DETRConfig¶

class lucid.models.DETRConfig(backbone: lucid.nn.module.Module, transformer: lucid.nn.module.Module, num_classes: int, num_queries: int = 100, aux_loss: bool = True, matcher: lucid.nn.module.Module | None = None, class_loss_coef: float = 1.0, bbox_loss_coef: float = 5.0, giou_loss_coef: float = 2.0, eos_coef: float = 0.1)¶

DETRConfig stores the backbone, transformer, query count, and loss settings used to build lucid.models.DETR.

Class Signature¶

@dataclass
class DETRConfig:
    backbone: nn.Module
    transformer: nn.Module
    num_classes: int
    num_queries: int = 100
    aux_loss: bool = True
    matcher: nn.Module | None = None
    class_loss_coef: float = 1.0
    bbox_loss_coef: float = 5.0
    giou_loss_coef: float = 2.0
    eos_coef: float = 0.1

Parameters¶

backbone (nn.Module): Feature extractor module. It must define a positive integer num_channels attribute used by the 1x1 input projection.
transformer (nn.Module): DETR encoder-decoder module. It must define a positive integer d_model attribute used by positional encoding, query embeddings, and prediction heads.
num_classes (int): Number of foreground classes.
num_queries (int): Number of learned object queries.
aux_loss (bool): Whether intermediate decoder outputs should contribute auxiliary losses.
matcher (nn.Module | None): Optional Hungarian matcher override. If omitted, the standard matcher is built from the configured loss coefficients.
class_loss_coef (float): Classification loss weight.
bbox_loss_coef (float): Box L1 loss weight.
giou_loss_coef (float): Generalized IoU loss weight.
eos_coef (float): Weight for the “no-object” class in cross-entropy.

Validation¶

backbone must be an nn.Module with positive integer num_channels.
transformer must be an nn.Module with positive integer d_model.
num_classes and num_queries must be greater than 0.
aux_loss must be a boolean.
matcher must be an nn.Module or None.
class_loss_coef, bbox_loss_coef, and giou_loss_coef must be non-negative.
eos_coef must be in [0, 1].

Usage¶

import lucid.models as models
from lucid.models.vision.detr import _BackboneBase, _Transformer

config = models.DETRConfig(
    backbone=_BackboneBase(models.resnet_50()),
    transformer=_Transformer(d_model=256, n_head=8),
    num_classes=91,
    num_queries=100,
)
model = models.DETR(config)