---

name: data-preparation description: "Routes ALAE dataset preparation, TFRecord layout validation, sample-image setup, style-mixing image layout, and face-alignment workflows." disable-model-invocation: true metadata: disco-role: operating license: NO_LICENSE

ALAE Data Preparation

Use this sub-skill when a task asks for ALAE data inputs: TFRecord conversion or splitting, config dataset paths, sample-image directories, style-mixing src/dst image folders, or face alignment for FFHQ/CelebA-style inputs.

Route by request

• Validate a config or checkout layout first: read data-layouts, then run scripts/validate_alae_data_layout.py against the user's config. This is the safest first step before training, reconstruction, style mixing, or metrics.
• Prepare TFRecords: read TFRecord preparation. The original dataset scripts are TensorFlow 1.x, raw-data dependent, and often large or network-bound; treat commands there as explicit, checkout-root operations.
• Align face images: read face alignment, then use scripts/align_faces_alae.py with explicit --input-dir, --output-dir, and --predictor paths.
• Debug data errors: read troubleshooting for TensorFlow 1.x API failures, PYTHONPATH issues, missing dlib predictors, stale /data/datasets assumptions, disk/network side effects, and malformed style/sample layouts.

Boundaries

• This sub-skill owns dataset paths from ALAE configs: DATASET.PATH, DATASET.PATH_TEST, PART_COUNT, PART_COUNT_TEST, SAMPLES_PATH, STYLE_MIX_PATH, MAX_RESOLUTION_LEVEL, and OUTPUT_DIR as they affect data readiness.
• Route full training launches and checkpoint interpretation to the training sub-skill (../training/SKILL.md).
• Route pretrained model downloads, generation, reconstruction, and style-mixing execution to the generation sub-skill (../generation/SKILL.md) after this sub-skill validates input layouts.
• Route FID/PPL/LPIPS metric execution to the metrics sub-skill (../metrics/SKILL.md) after TFRecords and paths are validated here.

Safe bundled helpers

python scripts/validate_alae_data_layout.py --help
python scripts/align_faces_alae.py --help

Both helpers avoid network, training, and hard-coded checkout paths by default. Pass an ALAE checkout root or explicit input/output paths when validating or transforming user data.