Created by: ZeHuiGong

Some architectures need to initialize two or more backbones (e.g. ResNet50) when extracting features from the input images, and these backbones DO NOT share weights but can load pre-trained weights (such as pre-trained on ImageNet). Because in the PaddleDetection project, the backbone or neck passed into the architecture, e.g., Faster R-CNN is a python class object. In this case, how do I create two or more backbones that do not share weights, in the architecture?