Changes This release adds two new APIs (UpdateWorkforce and DescribeWorkforce) to SageMaker Ground Truth service for workforce IP whitelisting.
Changes SageMaker ListTrialComponents API filter by TrialName and ExperimentName.
Changes You can now use SageMaker Autopilot for automatically training and tuning candidate models using a combination of various feature engineering, ML algorithms, and hyperparameters determined from the user's input data. SageMaker Automatic Model Tuning now supports tuning across multiple algorithms. With Amazon SageMaker Experiments users can create Experiments, ExperimentTrials, and ExperimentTrialComponents to track, organize, and evaluate their ML training jobs. With Amazon SageMaker Debugger, users can easily debug training jobs using a number of pre-built rules provided by Amazon SageMaker, or build custom rules. With Amazon SageMaker Processing, users can run on-demand, distributed, and fully managed jobs for data pre- or post- processing or model evaluation. With Amazon SageMaker Model Monitor, a user can create MonitoringSchedules to automatically monitor endpoints to detect data drift and other issues and get alerted on them. This release also includes the preview version of Amazon SageMaker Studio with Domains, UserProfiles, and Apps. This release also includes the preview version of Amazon Augmented AI to easily implement human review of machine learning predictions by creating FlowDefinitions, HumanTaskUis, and HumanLoops.
Changes Amazon SageMaker now supports multi-model endpoints to host multiple models on an endpoint using a single inference container.
Changes Adds support for the new family of Elastic Inference Accelerators (eia2) for SageMaker Hosting and Notebook Services
Changes Enable G4D and R5 instances in SageMaker Hosting Services
Changes Amazon SageMaker now supports Amazon EFS and Amazon FSx for Lustre file systems as data sources for training machine learning models. Amazon SageMaker now supports running training jobs on ml.p3dn.24xlarge instance type. This instance type is offered as a limited private preview for certain SageMaker customers. If you are interested in joining the private preview, please reach out to the SageMaker Product Management team via AWS Support."
Changes Amazon SageMaker introduces Managed Spot Training. Increases the maximum number of metric definitions to 40 for SageMaker Training and Hyperparameter Tuning Jobs. SageMaker Neo adds support for Acer aiSage and Qualcomm QCS605 and QCS603.
Changes The default TaskTimeLimitInSeconds of labeling job is increased to 8 hours. Batch Transform introduces a new DataProcessing field which supports input and output filtering and data joining. Training job increases the max allowed input channels from 8 to 20.
Changes Workteams now supports notification configurations. Neo now supports Jetson Nano as a target device and NumberOfHumanWorkersPerDataObject is now included in the ListLabelingJobsForWorkteam response.
Changes Amazon SageMaker Automatic Model Tuning now supports random search and hyperparameter scaling.
Changes SageMaker notebook instances now support enabling or disabling root access for notebook users. SageMaker Neo now supports rk3399 and rk3288 as compilation target devices.
Changes SageMaker Training Jobs now support Inter-Container traffic encryption.
Changes Batch Transform Jobs now supports TFRecord as a Split Type. ListCompilationJobs API action now supports SortOrder and SortBy inputs.
Changes Amazon SageMaker Automatic Model Tuning now supports early stopping of training jobs. With early stopping, training jobs that are unlikely to generate good models will be automatically stopped during a Hyperparameter Tuning Job.
Changes Amazon SageMaker now has Algorithm and Model Package entities that can be used to create Training Jobs, Hyperparameter Tuning Jobs and hosted Models. Subscribed Marketplace products can be used on SageMaker to create Training Jobs, Hyperparameter Tuning Jobs and Models. Notebook Instances and Endpoints can leverage Elastic Inference accelerator types for on-demand GPU computing. Model optimizations can be performed with Compilation Jobs. Labeling Jobs can be created and supported by a Workforce. Models can now contain up to 5 containers allowing for inference pipelines within Endpoints. Code Repositories (such as Git) can be linked with SageMaker and loaded into Notebook Instances. Network isolation is now possible on Models, Training Jobs, and Hyperparameter Tuning Jobs, which restricts inbound/outbound network calls for the container. However, containers can talk to their peers in distributed training mode within the same security group. A Public Beta Search API was added that currently supports Training Jobs.
Changes SageMaker now makes the final set of metrics published from training jobs available in the DescribeTrainingJob results. Automatic Model Tuning now supports warm start of hyperparameter tuning jobs. Notebook instances now support a larger number of instance types to include instances from the ml.t3, ml.m5, ml.c4, ml.c5 families.
Changes SageMaker notebook instances can now have a volume size configured.
Changes VolumeKmsKeyId now available in Batch Transform Job
Changes Added an optional boolean parameter, 'DisassociateLifecycleConfig', to the UpdateNotebookInstance operation. When set to true, the lifecycle configuration associated with the notebook instance will be removed, allowing a new one to be set via a new 'LifecycleConfigName' parameter.