Overview
Every Developer/data Scientist Machine Learning Resource
Amazon’s AWS SageMaker makes it possible to build, train and deploy machine learning models quickly. It is fully managed and covers the entire ML (machine learning) workflow. A lot of large companies use SageMaker resource. With SageMaker you can; collect and prepare training data, choose and optimize your ML algorithms, setup and manage training environments, and tune your model for optimization. Models can be deployed and managed in production. Use reinforcement learning to build smart outcomes. SageMaker is open and flexible and AWS will provide detailed instructions on how to use the SageMaker resource.
The entire ML workflow is covered with AWS SageMaker. Build your ML project efficiently. Label and and prepare data by collecting training data. SageMaker has a library for common problems that contain data labeling and prebuilt notebooks. SageMaker has a marketplace for algorithms and models, as well as built-in high performance algorithms. Train the model in environments you setup and manage in a one-click high performance infrastructure. Your model can be optimized and tuned for deployment in an environment that makes smart decisions and takes actions based off those predictions. Use one-click deployment on your model in a scalable managed environment that can save up to 75% with auto-scaling clusters.
There are numerous companies that already use AWS SageMaker. Companies like Siemens, State Farm, Intuit, NFL, Expedia Group, Liberty Mutual, Coinbase, GE Healthcare, Korean Air, and Change Healthcare. These companies take advantage of the artificial intelligence and ease of deployment in SageMaker.
SageMaker ML starts with collecting and preparing training data. You can quickly label training data with SageMaker Ground Truth. Ground Truth gives you the ability to build and manage highly accurate data sets. Public and private human labelers can get pre-built workflows and interfaces for common labeling tasks. Ground Truth learns from the labels to make high quality and automatic annotations which can lower the costs of ML instances by 70%. There is a preparation and training loop used by Ground Truth. The loop starts with the AWS S3 resource which provides the raw data to the human labelers. The data is then trained by the active learning model from the human labeled data. If the model is understood, then an accurately labeled training data set is sent to SageMaker. If the data is not understood, then it is sent back for human labeling and reran through the loop. Ground Truth has hosted notebooks like Jupyter. Though it is possible to use your own to visualize and develop the data you will be modeling. Pre-built notebooks can also be used as-is, or modified for specificity. SageMaker has available solutions for many business problems. Recommendation and personalization, detecting fraud, forecasting, classifying images, customer targeting, churn prediction, log processing and anomaly detection, and speech to text are common uses for SageMaker ML.
Choose and automatically optimize your ML algorithms from popular platforms like TensorFlow, Apache MXNet, PyTorch, Chainer, Scikit Learn, SparkML, Horovod, Keras, or Gluon. Sagemaker has common ML algorithms built in and tuned for scale. Explore the marketplace where there are over 200 additional pre-trained models and algorithms. If you want to build additional algorithms, you can use Docker containers or bring your own framework or other algorithm.
With one-click, you can setup and manage training environments that scale up to petabytes. This is all because SageMaker manages the heavy lifting of the underlying infrastructure. SageMaker truly is the fastest distributed cloud ML. It uses EC2 P3 instances that provide 8 NVIDIA Tesla GPUs. Get 64 scalable vCPUs on Intel Xeon Skylake with AVX-512, 25 GBPS networking throughput, and 16 GB of memory per GPU. SageMaker is the best place to run TensorFlow because of the optimized near linear scaling across hundreds of GPUs. You will be able to run cloud scale operations without a lot of overhead, and save time gaining access to more sophisticated models.
Tune and optimize your training models automatically for accuracy and speed. Avoid the manual parameter adjustments in the trial and error stage of your models. Hyperpearemeter optimization discovers interesting features in your training model data sets and trains those features in your model to save you time. Train your model once, and run it anywhere with SageMaker NEO. You will be able to use any framework with no loss in accuracy.
Through a process called “inference” you can deploy and manage your model production. This can be done in the cloud to start generating predictions. The models run on SageMaker auto-scaling clusters which are set-up for high availability and performance. A/B testing can be used for experimenting different scenarios to see which will yield the best results. Elastic inference accelerates SageMaker instances and reduces deep inference costs by 75%. Optimizing deep learning instances is hard to do with just instance, so elastic inference picks the best instance type for your workload to be optimized, and then picks the GPU acceleration that works best for your environment. AWS IoT Greengrass deploys SageMaker models on the edge and runs inference. Greengrass can run lambda functions on connected devices, keep device data insync, and communicate with other devices securely even when your connection is down.
RL (Reinforcement learning) can help you build specific outcomes without the need for pre-labeled training data. There will be some scenarios where you have a specific outcome but don’t have the “right” answer. Imagine times when you need to get a robot to drive a car, or when you need to model financial data for your company. In these cases you can use the simulator instead of historic data to get the desired outcome based on a rewards and penalties environment. RL has built-in and fully managed algorithms for frameworks like TensorFlow, MXNet. You can also use custom frameworks like Intel Coach or RayLlib to built from the ground up. SageMaker RL supports multiple learning environments; 2D, 3D, and commercial simulation environments (MATLAB, Simulink, Anything that supports OpenAI Gym Interface, custom developed environments.) You can also train in virtual 3D environments, like AWS Sumerina or AWS Robomaker, to model everything (advertising, financial systems, commercial controls, robotics, and autonomous vehicles.)
AWS SageMaker is open and flexible, so you can use ML your way. You can use a broad set of framework and tools to keep up with the fast nature of machine learning. There is also better edge performance offered with SageMaker. The open source NEO project provides the capabilities of SageMaker to every developer. By being open soured ML can realize its full potential. Hardware vendors can improve NEO with new optimizations thereby advancing the ML landscape. Remember, SageMaker fits your workflow because of the different components it is made of (Ground Truth, Notebooks, Training, Neo, Hosting.) The service provides an end-to-end ML atmosphere because they are designed to work together. The components can be used individually to supplement existing ML workflows or to support models that run in individual data centers or on the edge.
AWS encourages you to get involved with learning about ML. There is a DeepRacer autonomous robot that helps you learn about RL through autonomous driving. Use DeepLens to learn about computer vision through special cameras. If you want a step-by-step method, you can join the ML certification course or complete the ML solutions lab. Either way, you will have access to hands-on educational workshops, brainstorming sessions, and professional advisory services. So get involved with AWS SageMaker today!
