---

name: d-fine description: "Route D-FINE object-detection setup, training, architecture inspection, inference, export, and troubleshooting workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

D-FINE Repo Skill

Use this skill when the user is working with the D-FINE object-detection repository and needs command guidance, config selection, architecture inspection, inference/export recipes, or troubleshooting.

Quick install and smoke check

From a D-FINE checkout, install the runtime dependencies first:

pip install -r requirements.txt

Add optional backend packages only when the workflow needs them:

• pip install -r tools/inference/requirements.txt for ONNX Runtime, TensorRT, or OpenVINO inference helpers.
• pip install -r tools/benchmark/requirements.txt for FLOPs or TensorRT benchmark helpers.
• pip install onnx onnxsim before ONNX export if they are not already available.
• pip install matplotlib if solver/validator imports fail during inspection.

Then run the bundled smoke probe from the checkout root:

python scripts/dfine_environment_probe.py --repo-root . --config configs/dfine/dfine_hgnetv2_n_coco.yml --build-model

Read references/install-and-environment.md for the public install paths and optional backend prerequisites, and read references/repo-provenance.md before deciding whether this skill matches the checkout.

Route map

• Dataset/config selection, class counts, remap policy, and YAML preflight: use sub-skills/data-and-configs/SKILL.md.
• Training, --test-only, resume, tuning, AMP/EMA, DDP launch, and output dirs: use sub-skills/training-evaluation/SKILL.md.
• Registry, model graph, deploy mode, and API inspection: use sub-skills/architecture-api/SKILL.md.
• PyTorch / ONNX / OpenVINO / TensorRT inference, export, benchmark, and EMA extraction: use sub-skills/inference-export/SKILL.md.

What this skill covers

• Choosing the right D-FINE config family for COCO, Objects365, Objects365-to-COCO, custom COCO-format, CrowdHuman, and VOC workflows.
• Building and checking train.py commands without mixing resume and tuning.
• Inspecting YAMLConfig, registry-based object construction, and the D-FINE model graph.
• Generating safe command recipes for image/video inference, ONNX export, TensorRT/OpenVINO deployment, and FLOPs/latency checks.
• Diagnosing import, dependency, checkpoint, config, and backend failures using bundled troubleshooting references.

How to use the route map

1. Start with the sub-skill that matches the user-visible task.
2. Read the linked reference there for signatures, workflows, and failure modes.
3. Run the bundled helper script if the task needs a safe command generator or a config/model probe.
4. If a workflow crosses sub-skill boundaries, hand off instead of mixing unrelated logic in one answer.

Cross-cutting notes

• D-FINE uses src-based registry imports, so missing registration usually means the defining module was not imported.
• HGNetv2.pretrained should usually be disabled for inspection or export smoke unless the user explicitly wants pretrained lookup.
• TensorRT, OpenVINO, and benchmark workflows are optional backend paths; do not claim them as verified unless the matching runtime packages are installed.
• Keep generated runtime guidance self-contained; do not depend on opening the original repository docs at answer time.