---

name: cli-data-embedding description: "Use LightlySSL CLIs, Hydra overrides, data layouts, YOLO crops, embeddings, lightly.core, and embedding APIs safely." disable-model-invocation: true metadata: disco-role: operating license: MIT

cli-data-embedding

Use this sub-skill when a task needs LightlySSL console commands, safe command construction, input-folder checks, YOLO crop fixtures, embedding artifacts, or the high-level lightly.core / lightly.embedding APIs.

For a fresh public environment, start with pip install lightly. Add lightly[timm] only for TIMM / MAE / ViT-style optional modules, and lightly[video] only for direct video-file datasets requiring PyAV support.

Route here for

• lightly-version, lightly-ssl-train, lightly-embed, lightly-magic, and lightly-crop command planning.
• Hydra override spelling, default CLI config keys, checkpoint and embedding output paths, and LIGHTLY_LAST_CHECKPOINT_PATH / LIGHTLY_LAST_EMBEDDING_PATH behavior.
• Image or video folder layout decisions before a CLI or LightlyDataset workflow.
• YOLO label format for lightly-crop, label-name YAML files, and tiny synthetic crop fixtures.
• Practical use of lightly.core, lightly.embedding, and Lightly-compatible embedding CSV files.

Route elsewhere

• Custom PyTorch or PyTorch Lightning SSL loops, method recipes, or distributed training plans: use training-workflows.
• Low-level losses, transforms, datasets, collate functions, model heads, and tensor-shape debugging: use ssl-building-blocks.
• KNN / linear evaluation, benchmarking, tests, docs, notebooks, and maintainer workflows: use evaluation-maintenance.

First actions

1. Identify the intended surface: CLI command, input data layout, crop labels, embedding CSV, or high-level Python API.
2. Read the relevant bundled reference rather than relying on source-repo docs:
   • CLI reference
   • Data formats
   • API reference
   • Troubleshooting
3. Use bundled helpers for safe dry-run planning and validation:
   • python scripts/build_cli_command.py --help
   • python scripts/validate_lightly_image_folder.py --help
   • python scripts/create_tiny_yolo_crop_fixture.py --help
4. Treat train, embed, magic, and crop commands as artifact-writing operations. Build or validate the command first; execute only when the user has selected the input/output paths and runtime budget.

Common safe patterns

• Start with lightly-version to confirm the console entry points are installed.
• For command construction, prefer scripts/build_cli_command.py so high-frequency Hydra keys such as input_dir, loader.batch_size, trainer.max_epochs, collate.input_size, checkpoint, label_dir, and crop_padding are spelled consistently.
• For data issues, run scripts/validate_lightly_image_folder.py <dataset-dir> before invoking a training, embedding, magic, or crop command.
• For crop command tests, generate a tiny self-contained fixture with scripts/create_tiny_yolo_crop_fixture.py <out-dir> and then inspect the printed lightly-crop command.