---

name: model-customization description: "Guides SimpleDet symbolic detector composition, config-first model changes, component contracts, model-family selection, and custom operator or checkpoint compatibility decisions." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Model customization

Use this route when a task asks to add or change a detector, backbone, neck, RPN, RoI extractor, bbox/mask head, metric, custom operator, or model-family config. Read architecture-api.md for contracts and model-families.md for family routing.

Config-first versus code

Start with a copied config for exposed parameters: backbone depth, FPN/neck, anchors, proposal limits, ROI shape, class count, normalizer, FP16, schedule, dataset split, NMS, or checkpoint prefix. Use code only for new symbols, inputs/labels, custom operators, losses, branches, or output contracts. Read customization-workflows.md for the sequence and troubleshooting.md for shape/name and operator failures.

Architecture route

The usual composition is Backbone -> Neck -> RpnHead -> RoiExtractor -> BboxHead, with optional MaskHead, detector, postprocessor, metric, and input transform. Detector classes expose train/test symbol methods; configs own the actual composition and returned data/label names. External mxnext TVM/custom operators and CUDA support are required for many families, so source inspection is not runtime proof.

Use scripts/list_model_families.py to inventory available family/config signals without importing MXNet or constructing symbols. Route actual execution through detection-workflows and backend setup through setup-and-operations.