Summary
By realizing the benefit of parallelism that comes from multiple devices and machines, we have learned about various ways to train a DL model. First, we learned how to use multiple CPU and GPU devices on a single machine. Then, we covered how to utilize the built-in features of TF and PyTorch to achieve the training in a distributed fashion, where the underlying cluster is managed explicitly. After that, we learned how to use SageMaker for distributed training and scaling up. Finally, the last three sections described frameworks that are designed for distributed training: Horovod, Ray, and Kubeflow.
In the next chapter, we will cover model understanding. We will learn about popular techniques for model understanding that provide some insights into what is happening within the model throughout the training process.
