---

name: yolov7-d2 description: "Operate YOLOv7-d2, a Detectron2-based repository for YOLO-family, SparseInst, DETR-family training, inference, evaluation, and ONNX deployment workflows." metadata: disco-role: operating disable-model-invocation: true license: GPL 3.0

YOLOv7-d2 Repo Skill

Use this skill when the task is about the yolov7_d2 / yolov7-d2 package or about a Detectron2-style YOLOv7-d2 checkout: choosing configs, training object detectors or instance segmenters, running demos/evaluation, exporting models, converting DETR checkpoints, or debugging this repository's dependency/config/runtime issues.

This skill is self-contained operating guidance distilled from the repository. It does not assume the construction checkout is available. When a task needs runnable project code, use the user's own YOLOv7-d2 checkout or installed package plus the bundled helper scripts and references here.

First checks

1. Confirm the user's task names YOLOv7-d2, yolov7_d2, yolov7, Detectron2 configs from this repo, or repo-specific scripts such as train_det.py, demo.py, export.py, or DETR converters.
2. Check installation facts with scripts/smoke_import_and_config.py: it imports yolov7, injects add_yolo_config, and optionally merges a user config.
3. Read references/installation-and-environment.md before giving dependency or backend advice.
4. Read references/model-family-overview.md before choosing between YOLO-family, SparseInst/SOLOv2, DETR-family, YOLOF, or LazyConfig workflows.
5. Use references/troubleshooting.md for cross-cutting install/import/config failures.

Route by task

• For config selection, custom COCO datasets, training launchers, augmentations, optimizer settings, model-family config maps, or dataset-registration failures, read sub-skills/training-and-configuration/SKILL.md.
• For PyTorch demo inference, image/video visualization, confidence/NMS flags, W&B inference logging, benchmark planning, or COCO evaluation, read sub-skills/inference-and-evaluation/SKILL.md.
• For ONNX/TorchScript export, ONNXRuntime inference, DETR checkpoint conversion, TensorRT, or quantization, read sub-skills/deployment-and-export/SKILL.md.

Minimal install/import guidance

YOLOv7-d2 is built on Detectron2 and PyTorch. The repository metadata is minimal, so users usually need to install a compatible PyTorch/Detectron2 stack first, then install yolov7_d2 and the runtime packages used by their workflow.

Minimum import/config inspection typically needs:

python -m pip install torch torchvision
python -m pip install 'detectron2'  # use the Detectron2 install command matching the user's torch/CUDA stack
python -m pip install yolov7-d2 timm nbnb omegaconf pycocotools scipy alfred-py wandb
python - <<'PY'
from detectron2.config import get_cfg
from yolov7.config import add_yolo_config
cfg = get_cfg(); add_yolo_config(cfg)
print(cfg.MODEL.YOLO.CLASSES, cfg.SOLVER.OPTIMIZER, cfg.MODEL.DETR.NUM_OBJECT_QUERIES)
PY

Use CUDA only when the user needs real GPU training/inference/export. CPU is enough for config inspection, command planning, and most preflight checks; it is not a substitute for GPU throughput, TensorRT, or long training validation.

What this skill does not cover

• It does not verify mAP, reproduce benchmark tables, download model weights, or run long training jobs by default.
• It does not promise closed-source or WIP features mentioned in the README unless they are present in the user's code and verified separately.
• It treats TensorRT and quantization paths as optional, toolchain-dependent workflows unless the user provides the needed engine/model/calibration artifacts and hardware.

Provenance and router metadata

• Source baseline and evidence paths are in references/repo-provenance.md.
• Managed repo-skills router metadata is in references/repo-routing-metadata.json.