create_model(name: str, pretrained: bool = False, overrides: object = {})Look up name and call its factory — the timm-style entry point.
Parameters
namestrRegistered model name (case-insensitive,
- / _
interchangeable).pretrained(bool, optional, keyword - only)= FalseWhether to download and load registered checkpoint weights.
**overridesobject= {}Forwarded as keyword arguments to the underlying factory. Most
factories pipe these into the
ModelConfig constructor so
callers can adjust hyper-parameters at creation time without
defining a custom config (e.g. num_classes=10 for transfer
learning).Returns
PretrainedModelA fully constructed concrete model.
Raises
ValueErrorIf
name is not registered.Notes
Equivalent to model_entrypoint(name)(pretrained=pretrained, **overrides). For task-aware dispatch with the Auto* shells, see
AutoModel and the AutoModelFor* family.
Examples
>>> from lucid.models import create_model
>>> # CIFAR-10 transfer head — random init, 10 classes
>>> model = create_model("resnet_50", num_classes=10)
>>> model.config.num_classes
10