关键字: [Amazon Web Services re:Invent 2023, Amazon SageMaker Studio, Machine Learning Development, Sagemaker Studio, Data Science Workflows, Model Building, Model Deployment]
Number of words: 2800, reading time: 14 minutes
video
If the video cannot be played normally, please go to bilibili to watch this video. >> https://www.bilibili.com/video/BV12e411y77f
Introduction
Amazon SageMaker Studio provides comprehensive tools for end-to-end machine learning development, from preparing data to training models, tracking experiments, deploying models, and managing pipelines, all within an integrated development environment. To accelerate generative AI development, you need integrated, purpose-built build tools to train and tune base models (FMs), and a flexible environment for custom machine learning workflows. In this forum, learn about the latest features in SageMaker Studio to help you quickly build, test, fine-tune, and iterate models to improve efficiency and performance. Also learn how BMW Group accelerates AI/machine learning development with SageMaker Studio.
Highlights of speech
The following is the essence of this speech compiled by the editor for you. It has a total of 2,500 words and takes about 12 minutes to read. If you want to know more about the content of the speech or watch the full text of the speech, please watch the full video of the speech or the original text of the speech below.
Sumit Harker, a veteran product owner, held a 30-minute lecture at the Manley Bay Convention Center in Las Vegas that attracted about 500 machine learning practitioners. He warmly welcomed this audience and introduced himself as well as his two distinguished co-speakers: Giuseppe Porcelli, a master solutions architect with over 15 years of experience building AI solutions; Mark Neumann, product lead on the machine learning engineer team at BMW AG. Together, they discussed how to leverage the power of Amazon SageMaker Studio to effectively scale your organization's machine learning development initiatives.
Harker first introduced the background information of SageMaker Studio, pointing out that it was first publicly released in November 2019 and is the first fully integrated development environment designed specifically for machine learning methodologies. This means data scientists, developers, and engineers can access all the tools they need for the entire end-to-end machine learning method lifecycle in a unified visual interface. This gives them access to a full set of capabilities in one integrated environment - from data annotation and feature engineering to model building, hyperparameter tuning, deployment, monitoring and full-process orchestration.
With Studio, practitioners can seamlessly switch between different steps, make modifications, immediately observe results, and quickly iterate to implement new machine learning capabilities. Since its launch 3 years ago, tens of thousands of customers in nearly every industry have adopted SageMaker Studio.
This provides Amazon Web Services’ SageMaker product team with a valuable opportunity to work closely with these users and gain valuable insights into the trends, challenges, and needs that impact the machine learning developer experience.
Harker highlighted some particularly noteworthy trends that emerged from their deep dive customer research. The first major trend is the extremely rapid adoption of artificial intelligence and machine learning in every sector of the global economy.
According to the authoritative AI Index report released annually by Stanford University, global corporate investment in artificial intelligence has increased approximately 13 times in the past decade—from only US$6 billion in 2012 to more than US$79 billion in 2021. Additionally, the report predicts that this frenzied pace of AI investment is likely to accelerate further in the coming years as generative AI advances. For machine learning developers, this proliferation of mainstream applications means a strong need for high-performance tools and features to help developers achieve maximum productivity. Users need to be able to quickly go from data collection and preparation to training models to extracting insights and business value.
The second trend Sumit mentioned is the specialization and proliferation of data and related positions, including data engineers, data scientists, machine learning engineers, MLOps engineers, etc. Given the complexity of developing, deploying, and operating an end-to-end machine learning system, it is unlikely that any single practitioner will possess all the skills required across the entire lifecycle. Data engineers specialize in collecting, integrating, and transforming disparate data sets into a format that can be used for analysis and modeling. Data scientists focus on extracting features, trying out algorithms, tweaking hyperparameters, and building models iteratively. Machine learning engineers excel at bringing models into production by building robust data pipelines, monitoring systems, and automated retraining. These specialized AI fields have experienced rapid growth as demand for machine learning talent surges across industries. As a result, developers are seeking tools tailored to specific tasks relevant to their roles and associated experience levels.
The third major trend identified by Sumit is the ongoing challenge many organizations still face in successfully scaling machine learning models from initial prototypes to large-scale production deployments. According to Gartner industry survey data, only about 50-60% of models can successfully transition from proof of concept to live production environment, indicating that there is still a lot of room for improvement.
The skills, tools, infrastructure, and processes required to iteratively develop a model during the research phase can differ significantly from the rigorous software engineering techniques required to deploy the model into a scalable, reliable, and monitored production system. As a result, developers are looking for smarter tools and frameworks to help their models transition smoothly from exploration to implementation.
Sumit succinctly summarizes these key takeaways and requirements into four “P’s” that encapsulate the evolving machine learning developer experience:
Performance - As the application of AI accelerates in various fields, developers need access to high-performance tools to meet the rapidly growing demand.
Productivity - Eliminate undifferentiated heavy work, allowing developers to increase productivity by focusing on high-value tasks.
Preferences - There is a growing need for purpose-built tools tailored to the specific responsibilities and job roles of data professionals.
Production - Model builders want more self-service capabilities to smoothly transition models from prototype to production.
To directly address four key pillars of the modern machine learning developer experience, the next generation of SageMaker Studio introduces significant innovations in each area:
- An extremely fast startup experience that provides instant access to tools in 5 seconds or less
- More code editor and notebook options for different tasks, including JupyterLab, RStudio and Visual Studio Code
- Built-in AI-assisted features like Amazon CodeWhisperer to automate tedious coding tasks
- New automation tool to streamline the process of converting machine learning code into production pipelines
Sumit then presented to the audience of 500 people a live demonstration of some of the most impactful new features, starting with the streamlined process of configuring SageMaker Studio domains and spaces. He demonstrated that a complete domain can now automatically handle identity management, network configuration and allocation of underlying computing resources with a single click.
A feature-rich development environment is provided within the space for storing computing instances, storage volumes, and runtime preferences. Sumit shows how to create a JupyterLab space, customize the ML compute instance type and allocated storage capacity, and launch the notebook in the browser - the entire environment is ready in less than 60 seconds.
JupyterLab instances are preloaded with Amazon Cloud's SageMaker distribution for Python, an open source project launched by Amazon Cloud that packages the latest versions of the most popular Python data science and machine learning libraries into Docker containers.
Use SageMaker Distributed to ensure a consistent runtime environment between Studio, Amazon SageMaker training jobs, and local development. Of course, users can further install additional libraries and packages based on their workflow needs.
Next, Sumit introduced how to meet the broader needs of note developers by providing multiple code editors and integrated development environments (IDEs) for different tasks.
Machine learning workers can now choose between JupyterLab, RStudio, and Visual Studio Code, with each tool designed for specific activities. For example, a data scientist might like the flexibility of Jupyter notebooks, while a machine learning engineer might choose VS Code to build stable systems.
The VS Code editor integrated into Amazon SageMaker Studio includes pre-installed Amazon Cloud Toolkit extensions, making accessing Amazon Cloud Services simple. It also provides access to over 3000 additional extensions through integration with Open VSX (the open source VS Code registry).
Just like JupyterLab, VS Code instances run on configurable Amazon SageMaker spaces with adjustable compute and storage.
In order to significantly improve developer productivity, the latest Amazon SageMaker Studio version has built-in multiple AI-assisted features. For example, Amazon CodeWhisperer allows developers to automatically generate Python code by adding plain English comments that describe their intent. Developers' productivity increased by 57% when using CodeWhisperer.
Studio Notebooks also feature an integrated chat assistant that developers can engage in natural language conversations with to get suggestions, understand code, and accept recommendations. Developers can pick from different base models, such as Anthropic’s Claude or Hugging Face’s Stardust, to customize their functionality.
Finally, to simplify the transition from prototyping to real-world applications, SageMaker Studio introduces new automation capabilities. Today, the SageMaker Python SDK includes decorators that allow a function to be marked and automatically executed as a SageMaker training task. This enables scaling of prototypes from laptop-less to leveraging more powerful cloud infrastructure without writing any coding changes. Users can combine multiple decorated functions into a production-grade pipeline by applying the new @step decorator.
Notebook Scheduling creates a pipeline of notebook jobs, while the ModelBuilder class simplifies the deployment process by automatically identifying key information such as inferring the type of server and required computing resources.
At this stage, Sumit invited his colleague Giuseppe Porcelli to the stage for a live demonstration of how to use SageMaker Studio for stress testing on end-to-end machine learning problems.
Giuseppe briefly introduced a scenario about using historical sensor data of industrial machinery and equipment to predict failures. The data comes from the machine learning library of the University of California, Irvine, with more than 55,000 samples and 12 input features, such as temperature, pressure, rotation speed. and a binary target variable representing the fault status. He will demonstrate a complete process, including loading and preparing data in a Jupyter notebook, training an XGBoost binary classification model using the preprocessed data to predict failures, converting the training code into a SageMaker job, and deploying the model to the endpoint from VS Code , as well as coordinating the various steps in the SageMaker pipeline.
Giuseppe first created a JupyterLab workspace in Studio, customized the type and storage capacity of the ML computing instance, and pre-selected the SageMaker Docker image. He also showed how to integrate Amazon CodeWhisperer into a notebook to automatically generate relevant operation codes for the accounted data based on pure English annotations.
Data scientist Giuseppe first loads the dataset and performs exploratory analysis, then handles missing values and codes categorical variables. Next, he split the data into 80/20 training and validation sets. Use Scikit-Learn for preprocessing and operate via Pandas. Next, he designed features and trained an XGBoost classifier to predict binary machine failure states. To take full advantage of the more powerful SageMaker computing resources, he added the @remote decorator on the above Python function, thereby automatically running it as a SageMaker training job.
After switching to the Visual Studio Code editor, he uses the new ModelBuilder class to deploy the trained XGBoost model to the SageMaker endpoint. Additionally, he demonstrates real-time inference against deployed models to detect failures from sample input data.
Finally, Giuseppe uses the @step decorator to easily convert data preparation, training, evaluation, registration, and deployment functions from Python code into automated SageMaker pipelines. This pipeline workflow coordinates these 5 steps and requires very little additional code. He executes the pipeline in Studio and demonstrates monitoring progress in the UI, including a model evaluation report that is automatically captured and added to the SageMaker model registry.
After concluding this insightful presentation, Sumit invited Mark Neumann, a representative from BMW, to the stage to describe how the auto manufacturing giant leverages SageMaker Studio to accelerate machine learning development across the enterprise. Mark shared one of his team’s use cases: using computer vision models to automate quality inspections on vehicle assembly lines in BMW factories.
High-resolution cameras above the production line continuously capture images of vehicles passing through multiple inspection points. AI models analyze these images in real time to ensure correct assembly, detect anomalies, and meet specifications and tolerances. Globally, the BMW Group produces more than 500 unique vehicle models at its 31 manufacturing plants. As inspection points increase during production and tooling, lighting, and production line configurations change, these inspection models need to be continuously retrained and updated to maintain their accuracy and reliability.
To achieve this goal, Mark's team works to provide BMW's developers, data scientists, and engineers with a powerful tool and platform so that they can easily build, update, and monitor these AI solutions without having to worry about infrastructure setup , operational and compliance issues. In the past, BMW relied on a series of machine learning platforms tailored to specific teams and use cases, but this fragmented approach led to a number of issues such as high operating costs, difficulty scaling computing resources and the creation of idle capacity during periods of demand fluctuations. waste.
To address these challenges, Mark's team designed a new centralized SageMaker Studio solution to meet the needs of different data science teams within the enterprise. This automated supply chain workflow creates separate, pre-configured Amazon Cloud accounts for each team, including all necessary Studio components, Docker containers with attached file systems for sharing, and security controls. This allows BMW’s platform team to focus on integration, governance and compliance, while developers can quickly onboard and access data to build models.
BMW sees strategic value in Studio’s open source foundation (such as Jupyter and VS Code), the continued speed of innovation brought about by investments in Amazon Cloud Technologies, and the ability to choose the best instance type and size for different workloads. Currently, the company is working to migrate the team to the new SageMaker environment while integrating existing MLOps, model monitoring, interpretability and model governance solutions to achieve a unified end-to-end machine learning platform. Mark expressed his sincere thanks for demonstrating the expanded SageMaker Studio capabilities.
Overall, this insightful 30-minute speech revealed the strategic background and tactical details of how Amazon SageMaker Studio can fully realize the transformative potential of artificial intelligence and bring enterprise benefits by accelerating the process of machine learning projects. Come to value.
Based on years of research and extensive customer experience, Sumit summarizes key trends in developer experience, including a strong need for high performance, productivity, role-specific tools, and a streamlined model production process.
The latest version of SageMaker Studio directly addresses various issues in these areas with innovative features such as 5-second startup time, multiple editing options, AI coding assistant and automatic deployment capabilities.
Giuseppe demonstrates in an engaging way how to create a development environment, generate code, train models, execute tasks, deploy models, and orchestrate the entire process of machine learning pipelines in Studio.
Finally, BMW's Mark Neumann shared how they transitioned their business from decentralized legacy systems to a centrally managed SageMaker Studio platform, achieving scale of development while also focusing on aspects such as integration, governance, and compliance.
BMW plans to use Studio’s open source foundation, new features, and powerful automation capabilities to drive AI innovation across the enterprise and maintain a competitive advantage.
Here are some highlights from the speech:
Leaders discuss how to leverage SageMaker Studio to right-scale machine learning development.
With the SageMaker distribution, JupyterLab, Code Editor and SageMaker Studio are implemented in a unified cross-platform runtime environment.
Amazon Cloud Technology's SageMaker pipeline makes building end-to-end ML workflows simple and easy.
Leaders showcased SageMaker Studio's pipeline interface, which provides a complete ML workflow from data preprocessing to model deployment.
Leaders outlined the benefits of Amazon Cloud Platform services in artificial intelligence and machine learning, enabling developers to easily integrate these capabilities into their applications.
Leaders recalled how Amazon Cloud Technologies began providing a machine learning development environment many years ago so that data scientists and engineers could leverage transfer learning on GPUs to process unstructured data such as images, audio, and text.
Summarize
The video details the latest features of Amazon SageMaker Studio designed to help machine learning developers scale. SageMaker Studio provides an integrated interface for machine learning workflows. Speakers focused on the key trends impacting the machine learning developer experience—rapid AI adoption, the proliferation of data roles, and the challenges of model deployment.
To address these trends, the new SageMaker Studio offers a faster startup experience, including more options, integrated development environments (IDEs) like JupyterLab and Visual Studio Code, and AI-assisted features through tools like CodeWhisperer. Additionally, it simplifies the process of putting the model into use.
During the demonstration, viewers can see how to use SageMaker Studio, SageMaker Training, and SageMaker Pipelines to preprocess data, train XGBoost models, evaluate them, and deploy the models to their destination. Decorators and pipes make it easier to scale up your prototyping workflow.
Speakers from BMW explained how SageMaker Studio helped them consolidate disparate machine learning platforms into a centrally managed and controlled platform. Its main advantages are flexibility, reduced operating costs and a smoother model development process.
Original speech
https://blog.csdn.net/just2gooo/article/details/134996662
Want to know more exciting and complete content? Visit re:Invent official Chinese website now!
2023 Amazon Cloud Technology re:Invent Global Conference - Official Website
Click here to get the latest global product/service information from Amazon Cloud Technology with one click!
Click here to get the latest product/service information from Amazon Cloud Technology China with one click!
Register an Amazon Cloud Technology account now and start your cloud journey!
[Free] Amazon Cloud Technology "Free trial of more than 100 core cloud service products"
[Free] "Free trial of more than 40 core cloud service products" of Amazon Cloud Technology China
Who is Amazon Cloud Technology?
Amazon Cloud Technology (Amazon Web Services) is the pioneer and leader of global cloud computing. Since 2006, it has been characterized by continuous innovation, technology leadership, rich services, and wide application And well-known in the industry. Amazon Cloud Technology can support almost any workload on the cloud. Amazon Cloud Technology currently provides more than 200 full-featured services, covering computing, storage, network, database, data analysis, robotics, machine learning and artificial intelligence, Internet of Things, mobile, security, hybrid cloud, virtual reality and augmented reality, media , as well as application development, deployment and management; the infrastructure covers 99 availability zones in 31 geographical regions, and plans to build 4 new regions and 12 availability zones. Millions of customers around the world, from startups, small and medium-sized enterprises, to large enterprises and government agencies, trust Amazon Cloud Technology. They use Amazon Cloud Technology services to strengthen their infrastructure, improve agility, reduce costs, accelerate innovation, and enhance competitiveness. Achieve business growth and success.
