Amazon SageMaker

 

• A fully managed service that allows data scientists and developers to easily build, train, and deploy machine learning models at scale.
• Provides built-in algorithms that you can immediately use for model training.
• Also supports custom algorithms through docker containers.
• One-click model deployment.

Concepts

• Hyperparameters
  • It refers to a set of variables that controls how a model is trained.
  • You can think of them as “volume knobs” that you can tune to acquire your model’s objective.
• Automatic Model Tuning
  • Finds the best version of a model by automating the training job within the limits of the hyperparameters that you specified.
• Training
  • The process where you create a machine learning model.
• Inference
  • The process of using the trained model to make predictions.
• Local Mode
  • Allows you to create and deploy estimators to your local machine for testing.
  • You must install the Amazon SageMaker Python SDK on your local environment to use local mode.

Common Training Data Formats For Built-in Algorithms

• CSV
• Protobuf RecordIO
• JSON
• Libsvm
• JPEG
• PNG

Input modes for transferring training data

• File mode
  • Downloads data into the SageMaker instance volume before model training commences.
  • Slower than pipe mode
  • Used for Incremental training
• Pipe mode
  • Directly stream data from Amazon S3 into the training algorithm container.
  • There’s no need to procure large volumes to store large datasets.
  • Provides shorter startup and training times.
  • Higher I/O throughputs
  • Faster than File mode.
  • You MUST use protobuf RecordIO as your training data format before you can take advantage of the Pipe mode.

Two methods of deploying a model for inference

• Amazon SageMaker Hosting Services
  • Provides a persistent HTTPS endpoint for getting predictions one at a time.
  • Suited for web applications that need sub-second latency response.
• Amazon SageMaker Batch Transform
  • Doesn’t need a persistent endpoint
  • Get inferences for an entire dataset

Optimization

• Convert training data into a protobuf RecordIO format to make use of Pipe mode.
• Use Amazon FSx for Lustre to accelerate File mode training jobs.

Monitoring

• You can publish SageMaker instance metrics to the CloudWatch dashboard to gain a unified view of its CPU utilization, memory utilization, and latency.
• You can also send training metrics to the CloudWatch dashboard to monitor model performance in real-time.
• Amazon CloudTrail helps you detect unauthorized SageMaker API calls.

Pricing

• The building, training, and deploying of ML models are billed by the second, with no minimum fees and no upfront commitments.