CenterNet

CenterNet is a COCO-based object-detection repository with two main command-line entry points, train.py and test.py, plus legacy compiled extensions for NMS, custom pooling, and the COCO Python API.

Use this skill when you need to train a CenterNet model, run evaluation on COCO splits, choose between the CenterNet-52 and CenterNet-104 configs, or diagnose the repo's runtime/build prerequisites.

Start here

1. Read references/workflows.md for the end-to-end training and evaluation flow.
2. Read references/cli-reference.md for the train.py and test.py flags and output paths.
3. Read references/configuration.md before changing model or dataset settings.
4. Read references/data-layout.md before pointing the repo at COCO data or caches.
5. After building the COCO API and custom extensions, run scripts/check_install.py --repo-root <checkout> to verify imports, CUDA, and the compiled extensions.
6. Read references/troubleshooting.md when an import, build, GPU, or data-path check fails.
7. Read references/repo-provenance.md before deciding whether this skill is current for a checkout.

Installation and setup

• Use a CUDA-capable Python environment. The training and test code call .cuda() and expect a working NVIDIA backend.
• Install the runtime stack used by the repo: torch, numpy, opencv-python, matplotlib, tqdm, scipy, h5py, Cython, and pycocotools.
• Build the COCO Python API in data/coco/PythonAPI and the custom CenterNet extensions in external/ and models/py_utils/_cpools/.
• If the extension build fails on a modern toolchain, read references/troubleshooting.md before retrying with a different version mix.

Common routes

• Train a detector or resume from a checkpoint: references/workflows.md
• Run validation, testing, or multi-scale evaluation: references/workflows.md
• Compare CenterNet-52 and CenterNet-104, or review config keys: references/configuration.md
• Understand COCO directory layout, cache files, and result output: references/data-layout.md
• Diagnose missing compiled modules, CUDA, checkpoints, or COCO paths: references/troubleshooting.md

Quick checks

• python scripts/check_install.py --repo-root <checkout> after the compiled extensions are built
• python train.py --help
• python test.py --help

Notes

• The repository is not published as a packaged wheel; work from a repository checkout.
• config.py defines the runtime defaults that train.py and test.py load through the selected JSON config.
• The shipped dataset registry currently contains MSCOCO only.