This workshop demonstrates two methods of machine learning inference for globally-scalable production using AWS Lambda and Amazon SageMaker. Scikit-learn is a popular machine learning library that covers most aspects of shallow learning. With these techniques the library module and your custom inference code can be combined into a flexible package for immediate cloud deployment.
Scikit-learn is vast and deep. It's an essential tool for every data scientist's workbench. However, it can't do everything. Importantly, it can't call inference on deep learning models, especially those created with MXNet, TensorFlow, or pytorch. Amazon SageMaker is ideally suited to build and deploy models with those frameworks. The SageMaker feature batch-transform is demonstrated calling inference in the cloud on-demand. As such it can handle nearly every inference workload without the need for a costly persistent endpoint.
AI enables machine functionality that can be interpreted as human intelligence. In practice this is the ability to infer predictions based on data collected from previous experiences. Machine learning, a subset of AI, is currently enjoying a renascence from its dormant state since the mid-90s thanks to new GPU-driven software techniques.
Serverless computing enables you to build & run applications without thinking about servers and other network infrastructure. Relieving a developer from the burden of bootstraping, configuring, networking, clustering, updating, and all aspects of physical server management is an enormous benefit. It enables one to focus on the tasks at hand without sacrificing global access and scalability. In most cases serverless computing dramatically reduces the cost of application workloads. Benefits include:
This underlying compute infrastructure scales underneath your application It means:
Though the capacity for SageMaker is limitless there is a cost when using Batch Transformation: latency. When calling inference on Batch Transformation you must endure the cost of loading a Docker image for the workload. This usually takes 2-3 minutes. There's no way to speed this up. If millisecond latency is required for your application you should not choose this option. Instead take advantage of SageMaker's ability to deploy persistent endpoints.
We recommend you start with Lambda and SciKit Learn Inference and complete the workshop until you have successful results. Then, provide new test data to explore the robustness of your model. When you're satisfied with your ability to deploy scikit-learn move on to SageMaker Batch Transform. Here the possibilities open up with access to deep learning frameworks and limitless compute resources.
If any notebook entry is unclear please ask for help either in the workshop or at http://bit.ly/serverlessAI.