Finally, Keras once again provides its own ResNet implementations—for instance, tf.keras.applications.ResNet50() (refer to the documentation at https://keras.io/applications/#resnet50)—with the option to load parameters pretrained on ImageNet. These methods have the same signature as previously covered Keras applications.
The list of CNN architectures presented in this chapter does not pretend to be exhaustive. It has been curated to cover solutions both instrumental to the computer vision domain and of pedagogical value.
As research in visual recognition keeps moving forward at a fast pace, more advanced architectures are being proposed, building upon previous solutions (as Highway and DenseNet methods do for ResNet, for instance), merging them (as with the Inception-ResNet solution), or optimizing them for particular use cases (such as the lighter MobileNet...
