---

name: serialization-and-reproducibility description: "Save, load, replay, inspect, and reproduce Albumentations augmentation pipelines with JSON/YAML serialization, ReplayCompose, applied-parameter tracking, custom Lambda mappings, deterministic seeds, and Hub caveats." disable-model-invocation: true metadata: disco-role: operating license: MIT

Albumentations serialization and reproducibility

Use this sub-skill when a task involves persisting augmentation pipelines, comparing a loaded pipeline with the original, replaying exact random choices, recording transform parameters, or diagnosing reproducibility surprises.

Route first

• Build or edit the initial A.Compose / A.ReplayCompose structure with ../pipeline-composition/ before serializing it.
• Choose transform families and parameter names with ../transform-catalog/ before saving a config.
• Confirm image, mask, bbox, keypoint, volume, and additional-target schemas with ../targets-and-formats/ before replaying across data.
• Handle PyTorch Dataset and dataloader-worker reproducibility with ../framework-integration/ after the Albumentations pipeline itself is correct.

Use the bundled materials

• references/serialization-api.md: Use for A.to_dict, A.from_dict, A.save, A.load, JSON/YAML buffers/files, schema shape, custom transform mapping, and Hub save/load methods.
• references/reproducibility.md: Use for ReplayCompose, ReplayCompose.replay, save_applied_params, applied_transforms, seed, set_random_seed, and exact-output validation checks.
• references/troubleshooting.md: Use for nonserializable Lambda failures, missing nonserializable keys, YAML dependency errors, replay misuse, shape changes before replay, and Hugging Face Hub network/auth limitations.
• scripts/roundtrip_pipeline.py: Run a tiny local smoke check that saves and loads a simple pipeline, resets seeds, verifies equal output, and optionally writes JSON or YAML.

Standard workflow

1. Make the pipeline serializable: prefer built-in transforms; give A.Lambda a stable name; keep custom callables available in code that loads the config.
2. Round-trip the configuration with A.to_dict/A.from_dict or A.save/A.load, then compare to_dict() outputs and seeded augmentation outputs on a tiny fixture.
3. For exact reuse of one random draw, capture result["replay"] from A.ReplayCompose and pass it to A.ReplayCompose.replay with data of compatible targets and shapes.
4. For debugging, use A.Compose(..., save_applied_params=True) and inspect result["applied_transforms"] for transform names and sampled params.
5. Treat Hub methods as optional sharing helpers: local save_pretrained/from_pretrained use JSON config files; remote push/download requires huggingface_hub, network access, and credentials when needed.