inference-and-models

RF-DETR Inference and Models

Use this sub-skill when the task is to run or diagnose RF-DETR prediction code, choose a detection / segmentation / keypoint model class, load a checkpoint, interpret supervision outputs, handle class names, tune inference memory, or explain Plus model boundaries.

Do not use this sub-skill for fine-tuning, Lightning CLI, dataset schema work, export/deployment formats, or source-repository contribution policy; route those tasks to sibling sub-skills when available:

• ../training-and-cli/SKILL.md for train(), evaluate(), datasets, configs, and CLI.
• ../export-and-deployment/SKILL.md for ONNX, TensorRT, TFLite, ExecuTorch, CoreML, and Roboflow deployment bundles.
• ../repository-development/SKILL.md for modifying RF-DETR source, tests, docs, packaging, or CI.

Runtime map

• API surface, class tables, signatures, checkpoint trust rules, output fields, and label mapping: references/api-reference.md
• Recipes for image, batch, video, webcam, RTSP, keypoint visualization, checkpoint loading, and memory optimization: references/workflows.md
• Failure diagnosis for imports, downloads, checkpoints, shapes, labels, keypoints, video sources, devices, and Plus models: references/troubleshooting.md
• Safe installed-package inspector that does not download weights or instantiate models: scripts/inspect_rfdetr_models.py

Quick decision rules

1. Default new detection examples to RFDETRSmall / "rfdetr-small"; do not introduce new RFDETRBase or "rfdetr-base" examples.
2. Use a sized segmentation model (RFDETRSegNano, RFDETRSegSmall, RFDETRSegMedium, RFDETRSegLarge, RFDETRSegXLarge, RFDETRSeg2XLarge); do not use RFDETRSegPreview for new work.
3. Keypoints are preview-only: use RFDETRKeypointPreview / "rfdetr-keypoint-preview" only for keypoint tasks.
4. Prefer detections.data["class_name"] or key_points.data["class_name"] for labels. COCO_CLASSES is only for manually mapping COCO-pretrained sparse category IDs.
5. For fine-tuned checkpoints, load with RFDETR.from_checkpoint(path) or rfdetr.from_checkpoint(path) and leave trust_checkpoint=False unless the checkpoint source is fully trusted.
6. For memory-constrained prediction, consider include_source_image=False first; use model.inference(compile=False, inplace=True, dtype="float16") only when the destructive, inference-only trade-off is acceptable.