libreyolo-port-model

Port a model into LibreYOLO

This skill is executable. It assumes you've been asked to port a model and your job is to ship it. Reference material (architectural patterns, ABC contracts, validators) is in the back half; the front half is the path you follow to write code.

The legacy detection-only skill is preserved at skills/legacy_libreyolo-add-native-detection-model/SKILL.md for archaeology. Don't follow it.

0. Read this first — orientation

You have an upstream model. Six questions get you pointed at the right scaffold:

1. Tier: checkpoint-driven (a BaseModel factory family — everything below), or prompt-driven (promptable segmentation, open-vocab detection, VLM)? Prompt-driven models join a sibling factory (LibreSAM / LibreOpenVocab / LibreVLM), not the BaseModel registry — see §4.1.
2. Architecture: per-anchor head with NMS (YOLO-grid), set-prediction with Hungarian matching (DETR), or one-to-one head with top-K and no NMS (NMS-free YOLO-grid)?
3. Tasks shipped: detect / pose / segment (this skill's templates), or classify / semantic / depth / restore / point / gaze (clone the merged exemplar family — §10 lists them)?
4. Backbone source: standard PyTorch, vendored separately-licensed (e.g. DINOv3), or loaded from an optional dependency (e.g. transformers — RF-DETR's DINOv2 backbone)?
5. License: code (upstream LICENSE file) and weights (HF model card YAML). Both must be permissive (MIT / Apache-2.0 / BSD) for the family to live in core.
6. Maturity target: inference-only, gated experimental training, or production training with e2e parity?

The answers route you through §1 (license gate) → §3 (pick scaffold) → §4 (per-family ledger entry to clone) → §5 (commit sequence) → §6 (paste-ready templates). Reference sections fill in the contract details.

Two concrete starts:

• RTMO (one-stage pose) → YOLO-grid pose, Apache-2.0 (MMPose-based). Closest scaffolds: YOLOX for the CSPDarknet backbone, RTMDet (models/rtmdet/) as the worked example of the mm-series route — mm-series upstream, key remap in models/rtmdet/convert.py, gated-experimental training. (RTMDet itself used to be the hypothetical example here; it has since shipped following exactly this skill.)
• A random model on HF → Run §1 (license check) first. If permissive, find architecture in §4 ledger; if no row matches, fall back to §3 decision tree.

1. License check — both code and weights

Two absolute rules. No exceptions, no "just to check", no exploratory read.

1. Never read AGPL code — ever. Do not open, clone, curl, cat, or otherwise view the source of an AGPL-licensed project, not even to "understand the architecture" or "check a shape". Once you determine a repo is AGPL (or dual-licensed with AGPL as the open option), stop and port from a permissive source, the paper, or a clean-room spec instead. Reading it contaminates the port. This holds even when the AGPL repo is the most convenient reference.
2. Never point at another repo's assets — especially an AGPL repo's — for datasets or weights. Do not reference, hotlink, download-at-runtime, or document a URL into a third-party (and above all an AGPL) repository for dataset files, label files, or checkpoints. Ship dataset config that resolves to a permissively-licensed source we control or that the user supplies; if the only copy lives in an AGPL/non-redistributable repo, treat it as unavailable and surface it as a licensing decision.

If either rule would block the port, that is the correct outcome — raise it, don't route around it.

LibreYOLO core is MIT and stays MIT-compatible. Check both before writing any code.

curl -sL https://raw.githubusercontent.com/<org>/<repo>/<branch>/LICENSE | head -3

For weights, check the HF model card YAML at the top:

license: apache-2.0    # ← permissive
license: gpl-3.0       # ← copyleft
license: cc-by-nc-4.0  # ← non-commercial: usually a hard "no"

Already-shipped reference cases:

• D-FINE / DEIM / RF-DETR / EC — Apache-2.0 → clean.
• DEIMv2 — Apache-2.0 family code, but s/m/l/x sizes vendor DINOv3 (Meta's custom non-OSI license). Not "Apache-2.0 clean" wholesale.
• YOLOv9 — MIT, via MultimediaTechLab/YOLO (Kin-Yiu Wong & Hao-Tang Tsui). Not WongKinYiu/yolov9 (GPL-3.0). When upstream has multiple forks, pick the permissive one.
• YOLO-NAS — Apache-2.0 code + Deci CDN weights (non-redistributable) → weights linked, not rehosted.
• YOLO-World — GPL-3.0 → flagged plugin-only in #108, never merged.
• L2CS-Net — portable code but redistribution-restricted weights → inference-only family, weights not rehosted, and the family is skiplisted in runtime auto-conversion (_SKIP_FAMILIES in models/autoconvert.py).

If a license problem surfaces mid-port, never "fix" it by rewriting or renaming the offending code to obscure where it came from. Surface it as a licensing decision — plugin, link-out, or drop.

Vendored sub-components with separate licenses

If your family vendors a separately-licensed architectural sub-component (DEIMv2 vendors DINOv3 for s/m/l/x backbones; Meta's custom non-OSI license), the family's own license isn't sufficient. Required:

• LICENSE.md next to the vendored code (e.g. libreyolo/models/deimv2/engine/backbone/dinov3/LICENSE.md).
• A NOTICE entry citing the sub-component, its upstream, and its license.
• A docstring annotation on the family's model class explaining which size variants depend on the differently-licensed sub-component.
• A clear rule for the user: do the constraints of the sub-component flow through to the produced weights? Document the answer.

Don't quietly bundle a non-OSI sub-component without these three artifacts.

2. Pick your maturity target

Pick one and document scope explicitly:

1. Inference-only. No trainer wired, or trainer rejects all tasks in train(). The user can model.predict(...) but not model.train(...). YOLO-NAS pose, EC pose, EC seg, L2CS, CLIP are here.
2. Gated experimental training. Trainer exists and train() accepts the task, but raises unless the user passes allow_experimental=True. Use when loss + assignment match upstream but small-dataset fine-tune convergence is unvalidated. PicoDet detect (models/picodet/model.py:157-201), EC detect (models/ec/model.py:269-310), and RTMDet detect (allow_experimental in models/rtmdet/model.py) are gated this way today.
3. Production-grade training. test_rf1_training passes; row in MODEL_CATALOG; recipe gaps documented in family docstring. YOLOX, YOLOv9, YOLO-NAS detect, D-FINE, DEIM, DEIMv2, RT-DETR, RF-DETR sit here, plus the classify families (MobileNetV4 / ConvNeXt / EfficientNetV2 / ResNet) on the shared classify path.

Inference-only is a legitimate ship state. Don't gate the port on a working trainer.

3. Pick your scaffold

Use this decision tree to pick the family you'll clone as your starting point.

3.1 By architecture

3.2 By task scope

3.3 By non-PyTorch upstream

• Paddle / TensorFlow / safetensors: PicoDet ports a community PyTorch reimplementation (Bo's). For Paddle direct, you'd write a heavier conversion script that handles framework-specific buffer cleanup. For safetensors, see DEIMv2's weights/convert_deimv2_weights.py:19-43 — dispatch on Path(input).suffix == ".safetensors", build a fresh native model, safetensors.torch.load_model(model, path, strict=True).

4. Per-family ledger

Dense reference. Find the family closest to your port and clone its directory as your starting scaffold. Each row tells you what's already solved.

File-count signal: 6–7 files = single-task YOLO-grid scaffold. 16+ = a DETR family with non-trivial loss + matcher + transforms.

Families added since the table above was first written

Compact rows — clone these directly for the newer archetypes:

4.1 Sibling factory tiers (not BaseModel families)

Prompt-driven models do not join the checkpoint-driven BaseModel registry. They live in sibling factories with their own contracts:

• LibreOpenVocab (models/openvocab/) — text-conditioned open-vocabulary detectors returning standard detection Results: Grounding DINO (models/grounding_dino/) and OWLv2 (models/owlv2/) are aliased today; the native OMDet-Turbo port (models/omdet_turbo/) is built for this tier. Shared towers live in models/bert/ (text) and models/swin/ (vision).
• LibreSAM — promptable segmentation: SAM-1 and SAM-2 (models/sam/), MobileSAM (models/mobilesam/).
• LibreVLM (models/vlm/) — vision-language models.

If your port is prompt-driven, clone one of these tiers instead of a factory family. The license gate (§1), parity discipline (§12), and HF-upload rules (commit 10) still apply in full; the BaseModel ABC contract (§8) does not.

5. Walking the port — commit sequence

Each numbered commit is a self-contained PR-able unit. Don't combine. Don't write the trainer before the inference parity test passes (see §12).

Commit 1 — Skeleton + factory recognition

Create libreyolo/models/<family>/{__init__.py, model.py, nn.py, utils.py} using template §6.1. Implement:

• Libre<FAMILY> class with FAMILY, FILENAME_PREFIX, INPUT_SIZES, SUPPORTED_TASKS, DEFAULT_TASK, TRAIN_CONFIG = None (we'll wire later).
• can_load(state_dict) — pick a key unique to your architecture. Never "backbone" or "weight".
• detect_size(state_dict) — infer size from a shape signature.
• detect_nb_classes(state_dict) — read nc from the head.
• _init_model, _get_available_layers, _forward, _postprocess (stub OK), _preprocess (use shared letterbox).

Add from .<family>.model import Libre<FAMILY> to libreyolo/models/__init__.py in the right registry order (most distinctive markers first). Add Libre<FAMILY> to libreyolo/__init__.py exports.

Verify: python -c "from libreyolo import Libre<FAMILY>; m = Libre<FAMILY>(size='s'); print(m.task, m.family)" runs.

Commit 2 — nn.py + forward smoke

Port the model architecture into libreyolo/models/<family>/nn.py. Mirror upstream attribute names where possible — it makes conversion a metadata-wrap (see Commit 3) and keeps the parity diff readable.

Verify: model builds at all sizes, model(torch.zeros(1, 3, 640, 640)) returns the expected tensor shape (or dict for DETR).

Commit 3 — Conversion script + atomic write

Create weights/convert_<family>_weights.py using template §6.4 (single-task) or §6.5 (multi-task). Use wrap_libreyolo_checkpoint(...) with task, supported_tasks, default_task populated even for single-task ports — free disambiguation later.

Write atomically: tmp = output.with_suffix(".tmp"); save_checkpoint(wrapped, tmp); tmp.rename(output). Print a missing/unexpected-key diff after loading the wrapped dict into a fresh model.

Verify: python weights/convert_<family>_weights.py upstream/x.pth weights/Libre<FAMILY>s.pt --size s runs and produces a file. LibreYOLO("weights/Libre<FAMILY>s.pt") loads without errors.

Commit 3b — Runtime auto-conversion recognizer

libreyolo/models/autoconvert.py makes LibreYOLO("<upstream_file>.pth") work directly: the factory unwraps the common upstream layouts (ema.module / ema / net / model / state_dict / plain), asks every registered family via BaseModel.convert_upstream_state_dict whether it recognizes the tensors, wraps the winner in v1.0 metadata (size / task / nc read from the tensors), and writes <source>-<Prefix><size>[-task].pt beside the source.

• The BaseModel default (models/base/model.py::convert_upstream_state_dict) claims any layout your can_load accepts — if upstream keys already match your native port, auto-conversion is free and this commit is just tests.
• If upstream naming differs, add models/<family>/convert.py with is_upstream_state_dict() + convert_upstream() (shared by the runtime auto-converter and your offline weights/convert_<family>_weights.py), and override the hook on the model class (template §6.10). Precedents: models/{yolo9,rtmdet,rtdetr,rtdetrv4,picodet}/convert.py.
• Special cases: non-redistributable weights are skiplisted (_SKIP_FAMILIES — L2CS); RF-DETR registers a bespoke recognizer in autoconvert.py because it needs the full checkpoint (size detection + COCO class remap), not just the tensor dict.
• Arbitration when several families claim one file: subclass beats base, then registry order (= import order in libreyolo/models/__init__.py). The filename is consulted only for the D-FINE/DEIM tensor tie.

Verify: LibreYOLO("<downloaded upstream>.pth") loads your family end to end, AND your recognizer returns None for every sibling family's upstream checkpoints — a greedy recognizer steals other families' files (landmine #36).

Commit 4 — Inference parity proof

This is the gate. Before any postprocess work, prove the model produces identical outputs to upstream on identical inputs.

Use template §6.8 (cross-load script). Save it as tests/unit/test_<family>_parity.py or as a one-off under weights/. The test:

1. Imports the upstream model class and yours side by side.
2. Loads upstream weights into both.
3. Feeds the same torch.zeros(...) (or a fixed seed) through both.
4. Asserts max_abs_diff == 0 in eval() mode on layers present in both.

Verify: parity script passes for every size of your family. Don't proceed until it does.

Commit 5 — _postprocess returning the canonical dict

Implement _postprocess in models/<family>/utils.py. Return:

{"boxes": (N, 4), "scores": (N,), "classes": (N,)}                # detect
# + "masks": (N, H, W) for segment
# + "keypoints": (N, K, 3) for pose (xy + visibility)

If your model emits keypoints as (N, K, 2), append a column of ones for visibility — Keypoints.has_visible requires column 3 (libreyolo/postprocess/ec.py::postprocess_pose for the precedent).

Verify: tests/unit/test_<family>_postprocess.py smoke-tests shape contracts on synthetic input.

Commit 6 — End-to-end inference

LibreYOLO("Libre<FAMILY>s.pt").predict("test.jpg") returns a Results object with the right slots populated, and Results._select(idx) slices boxes ↔ masks ↔ keypoints in lockstep. Drawing dispatches on slot presence (if result.keypoints is not None: draw_keypoints(...)), no task field needed.

Add <Family>ValPreprocessor to libreyolo/validation/preprocessors.py (or inherit existing). Set uses_letterbox, custom_normalization, wants_unresized_image properties to match your training transform.

Verify: model.predict("test.jpg") runs end-to-end. model.val(data="coco128.yaml") runs.

Commit 7 — ONNX export

For YOLO-grid: 1 output "output", opset 13. For DETR: 2 outputs ["pred_logits", "pred_boxes"], opset ≥ 16 (deformable attention needs grid_sample).

DETR families: add yourself to BaseBackend._is_nms_free_family() at libreyolo/backends/base.py:211. NMS-free YOLO-grid (yolo9_e2e-style) must also be added or exported backends will wrongly apply NMS. NCNN doesn't work for DETR — block early with NotImplementedError.

Verify: model.export(format="onnx") produces a working graph. OnnxBackend("Libre<FAMILY>s.onnx").predict(...) matches PyTorch outputs.

Commit 8 — Trainer (skip if shipping inference-only)

Decide maturity (§2). Inference-only: TRAIN_CONFIG = None, override train() to raise NotImplementedError, and stop here. Otherwise:

Create models/<family>/trainer.py using template §6.6 (YOLO-grid) or §6.7 (DETR with per-group LR via lr_ratio). Append <Family>Config(TrainConfig) to libreyolo/training/config.py (or use a family-local config — RF-DETR / RT-DETR / YOLOv9-E2E do this).

If gated experimental: override train() with an allow_experimental=True flag that raises a RuntimeError with a detailed rationale otherwise. See models/picodet/model.py:157-201 for the pattern.

For non-detect tasks: override best_metric_key explicitly. Default is "metrics/mAP50-95" (bbox). Set "metrics/mAP50-95(M)" for segment, "metrics/keypoints_mAP50-95" for pose.

Verify: model.train(data="coco128.yaml", epochs=3) runs and produces a new .pt.

Commit 9 — e2e catalog row

Append your family's sizes to MODEL_CATALOG in tests/e2e/conftest.py:417. Run:

pytest tests/e2e/test_val_coco128.py -k <family>
pytest tests/e2e/test_rf1_training.py -k <family>

DETR families: skip the last_loss < first_loss assertion in test_rf1_training (DETR loss is too noisy on small datasets — RF-DETR and D-FINE both exempt themselves).

Verify: both e2e tests pass for every size.

Commit 10 — HuggingFace upload

Only for redistributable weights — re-check §1. Permissive-licensed weights are rehosted under LibreYOLO/; anything else follows the YOLO-NAS link-out or L2CS no-rehost path and this commit becomes documentation of where the user gets the weights instead.

Follow the libreyolo-upload-hf-model skill. Cross-check your filename against the whitelist there before uploading. The 5-file contract: .gitattributes, README.md, LICENSE, NOTICE, Libre<FAMILY><size>[-<task>].pt.

Multi-task families: one HF repo per task variant (LibreYOLO/Libre<FAMILY>s for detect, LibreYOLO/Libre<FAMILY>s-pose for pose, LibreYOLO/Libre<FAMILY>s-seg for segment).

Verify: LibreYOLO.from_pretrained("LibreYOLO/Libre<FAMILY>s") works on a fresh machine / cleared cache.

6. Paste-ready templates

Copy these, fill in the # TODO markers. Mirror upstream attribute names in nn.py whenever possible — it makes conversion a metadata-wrap.

6.1 Family directory layout

libreyolo/models/<family>/
├── __init__.py        # exports Libre<FAMILY>
├── model.py           # BaseModel subclass (template 6.2 or 6.3)
├── nn.py              # the actual nn.Module — port-specific, mirror upstream names
├── utils.py           # postprocess, preprocess_numpy
├── loss.py            # only if you ship a trainer
└── trainer.py         # BaseTrainer subclass (template 6.6 or 6.7) — optional

__init__.py:

"""Libre<FAMILY> family: <one-line architecture summary>."""
from .model import Libre<FAMILY>

__all__ = ["Libre<FAMILY>"]

6.2 model.py — single-task detect-only

"""Libre<FAMILY>: BaseModel subclass wiring <FAMILY> into the LibreYOLO factory."""

from __future__ import annotations
from typing import Any, Optional

import torch
import torch.nn as nn

from ...training.config import <FAMILY>Config  # delete if TRAIN_CONFIG=None
from ...validation.preprocessors import <FAMILY>ValPreprocessor
from ..base import BaseModel
from .nn import Libre<FAMILY>Model
from .utils import postprocess as _postprocess
from .utils import preprocess_numpy as _preprocess_numpy


class Libre<FAMILY>(BaseModel):
    """<one-line architecture summary>."""

    FAMILY = "<family>"                              # TODO: short lower-case ID
    FILENAME_PREFIX = "Libre<FAMILY>"                # TODO: PascalCase, no -det suffix
    INPUT_SIZES = {"s": 640, "m": 640, "l": 640}     # TODO: confirm with upstream
    SUPPORTED_TASKS = ("detect",)
    DEFAULT_TASK = "detect"
    TRAIN_CONFIG = <FAMILY>Config                    # or None for inference-only
    val_preprocessor_class = <FAMILY>ValPreprocessor

    @classmethod
    def can_load(cls, weights_dict: dict) -> bool:
        # TODO: pick a key UNIQUE to your architecture.
        # Never "backbone", "weight", or anything generic.
        # Cross-check against every existing family's state dict in tests.
        return any("<unique_token>" in k for k in weights_dict)

    @classmethod
    def detect_size(cls, weights_dict: dict) -> Optional[str]:
        # TODO: read a shape signature that disambiguates sizes.
        # Common pattern: a head conv's out_channels.
        key = "<head.cls_pred.weight>"
        if key not in weights_dict:
            return None
        out_ch = int(weights_dict[key].shape[0])
        return {<channels>: "s", ...}.get(out_ch)

    @classmethod
    def detect_nb_classes(cls, weights_dict: dict) -> Optional[int]:
        # TODO: read nc from a head weight shape, accounting for reg channels.
        key = "<head.cls_pred.weight>"
        if key not in weights_dict:
            return None
        return int(weights_dict[key].shape[0])  # adjust if reg channels are mixed in

    def _init_model(self) -> nn.Module:
        return Libre<FAMILY>Model(size=self.size, nc=self.nb_classes)

    def _get_available_layers(self) -> dict[str, nn.Module]:
        return {
            "backbone": self.model.backbone,
            "neck": self.model.neck,
            "head": self.model.head,
        }

    @staticmethod
    def _get_preprocess_numpy():
        return _preprocess_numpy

    def _preprocess(self, image, *, color_format=None, **kwargs):
        # TODO: pick the shared letterbox helper or implement a family-local one.
        # See models/yolox/model.py or models/picodet/model.py for precedents.
        ...

    def _forward(self, x: torch.Tensor) -> Any:
        return self.model(x)

    def _postprocess(self, raw, conf_thres: float, iou_thres: float, **kwargs):
        return _postprocess(raw, conf_thres, iou_thres, **kwargs)

6.3 model.py — multi-task (detect + pose [+ segment])

Add per-task class vars and dispatch. EC is the 3-task reference (models/ec/model.py).

class Libre<FAMILY>(BaseModel):
    FAMILY = "<family>"
    FILENAME_PREFIX = "Libre<FAMILY>"
    SUPPORTED_TASKS = ("detect", "pose", "segment")  # subset as needed
    DEFAULT_TASK = "detect"

    INPUT_SIZES      = {"s": 640, "m": 640, "l": 640}
    POSE_INPUT_SIZES = {"s": 640, "m": 640, "l": 640}  # may add asymmetric sizes (e.g. "n")
    SEG_INPUT_SIZES  = {"s": 640, "m": 640, "l": 640}
    TASK_INPUT_SIZES = {
        "detect":  INPUT_SIZES,
        "pose":    POSE_INPUT_SIZES,
        "segment": SEG_INPUT_SIZES,
    }

    # State-dict discriminators — cross-test against sibling families!
    _POSE_HEAD_KEY = "<unique pose key>"
    _SEG_HEAD_KEY  = "<unique seg key>"

    @classmethod
    def is_pose_state_dict(cls, sd) -> bool:
        return cls._POSE_HEAD_KEY in sd

    @classmethod
    def is_seg_state_dict(cls, sd) -> bool:
        return any(k.startswith(cls._SEG_HEAD_KEY) for k in sd)

    @classmethod
    def detect_task_from_state_dict(cls, sd) -> Optional[str]:
        if cls.is_pose_state_dict(sd): return "pose"
        if cls.is_seg_state_dict(sd):  return "segment"
        return None  # falls back to detect via DEFAULT_TASK

    def _init_model(self) -> nn.Module:
        if self.task == "pose":    return Libre<FAMILY>PoseModel(size=self.size, nc=1)
        if self.task == "segment": return Libre<FAMILY>SegModel(size=self.size, nc=self.nb_classes)
        return Libre<FAMILY>Model(size=self.size, nc=self.nb_classes)

    def _postprocess(self, raw, conf_thres, iou_thres, **kwargs):
        if self.task == "pose":    return _postprocess_pose(raw, conf_thres, iou_thres, **kwargs)
        if self.task == "segment": return _postprocess_seg(raw, conf_thres, iou_thres, **kwargs)
        return _postprocess(raw, conf_thres, iou_thres, **kwargs)

6.4 Conversion script — single-task metadata-wrap

"""Convert upstream <FAMILY> weights to LibreYOLO format.

Usage:
    python weights/convert_<family>_weights.py upstream/<file>.pth weights/Libre<FAMILY>s.pt --size s
"""
from __future__ import annotations
import argparse
from pathlib import Path

from _conversion_utils import (
    add_repo_root_to_path,
    extract_state_dict,
    load_checkpoint,
    save_checkpoint,
    wrap_libreyolo_checkpoint,
)


def convert(input_path: str, output_path: str, size: str, nc: int = 80) -> None:
    raw = load_checkpoint(input_path)
    state_dict = extract_state_dict(raw, prefer_ema=True)
    print(f"Extracted {len(state_dict)} parameter entries from {input_path}")

    # OPTIONAL — strip an upstream prefix if upstream wraps:
    # state_dict = strip_state_dict_prefix(state_dict, "model.")

    # OPTIONAL — print missing/unexpected diff after dry-load into your model:
    # add_repo_root_to_path(); from libreyolo import Libre<FAMILY>
    # m = Libre<FAMILY>(size=size); res = m.model.load_state_dict(state_dict, strict=False)
    # print("missing:", res.missing_keys); print("unexpected:", res.unexpected_keys)

    wrapped = wrap_libreyolo_checkpoint(
        state_dict,
        model_family="<family>",
        size=size,
        nc=nc,
        task="detect",
        supported_tasks=("detect",),
        default_task="detect",
    )

    out = Path(output_path)
    tmp = out.with_suffix(out.suffix + ".tmp")
    save_checkpoint(wrapped, tmp)
    tmp.rename(out)  # atomic
    print(f"Wrote {out}")


if __name__ == "__main__":
    p = argparse.ArgumentParser()
    p.add_argument("input")
    p.add_argument("output")
    p.add_argument("--size", required=True, choices=["s", "m", "l"])
    p.add_argument("--nc", type=int, default=80)
    args = p.parse_args()
    convert(args.input, args.output, args.size, args.nc)

6.5 Conversion script — multi-task with --task flag

EC pattern. One CLI invocation per task variant.

"""Convert upstream <FAMILY> weights to LibreYOLO format (multi-task)."""
from __future__ import annotations
import argparse
from pathlib import Path

from _conversion_utils import (
    extract_state_dict, load_checkpoint, save_checkpoint, wrap_libreyolo_checkpoint,
)

_SUPPORTED = ("detect", "pose", "segment")
_DEFAULT = "detect"


def convert(input_path, output_path, size, task, nc):
    raw = load_checkpoint(input_path)
    state_dict = extract_state_dict(raw, prefer_ema=True)

    # Per-task overrides — pose typically forces nc=1, names={0: "person"}.
    names = None
    if task == "pose":
        nc = 1
        names = {0: "person"}

    wrapped = wrap_libreyolo_checkpoint(
        state_dict,
        model_family="<family>",
        size=size,
        nc=nc,
        names=names,
        task=task,
        supported_tasks=_SUPPORTED,
        default_task=_DEFAULT,
    )

    out = Path(output_path)
    tmp = out.with_suffix(out.suffix + ".tmp")
    save_checkpoint(wrapped, tmp)
    tmp.rename(out)
    print(f"Wrote {out} (task={task}, size={size})")


if __name__ == "__main__":
    p = argparse.ArgumentParser()
    p.add_argument("input")
    p.add_argument("output")
    p.add_argument("--size", required=True)
    p.add_argument("--task", required=True, choices=list(_SUPPORTED), default=_DEFAULT)
    p.add_argument("--nc", type=int, default=80)
    args = p.parse_args()
    convert(args.input, args.output, args.size, args.task, args.nc)

6.6 Trainer subclass — YOLO-grid

"""<FAMILY> trainer: BaseTrainer subclass."""

from __future__ import annotations
import torch

from ...training.trainer import BaseTrainer
from ...training.config import <FAMILY>Config
from ...training.scheduler import WarmupCosineScheduler
from .loss import <FAMILY>Loss


class <FAMILY>Trainer(BaseTrainer):
    @staticmethod
    def _config_class():
        return <FAMILY>Config

    @staticmethod
    def get_model_family() -> str:
        return "<family>"

    def get_model_tag(self) -> str:
        return f"<FAMILY>-{self.config.size}"

    def create_transforms(self):
        # Return (val_preprocessor, train_dataset_class). Most YOLO-grid
        # families use the shared MosaicMixupDataset wrapper.
        from ...training.augment import MosaicMixupDataset
        from ...validation.preprocessors import <FAMILY>ValPreprocessor
        return <FAMILY>ValPreprocessor(...), MosaicMixupDataset

    def create_scheduler(self, iters_per_epoch: int):
        return WarmupCosineScheduler(
            base_lr=self.config.lr0,
            warmup_iters=self.config.warmup_epochs * iters_per_epoch,
            total_iters=self.config.epochs * iters_per_epoch,
            min_lr_ratio=self.config.min_lr_ratio,
        )

    def get_loss_components(self, outputs) -> dict:
        return {k: float(v) for k, v in outputs.items() if k.startswith("loss_")}

    def on_forward(self, imgs, targets, polygons=None):
        # Detect-only families ignore polygons.
        outputs = self.wrapper_model.model(imgs)
        loss_dict = <FAMILY>Loss(...)(outputs, targets)
        return loss_dict

6.7 Trainer subclass — DETR with per-group LR

Use the RT-DETR pattern (lr_ratio + _scale_lr) — it's smaller than D-FINE's _train_epoch fork and survives base-trainer changes automatically.

class <FAMILY>Trainer(BaseTrainer):
    # ... (other overrides as in 6.6) ...

    def _scale_lr(self, base_lr: float, param_group: dict) -> float:
        return base_lr * param_group.get("lr_ratio", 1.0)

    def _setup_optimizer(self) -> torch.optim.Optimizer:
        # Group params by regex; store lr_ratio so _scale_lr applies it each step.
        backbone_params, head_params = [], []
        for name, param in self.wrapper_model.model.named_parameters():
            if not param.requires_grad:
                continue
            if name.startswith("backbone."):
                backbone_params.append(param)
            else:
                head_params.append(param)

        base_lr = self.config.lr0
        bb_mult = getattr(self.config, "backbone_lr_mult", 0.1)
        groups = [
            {"params": backbone_params, "lr": base_lr * bb_mult, "lr_ratio": bb_mult},
            {"params": head_params,     "lr": base_lr,           "lr_ratio": 1.0},
        ]
        return torch.optim.AdamW(groups, lr=base_lr, weight_decay=self.config.weight_decay)

    def on_forward(self, imgs, targets, polygons=None):
        # DETR criteria want list[dict] targets, not padded (B, max_labels, 5).
        # Translate here. See models/dfine/trainer.py for the precedent.
        target_dicts = self._padded_to_dict_list(targets)
        outputs = self.wrapper_model.model(imgs, target_dicts)  # DETR forward needs targets in train()
        loss_dict = self.criterion(outputs, target_dicts)
        # DETR aux losses across decoder layers — keep them all in the dict.
        return loss_dict

6.8 Inference parity test

Save under weights/parity_<family>.py or tests/unit/test_<family>_parity.py. Skip CI (gate on env var).

"""Cross-load upstream weights into our port and assert max_abs_diff == 0."""
import os
import torch

from libreyolo import Libre<FAMILY>

UPSTREAM_DIR = os.environ.get("<FAMILY>_OFFICIAL_CKPT_DIR")
if not UPSTREAM_DIR:
    raise SystemExit("Set <FAMILY>_OFFICIAL_CKPT_DIR to upstream checkpoint directory")


def main():
    from upstream_pkg import build_model as build_upstream  # TODO: import upstream

    sizes = ["s", "m", "l"]
    for size in sizes:
        upstream = build_upstream(size=size)
        upstream.load_state_dict(torch.load(f"{UPSTREAM_DIR}/<family>_{size}.pth"))
        upstream.eval()

        ours = Libre<FAMILY>(size=size).model
        # If your strict-loading needs adjustment, do it here:
        ours.load_state_dict(upstream.state_dict(), strict=True)
        ours.eval()

        x = torch.randn(1, 3, 640, 640)
        torch.manual_seed(0)
        with torch.no_grad():
            up_out = upstream(x)
            our_out = ours(x)

        # Detect: tensor or per-scale list. DETR: dict.
        # Compare every leaf tensor.
        for k in (up_out if isinstance(up_out, dict) else range(len(up_out))):
            diff = (up_out[k] - our_out[k]).abs().max().item()
            assert diff == 0.0, f"size={size} key={k} max_abs_diff={diff}"
        print(f"size={size}: OK")


if __name__ == "__main__":
    main()

6.9 Cross-load rejection test (siblings)

# tests/unit/test_<family>_can_load.py
"""Bidirectional can_load rejection. Critical for sibling families."""
import torch

from libreyolo.models.<family>.model import Libre<FAMILY>
from libreyolo.models.<sibling>.model import Libre<SIBLING>


def test_<family>_rejects_<sibling>_state_dict():
    sibling_sd = torch.load("tests/fixtures/Libre<SIBLING>s.pt")["model"]
    assert Libre<FAMILY>.can_load(sibling_sd) is False


def test_<sibling>_rejects_<family>_state_dict():
    our_sd = torch.load("tests/fixtures/Libre<FAMILY>s.pt")["model"]
    assert Libre<SIBLING>.can_load(our_sd) is False

6.10 Runtime auto-convert recognizer (only if upstream keys differ)

Skip this when upstream keys already match your native port — the BaseModel default hook claims can_load-matching layouts for free. RTMDet is the smallest precedent (models/rtmdet/convert.py).

# models/<family>/convert.py
"""Upstream <FAMILY> checkpoint key remapping.

Shared by the runtime auto-converter and ``weights/convert_<family>_weights.py``.
"""
from __future__ import annotations
from typing import Dict
import torch

DROP_PREFIXES = ("data_preprocessor.",)  # TODO: buffers your port doesn't carry


def is_upstream_state_dict(state_dict: dict) -> bool:
    """True ONLY for your upstream's naming — be strict, not greedy."""
    # TODO: require a conjunction of tokens unique to the upstream layout.
    return any(k.startswith("bbox_head.") and "rtm_cls" in k for k in state_dict)


def convert_upstream(state_dict: dict) -> Dict[str, torch.Tensor]:
    """Remap upstream keys to the native port's naming."""
    out: Dict[str, torch.Tensor] = {}
    for k, v in state_dict.items():
        if any(k.startswith(p) for p in DROP_PREFIXES):
            continue
        # TODO: syntactic renames only — numerics must stay identical.
        out[k.replace("bbox_head.", "head.", 1)] = v
    return out

# in models/<family>/model.py
    @classmethod
    def convert_upstream_state_dict(cls, weights_dict: dict) -> Optional[dict]:
        """Remap upstream naming to Libre<FAMILY>'s, or None if unrecognized."""
        from .convert import convert_upstream, is_upstream_state_dict

        if not is_upstream_state_dict(weights_dict):
            return None
        return convert_upstream(weights_dict)

Pair it with rejection tests mirroring §6.9: feed every sibling family's upstream checkpoints through is_upstream_state_dict and assert False.

---

Reference

The rest of this document is reference material. You don't need to read it end-to-end. Look up what you need.

7. Architectural patterns

7.1 YOLO-grid (NMS-using)

• Output: per-scale tensor list. Shape varies. YOLOX is (B, 5+nc, H, W) per scale (4 reg + 1 obj + nc cls). YOLOv9 / YOLO-NAS / PicoDet drop objectness, emit (B, 4+nc, N). Confirm against upstream head.
• Targets: (B, max_labels, 5) padded with [class, ...4 box coords]. Box convention is family-specific — YOLOX uses cx, cy, w, h pixel; YOLOv9 uses xyxy normalized to [0, 1] (models/yolo9/loss.py:89, 305). Cross-check the loss's first computation against the dataset's emitted target.
• Loss: per-anchor + assignment (SimOTA / TaskAlignedAssigner / DFL).
• Augmentation: numpy/cv2, mosaic + mixup. Lives in libreyolo/training/augment.py.
• ONNX: 1 output "output", opset 13 default works.
• self.export: bool flag flipped by exporter at libreyolo/export/exporter.py:_model_context.
• NCNN / OpenVINO / TensorRT / TorchScript all work out of the box.

7.2 NMS-free YOLO-grid hybrid (yolo9_e2e)

• One-to-one head + top-K inference (no NMS).
• Backbone, neck, training infra inherited from YOLOv9.
• Postprocess accepts iou_thres for API compat but dels it (libreyolo/postprocess/yolo9_e2e.py::postprocess).
• Must add the family to BaseBackend._is_nms_free_family() (libreyolo/backends/base.py:211) or exported backends will wrongly apply NMS. Current set: {dfine, deim, deimv2, ec, rfdetr, rtdetr, rtdetrv2, rtdetrv4, yolo9_e2e}.

7.3 DETR

• Output: dict {"pred_logits": (B, Q, nc), "pred_boxes": (B, Q, 4)} cxcywh in [0, 1]. Multi-task variants add "pred_masks" / "pred_keypoints".
• Targets: list[dict{labels, boxes_cxcywh_normalized}] — no padding.
• Loss: Hungarian matching + auxiliary outputs across decoder layers.
• Augmentation: torchvision v2 with tv_tensors.Image + tv_tensors.BoundingBoxes. Family-local transforms.py.
• Multi-scale: per-batch random resize via custom collate.
• Backbone LR multiplier (0.1× / 0.5×) is standard. Per-group LR in _setup_optimizer + _scale_lr override (RT-DETR pattern, preferred) or _train_epoch override (D-FINE pattern, older).
• Gradient clipping (max_norm=0.1).
• ONNX: 2 outputs ["pred_logits", "pred_boxes"], opset ≥ 16.
• Export wrapper: small nn.Module calling model.deploy() (recursive convert_to_deploy) and flattening dict→tuple. D-FINE has the precedent at models/dfine/nn.py:203-209.
• NCNN doesn't work — topk not in NCNN's op registry. Block early.

7.4 Sibling / parent-child families

Three flavors, different disambiguation rules:

Parent-child (one extends another). YOLOv9-E2E extends YOLOv9. Rules:

• Re-declare FAMILY and FILENAME_PREFIX always.
• Inherit ClassVars that don't change.
• Override can_load to fingerprint a key the child has and the parent doesn't.
• Bidirectional rejection test (template §6.9) is mandatory.
• Trainer minimal-override: subclass parent trainer, override only _config_class.
• Reuse parent's converter by passing a different model_family at wrap.

Same architecture, different loss/matcher (D-FINE / DEIM). can_load collisions are inevitable. Disambiguate via:

1. Embed model_family in the converted checkpoint (the converter's job).
2. Use FILENAME_PREFIX in detect_size_from_filename as fallback.
3. Order registry imports in libreyolo/models/__init__.py so families with most distinctive markers load first; generic fallbacks last. Current order at libreyolo/models/__init__.py:47-75: EC → YOLOX → YOLOv9-E2E → YOLOv9 → YOLO-NAS → DEIMv2 → RT-DETRv4 → D-FINE → DEIM → PicoDet → RT-DETR → RT-DETRv2 → RTMDet → L2CS → FOMO → DepthAnythingV2 → NAFNet → EoMT → PIDNet → MobileNetV4 → ConvNeXt → EfficientNetV2 → ResNet → CLIP (RF-DETR and DINOv2 lazy-register behind the transformers dep). The same order arbitrates runtime auto-conversion claims (§ commit 3b).
4. Raise an explicit "ambiguous between {A, B}" error on a true tie. The D-FINE/DEIM-tie precedent lives in libreyolo/models/__init__.py (search "Ambiguous D-FINE/DEIM").

8. BaseModel contract

8.1 ClassVars (declare these explicitly, even when single-task)

__init_subclass__ auto-registers on import (models/base/model.py:62-69). Import order in libreyolo/models/__init__.py = can_load priority.

8.2 The 4 ABCs

Family configs are split: most append to libreyolo/training/config.py (YOLOX, YOLO9, D-FINE, DEIM, DEIMv2, EC, YOLO-NAS, PicoDet); RF-DETR / RT-DETR / YOLOv9-E2E use family-local models/<family>/config.py. Either is fine.

8.3 Task resolution flow

When LibreYOLO("LibreECs-pose.pt", task=None) is called:

1. Family resolution (libreyolo/models/__init__.py:288-331):
   • state_dict["model_family"] → match registered family.
   • Else iterate _registry; first class where detect_size_from_filename(name) is not None AND can_load(weights_dict) is True wins.
   • Else fall back to can_load only.
2. Task resolution (libreyolo/models/__init__.py:373-399 → libreyolo/tasks.py:109-127):
   • Precedence: explicit task= → checkpoint["task"] → state-dict sniff → filename suffix → DEFAULT_TASK.
   • State-dict sniff: family-specific is_pose_state_dict / is_seg_state_dict.
   • resolve_task() validates the resolved task is in SUPPORTED_TASKS.

_filename_regex (models/base/model.py) compiles {FILENAME_PREFIX}(?P<size>...)(?P<task>-<suffix>)?{WEIGHT_EXT}. Task suffix mapping at libreyolo/tasks.py:70-80:

TASK_TO_SUFFIX = {
    "segment": "seg", "semantic": "sem", "pose": "pose", "classify": "cls",
    "gaze": "gaze", "obb": "obb", "point": "point", "depth": "depth",
    "restore": "restore",
}

Detect has no suffix.

If nothing recognizes the file as a LibreYOLO checkpoint, the factory attempts runtime auto-conversion (autoconvert_upstream_checkpoint, called from libreyolo/models/__init__.py) before failing — see commit 3b.

9. Multi-task internals

Results slots default None:

Results(boxes=Boxes(...), masks=None, keypoints=None, probs=None, obb=None)

Results._select(idx) slices every non-None slot in lockstep. Drawing dispatches on slot presence (if result.keypoints is not None: draw_keypoints(...)), not on self.task. draw_keypoints uses COCO_KEYPOINT_EDGES skeleton at libreyolo/utils/drawing.py:222-286.

InferenceRunner._apply_classes_filter takes both masks_t and keypoints_t (models/base/inference.py:241-258). When you implement _postprocess, return only the present slots; _wrap_results threads them into Results.

Postprocess return contract:

If your model emits keypoints as (N, K, 2), append a column of ones for visibility (libreyolo/postprocess/ec.py::postprocess_pose).

10. Validation

BaseModel.val() dispatches by self.task (see the validator-selection block inside models/base/model.py::val; a family can also pin validator_class directly):

detect   -> DetectionValidator     (COCOeval iouType="bbox")
segment  -> SegmentationValidator  (iouType="bbox" + "segm", returns (B) and (M))
pose     -> PoseValidator          (iouType="keypoints")
classify -> ClassifyValidator      obb   -> OBBValidator
semantic -> SemanticValidator      point -> PointValidator
depth    -> DepthValidator         restore -> RestoreValidator

ValidationConfig (libreyolo/validation/config.py) exposes one of three input modes — exactly one must be set:

• data (YAML) — detect/segment with COCO128-style YAML.
• data_dir — detect/segment with a flat dataset.
• keypoints_json + images_dir — pose only.

This skill's templates are detect/pose/segment-shaped. For the other tasks, clone the merged exemplar family instead of adapting a template:

• classify → models/{mobilenetv4,convnext,efficientnetv2,resnet}/ — reuse the shared BaseTrainer classify path (_setup_classify_data / _run_classify_validation), return {"probs": ...} from _postprocess, set best_metric_key = "metrics/accuracy_top1".
• zero-shot classify → models/clip/ (set_classes, clip_validator.py).
• semantic → models/pidnet/ (CNN) or models/eomt/ (ViT).
• depth → models/depth_anything/.
• restore → models/nafnet/ (native-resolution, no letterbox).
• point → models/fomo/.
• gaze → models/l2cs/ (inference-only precedent).

Per-task best_metric_key override is mandatory for non-detect trainers. BaseTrainer.best_metric_key = "metrics/mAP50-95" (training/trainer.py:35) selects on bbox. Pose / seg trainers tracking the wrong metric is the silent issue this skill calls out as landmine #23.

11. Training, per task

11.1 Detection — production

Every existing family ships a detection trainer that subclasses BaseTrainer (or DFINETrainer for DETR). E2E covered by tests/e2e/test_val_coco128.py

• tests/e2e/test_rf1_training.py.

11.2 Segmentation — plumbing exists, native uptake limited

• BaseTrainer.on_forward(imgs, targets, polygons=None) accepts polygons (training/trainer.py:130-149).
• BaseTrainer batch-unpacks 5-tuples when segments are loaded (trainer.py:451-455).
• YOLODataset(..., load_segments=True) and COCODataset(..., load_segments=True) add a polygons stream.
• Polygons contract: batch -> image -> instance -> ring, each ring Nx2 float32 in pixel-space original-image coords (data/dataset.py:28-49).
• SegmentationValidator runs two pycocotools evaluators (bbox + segm), reports metrics/mAP50-95(B) and metrics/mAP50-95(M).
• RF-DETR is the only family with a multi-task trainer today, via compile-time RFDETR_TRAINERS vs RFDETR_SEG_TRAINERS selection (models/rfdetr/trainer.py:27-106).

To wire native segmentation training: proto/mask head + mask coefficients + polygon→mask rasterization + mask loss in family-local nn.py / loss.py. Override best_metric_key = "metrics/mAP50-95(M)". Disable mosaic/mixup or use a family-local dataset wrapper (D-FINE pattern).

11.3 Pose — inference-only territory

• PoseValidator exists (validation/pose_validator.py:35-270), uses COCOeval(iouType="keypoints").
• Results.keypoints plumbing is complete; _select preserves alignment.
• Data pipeline does not yet have load_keypoints=. No YOLODataset flag, no batch shape for keypoints. A pose trainer would have to add this first.
• Mosaic/mixup don't transform keypoints either.

If you ship a pose trainer, you're greenfield: land the load_keypoints= data path first, then keypoint loss, then the trainer. Document scope in your commit message.

12. Inference parity proof

Before writing a trainer, prove the model loads upstream weights and produces identical outputs.

1. Import the upstream model class side-by-side with yours.
2. Build both with the same config / size; cross-load the upstream state_dict into yours and inspect the missing/unexpected key diff. Use strict=True only if your port loads the full upstream state dict. For ports that intentionally drop layers, use strict=False and assert the missing/unexpected set matches a documented expected set.
3. Run identical inputs through both at FP32 and assert max_abs_diff == 0 on output tensors that come from layers present in both, in eval() mode.
4. Save the script as a one-off (template §6.8).

_strict_loading=False is the right default whenever upstream checkpoints carry buffers the port doesn't materialize identically (regenerated anchors, valid_mask). EC and D-FINE both return False unconditionally for this reason — don't conflate it with task-aware logic.

For sibling-tier integrations that load upstream implementations from an installed dependency (LibreSAM via transformers): substitute "import the upstream package and verify it produces the documented outputs." Run the parity check against the dependency versions users will actually install — see landmine #37 for what silent dep drift does.

For multi-task families: do the parity check for each task variant.

13. Files-touched matrix

Always edited:

Conditional:

14. Integration-proof tests

DETR families: skip the last_loss < first_loss assertion (loss too noisy).

Per-family unit smoke tests

The floor below the e2e gate. For each family:

• filename detection
• can_load discriminator (bidirectional for siblings — template §6.9)
• forward shape
• loss parity
• export smoke
• trainer smoke
• multi-task: one file per task

Optional faithfulness gate

test_val_coco128's mAP50-95 ≥ 0.18 floor is a sanity check, not a faithfulness check. For published-number matching: tests/nightly/test_<family>_official_ckpt_map.py, gated on <FAMILY>_OFFICIAL_CKPT_DIR env var, opt-in.

15. Silent-corruption landmines

In priority order. Each line: [which family hit it] — what to do.

1. Color space mismatch between training transform and val preprocessor (YOLOX BGR vs YOLOv9 RGB) — pin the convention in both docstrings.
2. Target format mismatch (D-FINE) — DETR criteria want list[dict] but pipeline yields padded (B, max_labels, 5); translate in on_forward.
3. can_load() too greedy (RF-DETR almost stole D-FINE checkpoints) — match on tokens unique to your architecture; never "backbone" or "weight".
4. Backbone LR multiplier missing (DETR families) — silent ~0.5 mAP loss in fine-tuning. Implies per-group LR + _scale_lr override (RT-DETR pattern).
5. Multi-scale collate epoch propagation (D-FINE) — collate needs set_epoch() from trainer at each epoch start.
6. Stop-epoch augmentation policy (D-FINE) — disable RandomZoomOut/RandomIoUCrop at epoch N.
7. labels_getter=lambda is unpicklable under Python 3.14 forkserver (D-FINE on macOS) — module-level function for SanitizeBoundingBoxes.
8. RandomIoUCrop has no p parameter in tv2 — wrap with RandomApply.
9. MPS-specific torch bugs in DETR backward (D-FINE) — _setup_device override falls back to CPU; CUDA stays unchanged.
10. Post-train device drift (D-FINE) — end train() with self.model.to(self.device).
11. ONNX opset 13 default — DETR with deformable attention needs ≥ 16. Set per-family default in BaseExporter.__call__.
12. NCNN can't handle DETR ops (D-FINE) — block the export early.
13. head.export flag missing (YOLO-grid) — without it, ONNX bakes static shapes.
14. strict=True state-dict loading — override _strict_loading() = False if upstream carries EMA buffers, profiling state, or aux heads.
15. Cross-family rejection on cross-family transfer — pop model_family from the donor checkpoint dict.
16. EMA decay too low — early-epoch evals show flat mAP because EMA hasn't settled.
17. Letterbox vs. plain resize — uses_letterbox property must match the training transform.
18. _train_epoch override drift (DETR) — leave a "kept in sync as of " comment if you fork the loop. Better: use _scale_lr (RT-DETR pattern) and don't fork.
19. Per-size defaults copy-pasted from one size to all. Build a side-by-side table of every override per size before assuming s applies to n.
20. min_lr_ratio cross-family disagreement. D-FINE: 0.05, DEIM: 0.5 (default), DEIMv2: 0.5 default but n overrides to 1.0 (training/config.py:304), EC: 0.5, YOLOv9: 0.01. Cross-check upstream YAML.
21. Wasted ImageNet backbone download in _init_model. When the user constructs from a LibreYOLO checkpoint, your weights overwrite anything _init_model initialised. Pattern (prototyped in bf16a2b but not currently in models/rtdetr/model.py): BaseModel.__init__ peeks at the checkpoint via cls.detect_size; inside _init_model, check self._loading_pretrained_checkpoint and pass backbone_pretrained=False when set. New backbone-heavy ports should implement this from day one.
22. is_pose_state_dict / is_seg_state_dict collision with sibling families. Cross-test against every sibling family's known state dict in unit tests.
23. best_metric_key left at default for non-detect trainers. BaseTrainer.best_metric_key = "metrics/mAP50-95" selects on bbox. Pose / seg trainers must override.
24. Mosaic/mixup applied to polygons/keypoints they don't transform. Disable, or family-local dataset wrapper (D-FINE).
25. TASK_INPUT_SIZES keys disagree with INPUT_SIZES keys. Per-task entries needed for asymmetric task sizes (YOLO-NAS pose-n).
26. Postprocess returning keypoints without a visibility column. Append a column of ones (libreyolo/postprocess/ec.py::postprocess_pose).
27. Multi-character size codes need length-descending regex sort. RT-DETR's r50 vs r50m — override detect_size_from_filename and sort sizes by length descending. Precedent: models/rtdetr/model.py:253-270.
28. Pose head with single-class user-facing override. DETR pose may pair multi-class internal head with single user-facing class. Converter must override nc=1, names={0:"person"} for the pose variant. EC: weights/convert_ec_weights.py:53-55.
29. DETR pose may skip the box head entirely. Some pose decoders derive boxes from keypoint extents rather than a dedicated regression head (EC: libreyolo/postprocess/ec.py::postprocess_pose). Document if yours does.
30. NMS-free family not added to backend allowlist. BaseBackend._is_nms_free_family() (backends/base.py:211) hardcodes {"dfine","deim","deimv2","ec","rfdetr","rtdetr","rtdetrv2","rtdetrv4","yolo9_e2e"}. Backends apply NMS post-export to anything not in this set. YOLOv9-E2E shipped with this miss and exported graphs were silently wrong until it was fixed — don't be the next entry.
31. Sibling-family can_load collisions are bidirectional. Both directions need tests (template §6.9).
32. Sniff-key prefix-match collisions across multi-task families. RF-DETR's "segmentation_head" is a shorter prefix than EC's "decoder.decoder.segmentation_head" — if a new family puts segmentation_head.* at a path not nested under decoder.decoder., RF-DETR steals it.
33. warmup_iters from upstream is recipe-tuned to a full epoch budget. When user overrides epochs, fall back to warmup_epochs. DEIMv2: models/deimv2/trainer.py:79-90 (commit 22f90518).
34. Asymmetric per-task sizes need explicit error messages. YOLO-NAS pose ships n but YOLO-NAS detect doesn't. The factory's TASK_INPUT_SIZES[task] validation needs a per-task error path.
35. Single-task families currently inherit SUPPORTED_TASKS defaults silently. §8.1's recommendation is aspirational today — D-FINE, DEIM, DEIMv2, PicoDet, RT-DETR, YOLOX, YOLOv9, YOLOv9-E2E all rely on BaseModel's ("detect",) default. New ports should declare explicitly.
36. Greedy convert_upstream_state_dict steals other families' upstream checkpoints. Runtime auto-conversion asks every registered family; subclass-beats-base then registry order arbitrates. Recognize on a strict conjunction of upstream-unique tokens, return None for everything else, and add rejection tests against sibling families' upstream checkpoints (mirror §6.9).
37. Dependency loads that silently no-op leave you training a random backbone. (RF-DETR) WindowedDinov2WithRegistersBackbone.from_pretrained(...) silently skipped the weight transfer under recent Transformers releases — classify/seg/depth variants trained on a randomly-initialized ViT and nobody crashed. Fix: explicit reference-model state-dict transfer (models/rfdetr/backbone.py, see the NOTE around line 534). Your commit-4 parity gate is the detector — run it against the dependency versions users actually install, not only your lockfile.
38. Third-party API drift breaks training while inference stays green. (all four timm-derived classify families, #496) a timm kwarg rename crashed model.train() for every shipped classify family while predict/val stayed perfect, so it shipped broken. Inference parity proves nothing about the training path — run a real model.train() smoke per family against current deps before release.
39. Augmentation knobs that silently don't fire. (YOLO9, #432/#484) degrees/translate/shear/scale were dead no-ops in the mosaic path, and mixup dropped the second image's labels — silently, for a long time. When you wire a training pipeline, unit-test that each documented aug knob measurably changes pixels and labels.

16. Cross-references

• Filename whitelist + HF 5-file contract: skills/libreyolo-upload-hf-model/SKILL.md. Authoritative list of all valid weight names.
• Nomenclature: docs/nomenclature.md. Family table, casing, size codes, task suffixes.
• Runtime auto-conversion: libreyolo/models/autoconvert.py (its module docstring is the spec); hook default models/base/model.py::convert_upstream_state_dict; per-family remaps models/{yolo9,rtmdet,rtdetr,rtdetrv4,picodet}/convert.py; skiplist _SKIP_FAMILIES; RF-DETR's bespoke recognizer lives in autoconvert.py itself.
• Sibling tiers (prompt-driven): models/openvocab/ (Grounding DINO, OWLv2, OMDet-Turbo; towers in models/bert/ + models/swin/), models/sam/ + models/mobilesam/ (LibreSAM), models/vlm/ (LibreVLM).
• Conversion tier examples:
  • metadata-wrap (single-task): weights/convert_dfine_weights.py, weights/convert_deim_weights.py
  • metadata-wrap (multi-task --task flag): weights/convert_ec_weights.py
  • light structural (key remap + EMA drop): weights/convert_rtdetr_hgnetv2_weights.py, weights/convert_picodet_weights.py
  • heavy structural (numbered upstream → semantic): weights/convert_yolo9_weights.py
  • safetensors handling: weights/convert_deimv2_weights.py:19-43
• Pattern references:
  • NMS-free YOLO-grid hybrid: libreyolo/models/yolo9_e2e/
  • Parent-child sibling: YOLOv9-E2E inheriting from YOLOv9
  • Same-architecture siblings (D-FINE / DEIM): tie-break in libreyolo/models/__init__.py (search "Ambiguous D-FINE/DEIM")
  • Lazy import for optional dep: _ensure_rfdetr in libreyolo/models/__init__.py (registers RF-DETR + DINOv2 behind the transformers dep)
  • Gated experimental training: libreyolo/models/picodet/model.py:157-201, libreyolo/models/ec/model.py:269-310
  • Vendored sub-component license: libreyolo/models/deimv2/engine/backbone/dinov3/
  • Per-group LR via lr_ratio + _scale_lr (preferred): libreyolo/models/rtdetr/trainer.py:185-260
  • Per-group LR via _train_epoch fork (older): libreyolo/models/dfine/trainer.py:148-212
• Legacy detection-only skill (do not follow for new ports): skills/legacy_libreyolo-add-native-detection-model/SKILL.md.