YOLOZU 4.4.0

What's Changed

• Add NGC publish targets to container workflow by @thinksyncs in #94
• Release v4.4.0 by @thinksyncs in #95

Full Changelog: v4.3.1...v4.4.0

v4.3.1

[4.3.1] - 2026-04-18

Fixed

• CI smoke bootstrap now chooses a Python interpreter that can actually start the repo-local yolozu CLI, preventing deep smoke failures caused by stale or broken local virtualenvs.
• Container bootstrap locks are aligned with current build targets: requirements-runtime.lock now carries a Python 3.14-compatible numpy, and the RT-DETR pose image lock aligns cuda-python with the pinned torch CUDA bindings.

Changed

• Release/publish automation is stricter: .github/workflows/publish.yml now validates package version, optional manual-release tag input, and CHANGELOG.md release heading before publishing.
• Workflow-only edits are no longer a blind spot in CI. .github/workflows/build_and_test.yml now runs release/security workflow regression tests on .github/workflows/**-only changes.
• Security/reliability workflows now fail earlier in pull requests: container builds run for container-related PRs, and Scorecard also runs on PRs targeting main.

Contract change

• Reference adapter regression metadata now records provenance/SBOM snapshots in baseline_meta.provenance (pip freeze, python -VV, OS/CPU/torch build hashes).
• Added matrix baseline layout support for reference adapter regression paths (baselines/<adapter>/<backend>/<device>/<version>/<profile>.json).
• Added robust behavior metrics for regression gates (map50, map50_95, worst-k/median class AP, recall@K, IoU quantiles, mismatch counts).

Added

• Multi-task SDFT distillation: yolozu/sdft.py now supports task-specific losses for
  6D pose (rot6d — geodesic MSE proxy), keypoints (smooth-L1), depth (scale-invariant L1),
  and segmentation (BCE with teacher sigmoid targets). Per-key weights configurable via
  SdftConfig fields (rot6d_weight, keypoints_weight, depth_weight, seg_weight, etc.).
• SDFT convenience constructors: make_pose_sdft_config(), make_keypoints_sdft_config(),
  make_depth_sdft_config(), make_seg_sdft_config(), make_full_sdft_config().
• Multi-task Tent TTT: TentRunner supports auxiliary consistency losses for pose, keypoints,
  depth, and segmentation heads via aux_pose_weight, aux_keypoints_weight,
  aux_depth_weight, aux_seg_weight in TentConfig.
• TTT task presets: Added pose_safe, keypoints_safe, depth_safe, seg_safe, pose_mim
  presets with task-tuned hyperparameters. _choose_default_preset_id() auto-selects
  task-specific presets when sdft_task is set.
• TTTConfig multi-task fields: aux_pose_weight, aux_keypoints_weight, aux_depth_weight,
  aux_seg_weight, aux_temperature, sdft_task.
• Tests: 65 new tests in test_sdft_multitask.py and test_tta_multitask.py covering all
  task-specific losses, dispatch, convenience constructors, aux consistency, and presets.
• TTT improvement demo: yolozu demo ttt runs a deterministic domain shift + few-shot training,
  then reports a simple mAP proxy delta (no TTT vs with TTT) with overlay PNG evidence in the
  predictions interface contract.

Fixed

• calibration package shadowing: Standalone yolozu/calibration.py was shadowed by
  yolozu/calibration/ package — merged apply_temperature() and
  calibrate_predictions_entries() into the package __init__.py so imports work correctly.
• calibration/distillation key preservation: calibrate_predictions_entries() and
  distill_predictions() now preserve all original entry keys (e.g. image_size, preprocess)
  instead of dropping them.

Changed

• Image format support: All image-loading code paths (dataset, CLI, predict, rtdetr_pose, demos,
  pascal_voc, ade20k, make_subset) now accept BMP, TIFF, WebP, and GIF alongside JPEG/PNG.
• image_size: Added native header parsers for BMP, GIF, TIFF, WebP; PIL fallback for other formats.
• letterbox: compute_letterbox() accepts rectangular (w, h) tuples in addition to int.
• geometry: Zero-division protection in recover_translation(); fixed variable shadowing in
  corrected_intrinsics().
• boxes: width/height params accept int | float.
• gates: final_score() uses safe .get() with defaults; handles non-dict weights gracefully.
• constraints: apply_constraints() validates r_mat shape (falls back to identity); accepts
  non-dict cfg.
• cli: _detect_config_source_from_path() raises clear error on unknown .py config instead of
  silently defaulting to mmdet.
• sdft: Removed duplicate if total is None: block.

Added

• tests/test_dataset_formats_quality.py: 26 unit tests covering new image format parsers,
  dataset discovery, letterbox, geometry, calibration, distillation, constraints, and gates.

v4.3.0

What's Changed

• Strengthen training platform handoff and TAO backend by @thinksyncs in #90
• Release v4.3.0 by @thinksyncs in #91

Full Changelog: v4.2.0...v4.3.0

YOLOZU 4.2.0

What's Changed

• Finish benchmark parity lanes and detect breadth docs by @thinksyncs in #85
• Release v4.2.0 by @thinksyncs in #86

Full Changelog: v4.1.0...v4.2.0

YOLOZU 4.1.0

What's Changed

• Add release versioning modes and security policy docs by @thinksyncs in #30
• Reduce CodeQL backlog in benchmark and validation paths by @thinksyncs in #31
• Tighten runtime guard fallbacks by @thinksyncs in #47
• Neutralize public runtime branding by @thinksyncs in #52
• Reduce CodeQL backlog core tranche by @thinksyncs in #48
• Add local repository governance audit tool by @thinksyncs in #51
• build(deps): bump xu-cheng/latex-action from 3.3.0 to 4.0.0 by @dependabot[bot] in #39
• build(deps): bump actions/download-artifact from 4.3.0 to 8.0.1 by @dependabot[bot] in #38
• build(deps): bump dorny/paths-filter from 4.0.0 to 4.0.1 by @dependabot[bot] in #37
• build(deps): bump docker/setup-buildx-action from 3.12.0 to 4.0.0 by @dependabot[bot] in #36
• build(deps): bump github/codeql-action from 3.32.6 to 4.34.0 by @dependabot[bot] in #35
• build(deps): bump nvidia/tensorrt from 24.05-py3 to 26.02-py3 in /examples/infer_cpp/docker by @dependabot[bot] in #34
• build(deps): bump python from 3.11-slim-bookworm to 3.14-slim-bookworm in /deploy/docker by @dependabot[bot] in #33
• Reduce CodeQL backlog in dataset and eval fallbacks by @thinksyncs in #55
• Finish CodeQL backlog cleanup tranche by @thinksyncs in #56
• Reduce residual security and CI findings by @thinksyncs in #57
• Fix repo-backed MIM TTT compare workflow by @thinksyncs in #65
• Add macOS MPS beta support surfaces by @thinksyncs in #66
• build(deps): bump github/codeql-action from 4.34.0 to 4.35.0 by @dependabot[bot] in #61
• Fix code scanning follow-ups by @thinksyncs in #67
• Harden TTT figure renderer for CI by @thinksyncs in #68
• Update Scorecard action pin by @thinksyncs in #69
• Raise ONNX floor to 1.21.0 for code scanning by @thinksyncs in #72
• build(deps): bump docker/build-push-action from 6.19.2 to 7.0.0 by @dependabot[bot] in #63
• build(deps): bump actions/upload-artifact from 4.6.2 to 7.0.0 by @dependabot[bot] in #62
• build(deps): bump docker/login-action from 3.7.0 to 4.0.0 by @dependabot[bot] in #64
• Fix semver auto-bump classification by @thinksyncs in #71
• Document macOS Miniforge MPS setup by @thinksyncs in #70
• build(deps): bump nvidia/tensorrt from 26.02-py3 to 26.03-py3 in /examples/infer_cpp/docker by @dependabot[bot] in #59
• build(deps): bump nvidia/tensorrt from 24.05-py3 to 26.03-py3 in /deploy/runpod by @dependabot[bot] in #58
• Document residual scorecard alerts and tidy root ignores by @thinksyncs in #73
• Relocate lockfiles into dedicated directory by @thinksyncs in #74
• Simplify README onboarding by @thinksyncs in #75
• Trim README into docs-first index by @thinksyncs in #76
• Prepare YOLOZU-synthgen integration handoff by @thinksyncs in #77
• Add SynthGen manual details and measured SDFT docs by @thinksyncs in #81
• Align container workflow triggers and Docker docs by @thinksyncs in #82
• [codex] Finish manual scaffold cleanup by @thinksyncs in #83
• [codex] Release v4.1.0 by @thinksyncs in #84

New Contributors

• @dependabot[bot] made their first contribution in #39

Full Changelog: v4.0.0...v4.1.0

YOLOZU 4.0.0

What's Changed

• Fix publish workflow release validation by @thinksyncs in #23
• Audit benchmark parity gap against Ultralytics docs by @thinksyncs in #24
• Add torchscript benchmark format support by @thinksyncs in #25
• Add real benchmark parity artifacts by @thinksyncs in #26
• Define benchmark task semantics by @thinksyncs in #27
• Harden benchmark validation and runtime matrix by @thinksyncs in #28
• Add OpenSSF Best Practices badges to READMEs by @thinksyncs in #29

Full Changelog: v3.0.0...v4.0.0

YOLOZU 3.0.0

What's Changed

• Reduce empty-except findings and harden fuzzing by @thinksyncs in #19
• Add Phase-1 benchmark CLI by @thinksyncs in #20
• Sync benchmark docs, manual, and manifest by @thinksyncs in #21
• Release v3.0.0 by @thinksyncs in #22

Full Changelog: v2.0.0...v3.0.0

YOLOZU 2.0.0

Full Changelog: v1.0.8...v2.0.0

YOLOZU 1.0.8

Full Changelog: v1.0.7...v1.0.8

v1.0.3

Practical PyTorch demos: keypoints, depth (MiDaS), and MNIST train demo.\n\n- yolozu demo keypoints: Keypoint R-CNN inference + overlay + JSON\n- yolozu demo depth: MiDaS via torch.hub (downloads weights on first run)\n- yolozu demo train: bounded MNIST fine-tune + checkpoint + JSON\n\nAlso: yolozu[demo] now includes timm + opencv-python; tools/yolozu.py can passthrough demo.