The Keras API has an official implementation of these architectures, accessible via its tf.keras.applications package (refer to the documentation at https://www.tensorflow.org/api_docs/python/tf/keras/applications). This package contains several other well-known models and provides pre trained parameters for each (that is, parameters saved from prior training on a specific dataset). For instance, you can instantiate a VGG network with the following command:
vgg_net = tf.keras.applications.VGG16(
include_top=True, weights='imagenet', input_tensor=None,
input_shape=None, pooling=None, classes=1000)
With these default arguments, Keras instantiates the VGG-16 network and loads the persisted parameter values obtained after a complete training cycle on ImageNet. With this single command, we have a network ready to classify images into the 1,000 ImageNet categories. If we would like to retrain the network from scratch instead, we should fix weights...
