Computer Vision News - December 2018

26 Computer Vision News Focus on: BOHB Focus on Code: In the following example we will show how to set up a small CNN network on Keras and train it on the MNIST dataset. The table below shows the common hyperparameters for which we will be seeking optimal values, including dependencies between hyperparameters (for instance, momentum only receives a value if SGD optimizer is used). The code is short, simple and pretty self-explanatory. The following are the three basic ingredients needed to apply HpBandSter to a new optimization problem: (1) Implementing a Worker -- the worker is responsible for evaluating a given model with a single configuration on a single budget at a time. (2) Defining the Search Space -- defining the parameters to be optimized. (2) Picking the Budgets and the Number of Iterations -- besides the number of iterations you also need to specify a meaningful budget (as will be discussed later, this is one of the drawbacks of HpBandSter). The __init__ function As its name implies, this is the initialization function -- it’s responsible for reading the data, uploading data and normalizing data -- to prepare it for training. In this case we are reading the well known MNIST dataset.