Kwcoco-native detector training, evaluation, mining, and packaging. KDK keeps rich kwcoco / large-image semantics in its data plane and uses pluggable model engines; LibreYOLO is the primary general backend for current detector and instance-segmentation families.
git clone https://github.com/Erotemic/kwcoco-detector-kit.git
cd kwcoco-detector-kit
git submodule update --init --recursive # includes tpl/libreyolo
pip install -e ".[dev,libreyolo]"
bash examples/kwcoco_demo/run_smoke.shrun_smoke.sh exercises the full pipeline (synth kwcoco → tile → train mock → ONNX export → eval → eligibility manifest) in <90 s on a 1-CPU laptop.
data/— kwcoco tile augmentation (three modes: full-only, quadrant grid, multi-scale fixed-size), positive + hard-negative merging, offline hard-negative mining, kwcoco → MSCOCO export.trainers/— pluggable trainer interface;libreyolois the general model-engine backend (DEIMv2, D-FINE, RF-DETR, YOLOv9, GTR, TinyFormer in the initial catalog), while direct adapters remain available as specialized/reference paths;mock_tinyis a CPU smoke detector.predictors/— trained-checkpoint inference adapters used by the eval + hard-neg mining paths.export/— ONNX export + modelspec sidecar, torch ↔ ONNX parity guard, deployment package YAML.eval/— kwcoco eval driver, checkpoint shortlist sweep, ONNX desktop benchmark.orchestration/— Pareto sweep state machine, round-based hard-negative mining driver, eligibility manifest, setup-time--check-envprobe.config-init/config-inspect/config-edit— editable environment + dataset YAML configs with host and kwcoco introspection; seedocs/configs.md.
All CLIs use kwconf; python -m kwcoco_detector_kit --help or kwcoco-detector-kit --help.
The KDK/LibreYOLO ownership boundary and backend catalog are documented in
docs/libreyolo_integration.md.
| Tier | Hardware | Recommended variants |
|---|---|---|
| S | 1× 12–16 GB (GTX 1080 Ti / Titan X) | DEIMv2 HGNetv2 Atto / Femto / Pico |
| M | 1× 24 GB (RTX 3090 / 4090) | DEIMv2 HGNetv2 N / S, DEIMv2 DINOv3-S |
| L | 1× 48 GB (L40S / RTX 6000 Ada) | DEIMv2 DINOv3-M, OGDino-Swin-Tiny |
| XL | 1× 80 GB (A100 / H100) | DEIMv2 DINOv3-L/X, OGDino-Swin-Base |
| 2-4×L / 4×XL | DDP cluster | full sweep, larger effective batches |
| cloud | SLURM / k8s | tier-L/XL × N nodes with cloud-mount kwcoco |
See docs/scale_tiers.md.
This kit ships agent-readable engineering memory in dev/ — 19 documented failure modes from the prior prototype + 4 distilled benchmark candidates. New >1 h-debug bugs land in dev/journals/lessons_learned.md above the seed divider.
Phase 1 in flight (port + RGB + tier M/L single-GPU). See CHANGELOG.md and PLAN.md for the full roadmap.
Phase-3 (webdataset / multispectral / cloud) ship-vs-deferred status is tracked in docs/phase3_status.md.
examples/— minimal, self-contained demonstrations of one capability (e.g.kwcoco_demo/the 90 s CPU smoke). Not maintained as live campaigns. Thesealion_aerial/andviame_sealions_2026/example dirs are superseded historical references (banners at their top).projects/— real, ongoing experimental campaigns with their own scripts, class schemes, and run registry.projects/viame_sealions_2026/is the live sea-lion project and the reference shape new projects copy.
Apache-2.0.
