PreTrainedModelMixin¶

PreTrainedModelMixin is a small utility mixin for model classes that inherit from lucid.nn.Module. It adds a chainable from_pretrained(…) method that resolves a weight entry and applies it to the current model instance.

Warning

Not all pre-built model base classes under lucid.models currently inherit PreTrainedModelMixin. As a result, from_pretrained(…) is not guaranteed to exist on every model class.

Important

The model configuration used to instantiate the class must match the configuration expected by the pretrained weights you load with from_pretrained(…) (for example, hidden size, number of layers/heads, vocabulary size, and other structural settings). If these do not match, state dict loading may fail (or require strict=False with partial loading).

class lucid.models.base.PreTrainedModelMixin¶

Class Signature¶

class PreTrainedModelMixin:
    def from_pretrained(self, weights: WeightEntry, strict: bool = True) -> Self

Parameters¶

weights (WeightEntry): A registered weight entry (for example BERT_Weights.PRE_TRAIN_BASE).
strict (bool, default True): Passed through to model.load_state_dict(…, strict=strict). Set False when loading a backbone into a different task head.

Returns¶

self: The same model instance, so calls can be chained.

Behavior¶

Works only on classes that are also lucid.nn.Module.
Raises RuntimeError if called on a non-module object.
Internally calls lucid.weights.apply(…).

Method¶

PreTrainedModelMixin.from_pretrained(weights: WeightEntry, strict: bool = True) → Self¶

Examples¶

Basic pretrained load¶

from lucid.models import SomeModel
from lucid.weights import SomeModel_Weights

model = SomeModel()
model = model.from_pretrained(SomeModel_Weights.DEFAULT)
model.eval()

Backbone transfer (strict=False)¶

from lucid.models import TargetTaskModel
from lucid.weights import SourceBackbone_Weights

model = TargetTaskModel()
model.from_pretrained(
    SourceBackbone_Weights.DEFAULT,
    strict=False,  # allow task-head key mismatch
)

Chaining usage¶

from lucid.models import SomeModel
from lucid.weights import SomeModel_Weights

model = (
    SomeModel()
    .from_pretrained(SomeModel_Weights.DEFAULT)
    .to("gpu")
    .eval()
)

Custom model using the mixin¶

import lucid.nn as nn
from lucid.models.base import PreTrainedModelMixin

class MyModel(PreTrainedModelMixin, nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(128, 128)