every batch returns the same metric keys. Skipped batches (non-finite loss in DetectionAdapter) returned only loss + loss_non_finite, which caused KeyError: 'loss_loss_classifier' when averaging epoch metrics (e.g. Colab detection notebook).
• Ultralytics YOLO: training images are scaled to Ultralytics’ 0–255 float range only when the batch is already in ~0–1 space. Feeding 0–255 tensors no longer multiplies by 255 again (avoids overflow, NaN loss, and the “divide by 255” warning storm). Non-finite loss batches are skipped like torchvision detection.
• Dashboard (React): task: "detection" from the run state is honored — KPIs and charts use map_50 / map_50_95 instead of treating detection as classification (which led to empty charts, NaNs, or a blank UI when combined with bad loss data).

[0.2.1] — 2026-04-09

Fixed

• PyPI publish: remove redundant wheel.force-include for dashboard frontend dist that caused duplicate ZIP entries rejected by PyPI.
• Security: bump vite 5.4 to 6.4.2 (CVE path-traversal in optimized deps .map handling).
• Testing: add regression test for detection data quality analysis silent failure.

[0.2.0] — 2026-04-08

Added — Object Detection

• Object detection task (task: detection) — full training, evaluation, and XAI pipeline for detection models.
• Detection adapters: DetectionAdapter (torchvision Faster R-CNN / FCOS / RetinaNet / SSD) and UltralyticsDetectionAdapter (YOLO v8).
• Detection augmentations: bbox-aware transforms (DetectionHorizontalFlip, DetectionVerticalFlip, DetectionRandomRotate90, DetectionRandomScale), MosaicAugmentation, DetectionMixUp, Albumentations bbox integration, preset system (get_detection_preset).
• Detection ICD/AICD: iterative dataset curation for detection using bounding-box saliency priors with multiple fill strategies (gaussian blur, noise, global/local mean, solid).
• Detection metrics: mAP@0.5, mAP@[.50:.95], per-class AP (calculate_per_class_ap), detection confusion matrix; falls back to built-in implementation when torchmetrics is unavailable.
• Detection XAI: activation-based saliency maps (generate_detection_saliency), occlusion sensitivity (compute_detection_box_saliency_occlusion), three-panel visualization (GT overlay, saliency heatmap, prediction overlay).
• Pipelines: yolo and coco_mini built-in dataset pipelines for detection quickstart.
• Dashboard: detection triptych XAI panels in HTML standalone export.
• Events: detection metric units, split XAI artifact keys (xai_gt, xai_saliency, xai_pred), detection_details in sample prediction snapshots.
• Example configs, scripts, and notebook for detection (examples/detection/).

Fixed

• is_detection variable missing in _standalone_report_html (dashboard exporter crash).
• mypy errors in detection code paths (core.py predict_ultra narrowing, forward_layout Literal type, removed invalid is_detection kwarg in run_data_quality_analysis call).

[0.1.2] — 2026-04-04

• Dashboard (npm): lodash forced to ^4.18.0 via package.json overrides (Dependabot GHSA-f23m-r3pf-42rh, GHSA-r5fr-rjxr-66jc); regenerated package-lock.json.
• Align __version__, FastAPI dashboard version, README, and docs with release 0.1.2 (Python >=3.10 in user-facing docs).

[0.1.1] — 2026-04-04

• Security: constraint-dependencies in pyproject.toml (uv) — filelock>=3.20.3, pillow>=12.1.1, pygments>=2.20.0 to address Dependabot alerts; regenerated uv.lock.
• Breaking: minimum Python is now 3.10 (required for patched Pillow 12.x).

[0.1.0.post4] — 2026-04-01

• CI: run npm ci + Vite build before every pip install -e ".[dev,dashboard]" so src/bnnr/dashboard/frontend/dist exists; hatch force-include no longer breaks editable installs on lint/test jobs.

[0.1.0.post3] — 2026-04-01

• Dashboard: production Vite build is bundled in the PyPI wheel under bnnr/dashboard/frontend/dist, so pip install "bnnr[dashboard]" can serve the full React UI without cloning the repo.
• dashboard_web build output targets that path; CI runs npm ci + npm run build before python -m build so releases always include the frontend.

[0.1.0.post2] — 2026-04-01

Post-release: nowe artefakty na PyPI (bnnr-0.1.0.post2-*) — unika konfliktu z już wgranymi plikami 0.1.0. Zawartość funkcjonalna jak 0.1.0.

[0.1.0] — 2026-04-01

Initial Release (PyPI)

• BNNR core training loop — automated augmentation evaluation and selection
• Seven custom augmentations: ChurchNoise, Drust, LuxferGlass, ProCAM, Smugs, TeaStains, BasicAugmentation
• External augmentation integration: Albumentations, Kornia, torchvision
• XAI: OptiCAM, GradCAM saliency maps
• Advanced XAI: CRAFT, RealCRAFT, RecursiveCRAFT, NMF concept decomposition
• ICD / AICD — XAI-driven intelligent augmentation
• Live dashboard (React + FastAPI + WebSocket)
• CLI: train, report, dashboard, list-augmentations, list-presets, list-datasets
• Config system (YAML/JSON) with presets: auto, standard, aggressive, light, gpu
• Data quality helpers (near-duplicate hashing, basic image sanity checks)
• Tasks: single-label classification, multi-label classification, and object detection (this branch)
• Checkpointing and training resume
• Event system (JSONL) for training replay