---

name: fcos description: "Routes FCOS object-detection repo tasks for inference demos, config/data setup, training/evaluation, ONNX export, and legacy PyTorch internals." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

FCOS Repo Skill

Use this skill when a task is about FCOS: Fully Convolutional One-Stage Object Detection, the legacy PyTorch/maskrcnn-benchmark implementation for anchor-free object detection. It covers using the public fcos package/API, selecting and validating configs, preparing dataset layouts, building training/evaluation commands, exporting ONNX models, and maintaining FCOS internals.

First Checks

1. Confirm the user is asking about FCOS-style object detection, FCOS configs, fcos / fcos_core, maskrcnn-benchmark-derived training scripts, or errors such as missing fcos_core._C.
2. Check the runtime with scripts/check_fcos_environment.py when import, CUDA, extension, or config availability matters.
3. Read references/repo-provenance.md before deciding whether this skill matches a current checkout or should be refreshed.
4. For cross-cutting install/import/build failures, read references/troubleshooting.md before changing code or dependencies.

Public Package Baseline

FCOS is a legacy source-install package named fcos with import roots fcos and fcos_core. A practical installation usually needs:

pip install torch torchvision ninja yacs cython matplotlib tqdm opencv-python scikit-image
pip install git+https://github.com/tianzhi0549/FCOS.git
python - <<'PY'
from fcos_core.config import cfg
print(cfg.MODEL.FCOS_ON)
PY

For real model inference/training, a compiled fcos_core._C extension and a PyTorch/CUDA combination compatible with this older maskrcnn-benchmark code are required. Modern PyTorch releases can need compatibility patches; do not promise that a CPU import proves real detector inference.

Route Map

• Image inference, demos, and public FCOS API → sub-skills/inference-demo/SKILL.md. Use this for installed fcos CLI usage, FCOS.detect, model-name choices, image preprocessing, CPU/GPU selection, output boxes, visualization, and demo troubleshooting.
• Training, evaluation, distributed launch, checkpoints → sub-skills/training-evaluation/SKILL.md. Use this for --config-file, MODEL.WEIGHT, OUTPUT_DIR, TEST.IMS_PER_BATCH, multi-GPU launch, OOM mitigation, and stripping solver states.
• Configs, model catalog, and dataset layouts → sub-skills/data-configs/SKILL.md. Use this for configs/fcos selection, MODEL.FCOS keys, COCO/VOC/Cityscapes layout validation, custom dataset planning, and config merge diagnostics.
• ONNX export and ONNX evaluation planning → sub-skills/onnx-export/SKILL.md. Use this for FCOS ONNX export commands, output tensor names, dummy input constraints, and Caffe2/backend limitations.
• Source internals and repo maintenance → sub-skills/internals-maintenance/SKILL.md. Use this for FCOS head/loss/postprocess internals, BoxList, compiled extension issues, tests, and porting to newer PyTorch.

Common Decisions

• If the user only wants to detect objects in one image, route to inference-demo and prefer the high-level fcos.FCOS API or the installed fcos CLI.
• If the user mentions COCO/VOC/Cityscapes paths, config overrides, class counts, or YAML errors, route to data-configs before training.
• If the user asks for AP numbers, model weights, batch sizes, or an evaluation command, route to training-evaluation and remind them that datasets and weights are not bundled.
• If the user asks for ONNX, do not treat it as normal PyTorch evaluation; route to onnx-export because the export script emits backbone/head tensors and the test script does PyTorch post-processing.
• If the user is editing FCOS code or seeing _C, compiler, torch._six, libc10, or NumPy ABI errors, route to internals-maintenance plus the root troubleshooting reference.

Safety and Scope Notes

• Do not download pretrained weights, datasets, or run long training/evaluation unless the user approves the cost and runtime.
• Do not open GUI windows or webcams by default; use no-display/API paths for automated agents.
• Do not claim benchmark reproduction unless the required dataset, weights, compiled extension, and GPU runtime were actually verified.
• The bundled scripts are safe diagnostics and command builders. They are replacements for routine guidance, not large model weights or datasets.