DSVMs and DLVMs are good at carrying out single node-based computing. In scenarios where we need to distribute training, however, we can use the Batch AI service, which allows us to focus on training instead of having to worry about managing the cluster. A Batch AI service has VMs that use the same base image as the DSVM, meaning that all the libraries, tools, and frameworks that are available in a DSVM are available in the Batch AI service as well. The Batch AI service allows us to use parallel training and GPU-based VMs for deep learning, and we can also deploy a Docker container to a Batch AI node. When using the Batch AI service, we can mount our Azure Blob or Azure Data Lake Storage with our cluster. This means that we can train with a huge amount of data without having to copy the data to the cluster because it can be streamed instead.
At the time of writing...
