posted 1 year ago
Thanks for your question Michael.
By "the normal way" you could mean one of several things, so I will try to explain what I think are the important features of Sagemaker that make it stand out:
* Sagemaker provides a cloud hosted Jupyter environment. You can start up or shut down a Sagemaker "Jupyter instance" whenever you like and only pay for the time you are using it. You can choose a very wide variety of instance types, from small and cheap, to fast and expensive. Notebook instances can also be scheduled to shut down and start up as you see fit according to rules that suit your organisation. You can choose from a wide variety of pre configured docker containers for these instances, and you can also customise the packages that get installed when an instance is started.
* Sagemaker provides a wide variety of algorithms that have been built by Amazon and preconfigured to perform well in the cloud. Sagemaker allows you to train models built using these algorithms in a serverless way, where you can specify the instance type to use during training. Sagemaker will spin up an instance just to run the training job, before automatically shutting it down. Again, you only pay for the time the training instance is running. This can make it economical to run huge training instances that you would not otherwise be able to afford. It may even be cheaper to run a neural net model with 16 GPUs in parallel than with no GPUs, as well as being much faster.
* The Sagemaker API also has a prediction step. This allows you to either preconfigure a production ready end point that stays up until you choose to shut it down, or to run once off batch transform jobs just when you need to do a prediction on a batch of rows. This aspect of Sagemaker is very valuable, because the process of getting a model into production is usually the most time consuming and costly part of building any ML model for business.
There are other benefits as well, some of which we discuss in the book, but I think that hits the main ones
