Implementing NAS

In this section, we will implement NAS. In particular, our Controller is tasked with generating child network architectures that learn to classify images from the CIFAR-10 dataset. The architecture of the child network will be represented by a list of numbers. Every four values in this list represent a convolutional layer in the child network, each describing the kernel size, stride length, number of filters, and the pooling window size in the subsequent pooling layer. Moreover, we specify the number of layers in a child network as a hyper-parameters. For example, if our child network has three layers, its architecture is represented as a vector of length 12. If we have an architecture represented as [3, 1, 12, 2, 5, 1, 24, 2], then the child network is a two-layer network where the first layer has kernel size of 3, stride length of 1, 12 filters, and a max-pooling window size of 2, and the second layer has kernel size of 5, stride length of 1, 24 filters, and max-pooling...