Deploying Machine Learning Models on AWS SageMaker: The Power of Built-in Algorithms
The field of machine learning has progressed rapidly in recent years, with breakthroughs in deep learning and other techniques enabling powerful new applications. But for many organizations, the path from promising experiments to production-ready ML systems remains challenging.
Data scientists often spend inordinate amounts of time on infrastructure and DevOps rather than modeling. One survey found that data scientists spend 60% of their time on data cleaning and organizing, and only 12% on mining data for patterns. Another study found they spend 80% of their time on "data preparation."
Fortunately, the rise of managed ML platforms in the cloud is making the end-to-end machine learning workflow faster and easier than ever before. These platforms automate the undifferentiated heavy lifting of infrastructure management and allow data scientists to focus on what they do best – iterating on features and algorithms to build great models.
Enter Amazon SageMaker
Launched in 2017, Amazon SageMaker is a fully managed service that enables data scientists and developers to quickly build, train, and deploy machine learning models at scale. It‘s achieving rapid adoption – over 10,000 customers including Intuit, Tinder, Voodoo, ADP, Cerner, Ryanair, and many more use SageMaker to power their ML initiatives.
SageMaker provides a complete set of tools for the ML workflow:
- Data Preparation: Convert raw data into features for model training
- Model Training: Choose an ML algorithm, specify hyperparameters, and train on GPUs or CPUs
- Model Hosting: Deploy trained models to auto-scaling endpoints for real-time inference
- Model Evaluation: Monitor model quality and detect concept drift in production
One of the key value propositions of SageMaker is the ability to train and deploy models without having to manage servers or containers. Everything runs in a secure, scalable environment optimized for machine learning.
Built-in Algorithms: A Leg Up on Data Science
Among SageMaker‘s most powerful features are the built-in algorithms – a suite of optimized, scalable algorithms that take the guesswork out of model building. Teams can use these as a starting point instead of coding algorithms from scratch, often getting better performance out-of-the-box.
At the time of writing, SageMaker provides over 15 built-in algorithms covering use cases such as:
- Supervised learning for classification and regression
- Unsupervised learning for anomaly detection and clustering
- Reinforcement learning
- Deep learning for vision and language
- Recommendations and personalization
- Forecasting and predicting rare events
Here are some of the most popular built-in algorithms and examples of problems they solve:
| Algorithm | Use Cases |
|---|---|
| XGBoost | Classification, regression, ranking |
| DeepAR | Time series forecasting |
| BlazingText | Text classification, Word2Vec |
| Object2Vec | Embeddings for recommendations |
| Image Classification | Image classification |
| Seq2Seq | Neural machine translation |
These algorithms have been battle-tested and optimized to run at scale across different infrastructure configurations. Data scientists can focus on feature engineering and hyperparameter tuning rather than debugging custom code.
For example, SageMaker‘s XGBoost consistently achieves benchmark performance on gradient boosted trees. It is 4x faster than open source XGBoost on datasets with millions of features. And teams at Amazon use it in mission critical applications – it powers many customer facing services for Amazon.com.
A Closer Look at SageMaker Features
Beyond the built-in algorithms, what makes SageMaker so powerful for end-to-end machine learning? Let‘s take a closer look at some key features and how they accelerate ML projects.
Notebooks and Hosted Environments
SageMaker provides hosted Jupyter notebooks that make it easy to explore data and build models. These come pre-configured with core libraries like TensorFlow, PyTorch, Scikit-learn. No more worrying about package management or environment drift.
You can also Bring Your Own Containers with customized environments, or use SageMaker Studio for a fully integrated development environment.
Automatic Model Tuning
Training high-quality models requires experimenting with different combinations of hyperparameters. But given the large number of possibilities, it‘s impractical to do this manually.
SageMaker provides automatic model tuning that makes it easy to find the best version of a model. Just specify the hyperparameters you want to vary and their ranges, and SageMaker will search through them to find the optimal configuration.
Under the hood, SageMaker uses a technique called Bayesian optimization along with gradient-free optimization to search the space efficiently. In benchmarks, this finds high-performing models in 10-20 training jobs, compared to hundreds when using random search.
Distributed Training
Training on large datasets is key to building high-quality models, but it can take a long time on a single machine. SageMaker enables distributed training across fleets of machines to drastically reduce training times.
SageMaker‘s data parallel library makes it easy to distribute deep learning training across multiple GPUs and machines. It automatically handles details like gradient synchronization and checkpointing. In benchmarks, this scaled nearly linearly up to 256 GPUs, reducing training time from days to hours.
For some algorithms like XGBoost, SageMaker supports distributed training out-of-the-box. You simply specify the number and type of instances, and SageMaker will automatically split the data and coordinate training.
Experiments and Tracking
SageMaker Experiments make it easy to organize and track iterations when building models. You can automatically capture input parameters, configurations, and model metrics, and compare experiments visually in Amazon SageMaker Studio.
SageMaker also integrates with tools like MLflow for experiment tracking and model management. You can use MLflow with SageMaker to track model training runs, store model artifacts, and deploy models to endpoints.
Debugger and Model Monitor
SageMaker Debugger provides full visibility into the model training process, making it easier to diagnose issues and optimize model performance. It automatically identifies common problems like vanishing gradients and captures relevant data for analysis.
SageMaker Model Monitor continuously monitors ML models in production, alerting you to quality issues and concept drift. It compares live traffic to a baseline and provides reports and alerts when drift is detected, so you can update models proactively.
AutoML with Autopilot
For teams that want to go even faster, SageMaker Autopilot provides a one-click AutoML solution to build high-quality models automatically.
Simply specify the input data and prediction target – Autopilot will preprocess the data, select the best algorithm & hyperparameters, and deploy a model endpoint. Data scientists stay in full control and can inspect model details and refine them as needed.
Autopilot leverages AutoGluon, an open-source AutoML library built by Amazon, to quickly produce high-quality models with minimal code. In benchmarks, Autopilot-generated models achieve accuracy close to the state-of-the-art across datasets.
Customer Success Stories
Hundreds of customers are using SageMaker‘s built-in algorithms to achieve impactful results with machine learning. Here are a few examples:
-
Intuit uses DeepAR in SageMaker to provide more accurate, granular forecasts to power its innovations like smart money decisions in Mint, effortless invoice tracking in QuickBooks, and salary insights in TurboTax. They have seen a 50% improvement in forecast accuracy.
-
Tinder uses SageMaker for personalized recommendations in over 190 countries. They use embeddings generated by the built-in Object2Vec algorithm to represent users and train deep neural networks to predict likely matches. This has led to a 20% increase in engagement.
-
Equinox Fitness uses SageMaker‘s XGBoost for propensity modeling to predict which members are likely to churn. They then design personalized interventions to keep members engaged and reaching their fitness goals. This has resulted in a 55% reduction in member churn.
The Future is AutoML
As powerful as SageMaker‘s built-in algorithms are today, we‘ve only begun to scratch the surface of what‘s possible with AutoML. Techniques like neural architecture search, automated feature engineering, and learning-to-learn (i.e. learning how to learn) are poised to accelerate data science even further.
Amazon continues to innovate in this direction with services like SageMaker Autopilot. The long-term vision is what Jeff Bezos calls "Machine learning for everyone" – enabling developers and analysts to build ML-powered applications without needing deep data science expertise.
As Swami Sivasubramanian, VP of AI/ML at AWS, puts it: "Our goal is to democratize machine learning and make it available to all developers." With tools like SageMaker and AutoML, that future is quickly becoming a reality.
Getting Started
Now that you‘ve seen the power of SageMaker‘s built-in algorithms, you‘re probably eager to get started. The best way is to dive in with hands-on tutorials and examples. Here are some great resources:
With SageMaker, you don‘t need to be an expert in infrastructure or DevOps to do machine learning. Focus on the data science, and let SageMaker handle the rest. Happy modeling!
