Libreyolo Checkpoint Metadata

Work with LibreYOLO checkpoint metadata

Metadata is the model-loading source of truth (REVIEW.md axiom): a .pt is a torch.save() dict whose top-level keys identify family, size, task, classes, and input size, so loading never depends on filename parsing or state-dict sniffing. The contract is docs/checkpoint_schema.md (v1.0 required keys, task-specific extras, training-checkpoint fields, export runtime metadata). Read that doc for the schema; do not trust a from-memory key list, including this skill's examples.

Inspect (first move for any load problem)

libreyolo metadata path=weights/LibreYOLO9t.pt      # raw embedded metadata
libreyolo info model=LibreYOLO9t.pt                 # resolved family/size/task/nc/device

Python, when you need it programmatic:

from libreyolo.utils.serialization import load_untrusted_torch_file, validate_checkpoint_metadata
ckpt = load_untrusted_torch_file("file.pt")   # weights_only-safe load for unknown files
validate_checkpoint_metadata(ckpt)            # raises CheckpointMetadataError with specifics

Produce (converters, trainers, tools)

The helpers in libreyolo/utils/serialization.py are the only sanctioned writer/reader; hand-assembled metadata dicts drift from the schema:

• wrap_libreyolo_checkpoint(...): builds the v1.0 wrapper around a state dict. Every weights/convert_*_weights.py script ends in this.
• validate_checkpoint_metadata(...): strict check; call it in the converter after writing and in tests.
• normalize_checkpoint_names(...) / build_class_names(...): names-dict hygiene (dict[int, str], keys 0..nc-1).
• load_untrusted_torch_file vs load_trusted_torch_file: use the untrusted loader for anything a user hands you.

Two checkpoint shapes, one schema core:

• Lean/published: state dict + required metadata only. What HF repos ship. No optimizer, epoch, config, or EMA-resume keys.
• Training: adds epoch, optimizer, config, EMA fields (is_ema_weights declares whether model is the EMA weights). best.pt and last.pt are these. Strip to lean before publishing (libreyolo-upload-hf-model validates this).

Debug a load failure (the decision tree)

1. "Foreign/unrecognized checkpoint": it is an upstream file, not a LibreYOLO one. Route: the family's weights/convert_*_weights.py, or the runtime auto-converter (libreyolo/models/autoconvert.py + per-family recognizers) which handles known upstream layouts transparently. If a known upstream format is rejected, the recognizer is the bug.
2. Cross-family or cross-task rejection ("checkpoint is family X, model is Y"): working as designed; these must fail loudly. Never "fix" by loosening the check; load with the right class/task instead.
3. Metadata invalid (missing keys, bad names range, wrong schema_version): the producer is at fault; fix the converter/trainer and regenerate, do not hand-patch the file (a hand-patched checkpoint in the wild is a provenance mystery later).
4. Legacy LibreYOLO checkpoint (pre-v1.0, loads with a warning): the compatibility path works but is legacy; recommend re-converting. State- dict sniffing exists only on this path and must not spread to new code.
5. Class-count surprises (nc=90 vs 80 RF-DETR-style): read the "Legacy And Foreign Weights" section of the schema doc before touching anything; the COCO-normalization rules there are subtle, deliberate, and test-covered.

Change the contract (rarely, and never quietly)

Adding/removing/reinterpreting a metadata key is a contract change. CONTRIBUTING.md hard-requires: update docs/checkpoint_schema.md and the helpers in libreyolo/utils/serialization.py in the same PR. Also:

• New required keys break every existing published weight; prefer optional keys with reader defaults (the crop_pct/interpolation classify keys are the pattern).
• Exported-runtime metadata (ONNX props etc.) is a parallel surface with its own section in the schema doc; exporter-writes and backend-reads must change together (libreyolo-export-formats).
• Task-specific extras (pose keypoint fields, restore degradation) belong in the task's schema section, defined when the task lands (libreyolo-add-task step 9).
• The unit tests around serialization and per-family load/reject (tests/unit/test_serialization*.py, test_*_can_load*-style) are the gate; extend them with the change.

Related

• docs/checkpoint_schema.md: the contract (always read it, it moves).
• skills/libreyolo-port-model/: converter templates that emit v1.0.
• skills/libreyolo-upload-hf-model/: pre-upload validation of the file.
• skills/libreyolo-export-formats/: the runtime-metadata twin surface.