AWS Bolsters SageMaker with New Capabilities
Amazon Web Services unveiled a half-dozen new SageMaker services today at its re:Invent conference in Las Vegas. The cloud giant bolstered its flagship AI development tool with new capabilities for data labeling, integration with data engineering and analytics workflows, and serverless deployments, as well as an entry-level version that’s free.
Amazon VP of AI Swami Sivasubramanian unveiled the SageMaker news during his two-hour keynote today at re:Invent, which is in its third day in returning to Las Vegas after a one-year hiatus due to the COVID-19 pandemic.
In a nod to the critical importance of data quality, AWS launched Amazon SageMaker Ground Truth Plus, which essentially is a professional services version of SageMaker Ground Truth, which is already available.
This new service enables customers to tap into a pool of expert data labelers who have been curated by AWS, and to have the data labeling process directly integrated with their SageMaker environment. AWS says the new offering can cut data labeling costs by up to 40%. You can find more information here.
Amazon SageMaker Studio, meanwhile, has been bolstered with new integrations to EMR, the company’s Hadoop-based service that provides access to frameworks like MapReduce, Spark, Presto, and Hive. SageMaker Studio users can now create, terminate, manage, discover, and connect to EMR clusters directly from within their SageMaker Studio environment, which should streamline workflows for data scientists.
There was some integration between the environments previously, but SageMaker Studio users could only access EMR directly if they were in the same account. AWS has also introduced templates, which is a new way to configure and provision clusters with support from DevOps pros. It also added the capability for data scientists to connect to, debug, and monitor EMR-based Spark jobs from within a SageMaker Studio Notebook. Check out this link for more information on the SageMaker Studio enhancements.
Training of deep learning models on GPUs will get faster with the new Amazon SageMaker Training Compiler. This capability will automatically compile your Python training code (PyTorch or TensorFlow) and generate GPU kernels specifically for your model, AWS says. By making “incremental optimizations” beyond what the native PyTorch and TensorFlow frameworks offer to maximize compute and memory utilization of GPUs, the software can cut training time by up to 50%.
AWS says it can take up to 25,000 GPU-hours to train the RoBERTa natural language processing (NLP) model. Skilled machine learning engineers can cut that time, but not everybody has those skills. AWS says the SageMaker Training Compiler helped to fine-tune Hugging Face’s GPT-2 model and cut training time from about 3 hours to 90 minutes. You can learn more about it here.
Deployment of machine learning models should improve with the new Amazon SageMaker Inference Recommender. It could take a bit of trial and error to figure out the right instance and configuration for a given ML model. That can be shortcut with this new offering, which provides optimized recommendations for the ML inference.
Once an MLOps engineer has received the recommendations, she can instantly deploy it to the selected instance type with only a few clicks, AWS says. For more information about this from AWS, click here.
Finally, for the utmost in speed and simplicity, AWS offers Amazon SageMaker Serverless Inference. As its name suggests, this new tool eliminates the need for a SageMaker user to make any decisions at all about which instance to choose for their deployed model.
AWS says Serverless Inference is ideal for workloads that are erratic and can’t be predicted, such as a chatbot used by a payment processing company. Customers pay only for the compute they’re used (billed to the millisecond). This offers a fourth option for inference, along with SageMaker Real-Time Inference, SageMaker Batch Transform, and SageMaker Asynchronous Inference. For more info from AWS, click here.