faster_rcnn_resnet_101_fpn¶

lucid.models.faster_rcnn_resnet_101_fpn(num_classes: int = 21, backbone_num_classes: int = 1000, **kwargs) → FasterRCNN¶

The faster_rcnn_resnet_101_fpn function constructs a two-stage object detection model based on the Faster R-CNN architecture, using a deeper ResNet-101 backbone and Feature Pyramid Network (FPN). This variant offers improved accuracy over ResNet-50 variants, particularly for more complex or large-scale detection tasks.

Total Parameters: 62,508,030

Function Signature¶

@register_model
def faster_rcnn_resnet_101_fpn(
    num_classes: int = 21,
    backbone_num_classes: int = 1000,
    **kwargs
) -> FasterRCNN

Parameters¶

num_classes (int, optional): The number of object categories to detect. Default is 21 (e.g., PASCAL VOC). Used in the classification head of the detection model.
backbone_num_classes (int, optional): Number of output classes for the ResNet-101 backbone used during initialization. This should match the classification task used for pretraining. Default is 1000.
kwargs (dict, optional): Additional keyword arguments for configuring the FasterRCNN model, such as anchor settings, thresholds, etc.

Returns¶

FasterRCNN: A Faster R-CNN model with a ResNet-101 backbone and FPN.

Examples¶

Basic Usage

from lucid.models import faster_rcnn_resnet_101_fpn

# Construct model for 21 object classes
model = faster_rcnn_resnet_101_fpn(num_classes=21)

# Dummy input image
x = lucid.random.randn(1, 3, 224, 224)

# Inference
cls_logits, bbox_deltas = model(x)

print(cls_logits.shape)   # (1, 300, 21)
print(bbox_deltas.shape)  # (1, 300, 4)

Training Notes¶

Lucid does not provide pretrained weights for ResNet-101 detection backbones. You are expected to train the backbone separately and reuse it for Faster R-CNN training.

Pretrain ResNet-101 Backbone

Access and train the internal backbone from the detector:

from lucid.models import faster_rcnn_resnet_101_fpn

model = faster_rcnn_resnet_101_fpn(num_classes=21)
backbone = model.backbone.net  # ResNet-101

You can use this as a standard classifier and train it using a dataset like ImageNet:

# Train backbone here
# Save weights:
state_dict = backbone.state_dict()

Use Pretrained Backbone for Detection

Load the pretrained weights back before training object detection:
```
model.backbone.net.load_state_dict(state_dict)
```
Then proceed to train the full FasterRCNN model.

Tip

Consider using a lower learning rate for the backbone compared to the detection heads when fine-tuning the entire model.

Warning

The backbone_num_classes argument must match the number of classes used when pretraining the backbone. A mismatch may lead to incompatible final layers or parameter shapes.