8. SSD model in Keras
Listing 11.8.1 shows the SSD model creation function build_ssd(). The model is illustrated in Figure 11.5.1. The function retrieves n_layers of output features from the backbone or base network by calling base_outputs = backbone(inputs).
In this book, the backbone() is build_resnet(). The ResNet models that can be generated by build_resnet() are similar to the residual networks discussed in Chapter 2, Deep Neural Networks. The build_resnet() function can be replaced by any function name that builds the base network.
As shown in Figure 11.5.1, the return value base_outputs is a list of output features that will be the input to class and offset prediction layers. For example, the first output base_outputs[0], is used to generate n1 class predictions and n1 offset predictions.
In the for loop of build_ssd() the class prediction is the classes variable, while the offsets prediction is the offsets variable. After the for loop iteration, the class predictions...
