detr_r101¶

lucid.models.detr_r101(num_classes: int = 91, num_queries: int = 100, pretrained_backbone: bool = False, **kwargs) → DETR¶

The detr_r101 function returns an instance of the DETR model configured with a ResNet-101 backbone.

Total Parameters: 60,570,528 (MS-COCO)

Function Signature¶

@register_model
def detr_r101(
    num_classes: int = 91,
    num_queries: int = 100,
    pretrained_backbone: bool = False,
    **kwargs,
) -> DETR

Parameters¶

num_classes (int, default=91): Number of foreground categories. For COCO this is 91; the model internally adds one “no-object” class.
num_queries (int, default=100): Number of learned object queries (maximum detections per image). This sets the second dimension of the outputs.
pretrained_backbone (bool, default=False): If True, initializes the ResNet-101 backbone with pretrained weights and (by default) fine-tunes it.
kwargs: Additional keyword arguments forwarded to the underlying DETR constructor (e.g., aux_loss, loss coefficients, custom matcher, etc.).

Returns¶

DETR:

A DETR-R101 detector with class and box heads. In eval mode with aux_loss=False, a forward pass returns:
- pred_logits: (B, num_queries, num_classes + 1) - raw class scores; the last logit is no-object.
- pred_boxes: (B, num_queries, 4) - normalized boxes in (cx, cy, w, h) with values in [0, 1].

Example Usage¶

>>> from lucid.models import detr_r101
>>> import lucid
>>>
>>> # COCO-style config (91 classes), 100 queries
>>> model = detr_r101(num_classes=91, num_queries=100, pretrained_backbone=True, aux_loss=False)
>>> print(model)
>>>
>>> # Dummy input
>>> x = lucid.random.randn(1, 3, 800, 800)
>>> logits, boxes = model(x)  # aux_loss=False returns a tuple
>>> print(logits.shape, boxes.shape)
(1, 100, 92) (1, 100, 4)
>>>
>>> # Convert logits to probabilities (drop no-object column)
>>> probs = lucid.softmax(logits, axis=-1)[..., :-1]
>>> scores = probs.max(axis=-1)  # (B, num_queries)