detr_resnet101(pretrained: bool | str = False, weights: DETRResNet101Weights | None = None, overrides: object = {})DETR with ResNet-101 backbone (Carion et al., ECCV 2020).
Builds a DETRForObjectDetection with the same transformer
head as detr_resnet50 but a deeper ResNet-101 backbone
([3, 4, 23, 3] bottleneck blocks). The COCO config uses
num_classes = 91. Approximately 60.5M parameters and COCO box
mAP of 43.5 (paper Table 1, DETR-R101 row).
Model Size
Parameters
pretrainedbool or str= FalseFalse → random init; True
→ the DEFAULT tag (DETRResNet101Weights.COCO_2017); a
tag string → that specific checkpoint. Mutually exclusive with
weights (which wins if both are given).pretrained.**overridesobject= {}DETRConfig.Returns
DETRForObjectDetectionDetector with the DETR ResNet-101 configuration applied (or with
overrides merged on top of it).
Notes
See Carion et al., "End-to-End Object Detection with Transformers",
ECCV 2020 (arXiv:2005.12872). Switching to ResNet-101 buys ~1.5
points of AP at the cost of ~50% more parameters in the backbone;
the transformer head is unchanged. Pretrained weights are converted
from the original Facebook DETR detr_resnet101 checkpoint and
hosted under lucid-dl/detr-resnet-101.
Examples
>>> import lucid
>>> from lucid.models.vision.detr import detr_resnet101
>>> model = detr_resnet101()
>>> x = lucid.randn(1, 3, 800, 800)
>>> out = model(x)
>>> out.logits.shape
(1, 100, 92)