This recipe uses Keras because the framework has a pre-cooked implementation of the above modules. Keras automatically downloads the weights of each network the first time it is used and stores these weights on your local disk. In other words, you don't need to retrain the networks but can leverage training already available on the internet. This is true under the assumption that you would like to classify your network on the 1,000 predefined categories. In the next recipe, we will see how to start with these 1,000 categories and extend them to a customized set with a process known as transfer learning.