DeepDanbooru

Use this repo skill when a task names DeepDanbooru or deepdanbooru, asks for anime-style multi-label image tagging, or needs the package's project, dataset, training, evaluation, TensorFlow Lite, or Grad-CAM workflows. This skill covers the 1.0.0 source contract represented by the provenance snapshot; read it before refreshing against a changed checkout.

Route by task

• Project, tags, or dataset layout: read project-data-setup. It owns create-project, download-tags, make-training-database, project.json, tags.txt, SQLite posts, and images/<prefix>/ validation.
• Training or checkpoint recovery: read model-training. It owns config validation, supported ResNet/optimizer/loss values, data preflight, train-project, checkpoints, and .keras exports.
• Tag inference or evaluation: read inference-evaluation. It owns evaluate, evaluate-project, Python inference APIs, preprocessing, thresholds, recursive folders, and .txt sidecars.
• Export or attribution artifacts: read post-training-tools. It owns conv2tflite, TFLite checks, and experimental grad-cam output.

Use the owning sub-skill rather than trying to make the root a full API manual. The root references below contain cross-cutting install, CLI, and failure information.

Install and verify

The source package declares a base install plus a TensorFlow extra. For the repository's documented full workflow, install the TensorFlow requirements in a fresh compatible Python environment:

python -m pip install "deepdanbooru[tensorflow]"
# Or, when the distribution's published requirements file is available:
python -m pip install -r requirements.txt

The package imports TensorFlow and TensorFlow I/O through its public module and CLI paths. Do not call a CPU import proof of CUDA readiness. Start with the bundled no-download diagnostic:

python scripts/environment_smoke.py
python -m deepdanbooru --help
python -m deepdanbooru evaluate --help

The console entry point is deepdanbooru. CPU TensorFlow is the verified correctness backend for this skill. evaluate is CPU-first unless --allow-gpu is explicitly supplied; this source snapshot did not verify GPU libraries or GPU performance.

Operating boundaries

• Do not download Danbooru tags, model weights, or datasets without explicit network/credential approval. The bundled scripts are offline and safe by default.
• Do not launch full training as a smoke test. Validate the project, tags, SQLite schema, image paths, and package imports first; training is expensive and writes checkpoints/exports.
• Treat the known 1.0.0 evaluate-project loader defect as real: its native route calls the absent dd.data.load_tags_from_project symbol. For production folder inference, use evaluate --project-path ... --allow-folder, which uses the available project tag loader. The bundled project smoke helper exercises that fallback and does not claim an unverified native run.
• The native evaluate --save-txt path can fail for an empty selected-tag list and overwrites an existing sibling sidecar with w. Use the skill helper sub-skills/inference-evaluation/scripts/save_txt_guard.py when sidecar output needs a local safety boundary: it rejects empty selections and sibling directories, and requires an explicit flag to overwrite a regular file.
• Read repo-provenance.md before deciding whether this graph is stale; package/source version drift remains possible.
• Read install-and-cli.md for the full command map and dependency/version caveat.
• Read troubleshooting.md for package-wide recovery.