Training AI Models

Abstract

Training AI models is usually more demanding than training standard ML models because they are processing intensive and often the data sets involved are larger. That is why if you are serious about deep learning you have to have access to GPUs. In Azure there are a number of ways you can make use of GPUs, on single VMs or in orchestrated clusters of them. In this chapter, we summarize several of the most common methods available as well as the pros and cons of each. Then we expand on the code we wrote in Chapter 6, which used a VGG-like CNN to tackle the CIFAR10 data set using the DLVM as the computing environment. In this chapter, we extend to other training options such as Batch AI and Batch Shipyard, which can both be useful for scaling up or scaling out training. We finish by highlighting briefly some of the other methods of training AI models on Azure that are not as common but might be useful depending on the problem at hand.