register_model

→Callable[[ModelFactory], ModelFactory]

register_model(task: str = 'base', family: str | None = None, model_type: str | None = None, model_class: type[PretrainedModel] | None = None, default_config: ModelConfig | None = None, params: int | None = None, summary: object = 'auto')

source edit

Decorator that registers a factory function under its __name__.

The global registry is the single source of truth for which models Lucid exposes. Every public factory in lucid/models/**/_pretrained.py is decorated with @register_model so it becomes discoverable via create_model, list_models, and the AutoModelFor* family.

Parameters

taskstr= "base"

Task tag used by Auto classes for typed dispatch. One of "base" (a backbone with no head) or the eight task tags declared in lucid.models._tasks: "image-classification", "image-generation", "object-detection", "semantic-segmentation", "sequence-classification", "token-classification", "language-modeling", "world-modeling". The tag a factory passes must equal its model_class.task, which the suite asserts — the two used to be independent strings and drifted.

familystr or None= None

Architecture family identifier (e.g. "resnet"). When None, inferred from the factory's parent module name (the second-to-last component of fn.__module__).

model_typestr or None= None

Persistent ModelConfig.model_type value this factory produces. Defaults to family. Used by directory-based loading to match on-disk config.json files against registered factories.

model_classtype[PretrainedModel] or None= None

The concrete PretrainedModel subclass returned by this factory. When supplied, directory-based loading (AutoModel.from_pretrained against a path) avoids a redundant factory call — strongly recommended.

default_configModelConfig or None= None

The default config produced when pretrained=False. When supplied, AutoConfig.from_pretrained returns it without instantiating the model.

paramsint or None= None

Paper-cited trainable-parameter count for this factory's default config (e.g. 61_100_840 for alexnet_cls). Surfaced on the API docs site as a "61.1M"-style tag on the factory card and as a "Model Size" section on the detail page. Omit (None) when no authoritative count is available — the UI then hides the tag rather than guessing.

summaryobject= "auto"

Layer-tree summary used by the docs site's collapsible "Model Size" section. "auto" (the canonical value) tells tools/build_model_summaries.py to introspect the model via shadow_alloc and emit a real tree. Pass an already-built summary dict to override the introspection (rare — used only when shadow-alloc introspection fails for a model family).

Returns

Callable[[ModelFactory], ModelFactory]

Decorator that returns the original factory unmodified (only the registry side-effect occurs).

Raises

ValueError

If a model is already registered under the same normalised name.

Notes

The decorator uses the factory's __name__ (normalised through _normalize) as the registry key. Same name twice raises — this catches accidental shadowing across families.

Examples

>>> from lucid.models import register_model, PretrainedModel
>>> # Inside lucid/models/vision/myfamily/_pretrained.py:
>>> @register_model(
...     task="image-classification",
...     family="myfamily",
...     model_type="myfamily",
... )
... def myfamily_small(pretrained: bool = False, **overrides):
...     cfg = MyFamilyConfig(depth=12, **overrides)
...     return MyFamilyForImageClassification(cfg)

Used by 76

… 64 more