Leveraging pretrained VGG models for new classes

In 2014, the paper Very Deep Convolutional Networks for Large-Scale Image Recognition (https://arxiv.org/abs/1409.1556) was published. At that time, both the models in the paper, VGG16 and VGG19, were considered very deep, with 16 and 19 layers, respectively. That included weights, in addition to the input and output layers, and a couple of max pooling layers. The network architecture of VGG stacks multiple 3×3 convolutional layers on top of each other. In total, the VGG16 network architecture has 13 convolutional layers and three fully connected layers. The 19-layer variant has 16 convolutional layers and the same three fully connected layers. In this recipe, we will use the bottleneck features of VGG16 and add our own layers on top of it. We will freeze the weights of the original model and train ...