module

weights

9 members

lucid.weights

`lucid.weights` — pretrained-weight system for the model zoo.

A standalone package (sibling to lucid.models) that provides the infrastructure for pretrained checkpoints: the tagged-variant enum base, hub download + SHA verification, loading, and a discovery registry. Preprocessing lives in lucid.utils.transforms; a checkpoint's WeightEntry carries one of those transforms.

It deliberately contains no architecture-specific weight declarations. Each model family declares its own checkpoints in a _weights.py module inside the model package (e.g. lucid.models.vision.resnet._weights.ResNet18Weights), importing the primitives below. This keeps lucid.weights a pure porting substrate and the dependency one-directional (models → weights, never the reverse).

Notes

Each architecture declares a WeightsEnum (e.g. ResNet18Weights) whose members are WeightEntry manifests — one per tagged checkpoint (IMAGENET1K_V1, …) plus a DEFAULT alias. This mirrors the torchvision Weights enum so call sites are familiar:

import lucid.models as models
import lucid.weights as weights

# default tag
model = models.resnet_18(pretrained=True)

# explicit tag — enum or string
from lucid.models.vision.resnet import ResNet18Weights
model = models.resnet_18_cls(weights=ResNet18Weights.IMAGENET1K_V1)
model = models.resnet_18_cls(pretrained="IMAGENET1K_V1")

# discover + preprocess (after the model package is imported)
weights.list_pretrained("resnet_18")        # ['IMAGENET1K_V1']
w = weights.get_weight("ResNet18Weights.IMAGENET1K_V1")
x = w.transforms()(image)

The package depends only on lucid.serialization (for SafeTensors loading); model factories depend on it, never the reverse, so there is no circular import. Crucially, WeightEntry carries no model config — named factories already pin their own.

Checkpoints are hosted on the Hugging Face Hub under the lucid-dl org, converted from torchvision / timm / transformers sources by the offline tools/convert_weights pipeline.

Classes

dataWeightEntry1 methods

Immutable manifest describing one pretrained checkpoint.

classWeightsEnum7 methods

Base class for per-architecture pretrained-weight enums.

Functions

fndownload→ pathlib.Path

Download url to the local cache, verify SHA-256, return the path.

fnload_weight_entry→ object

Download the checkpoint described by weights and load it.

fnresolve_weights→ WeightsEnum or None

Resolve a factory's pretrained / weights args to a member.

fnget_weight→ WeightsEnum

Resolve a dotted "EnumName.TAG" string to a weights member.

fnlist_pretrained→ list of str

List the available weight tags for a model.

fnregister_weights→ Callable[[type[WeightsEnum]], type[WeightsEnum]]

Class decorator registering a WeightsEnum for discovery.

fnweights_for→ type[WeightsEnum] or None

Return the WeightsEnum registered for model_name.