Create and Start a Human Loop - Amazon SageMaker

Create and Start a Human Loop

A human loop starts your human review workflow and sends data review tasks to human workers. When you use one of the Amazon A2I built-in task types, the corresponding AWS service creates and starts a human loop on your behalf when the conditions specified in your flow definition are met. If no conditions are specified in your flow definition, a human loop is created for each object. When using Amazon A2I for a custom task, a human loop starts when your application calls StartHumanLoop.

Use the following instructions to configure a human loop with Amazon Rekognition or Amazon Textract built-in task types and custom task types.

Prerequisites

To create and start a human loop, you must attach the AmazonAugmentedAIFullAccess policy to the AWS Identity and Access Management (IAM) user or role that configures or starts the human loop. This is the identity that you use to configure the human loop using HumanLoopConfig for built-in task types. For custom task types, this is the identity that you use to call StartHumanLoop.

Additionally, when using a built-in task type, your user or role must have permission to invoke API operations of the AWS service associated with your task type. For example, if you are using Amazon Rekognition with Augmented AI, you must attach permissions required to call DetectModerationLabels. For examples of identity-based policies you can use to grant these permissions, see Amazon Rekognition Identity-Based Policy Examples and Amazon Textract Identity-Based Policy Examples. You can also use the more general policy AmazonAugmentedAIIntegratedAPIAccess to grant these permissions. For more information, see Create a User With Permissions to Invoke Amazon A2I, Amazon Textract, and Amazon Rekognition API Operations.

To create and start a human loop, you need a flow definition ARN. To learn how to create a flow definition (or human review workflow), see Create a Human Review Workflow.

Important

Amazon A2I requires all S3 buckets that contain human loop input image data to have a CORS policy attached. To learn more about this change, see CORS Permission Requirement.

Create and Start a Human Loop for a Built-in Task Type

To start a human loop using a built-in task type, use the corresponding service's API to provide your input data and to configure the human loop. For Amazon Textract, you use the AnalyzeDocument API operation. For Amazon Rekognition, you use the DetectModerationLabels API operation. You can use the AWS CLI or a language-specific SDK to create requests using these API operations.

Important

When you create a human loop using a built-in task type, you can use DataAttributes to specify a set of ContentClassifiers related to the input provided to the StartHumanLoop operation. Use content classifiers to declare that your content is free of personally identifiable information or adult content.

To use Amazon Mechanical Turk, ensure your data is free of personally identifiable information, including protected health information under HIPAA. Include the FreeOfPersonallyIdentifiableInformation content classifier. If you do not use this content classifier, SageMaker does not send your task to Mechanical Turk. If your data is free of adult content, also include the 'FreeOfAdultContent' classifier. If you do not use these content classifiers, SageMaker may restrict the Mechanical Turk workers that can view your task.

After you start your ML job using your built-in task type's AWS service API, Amazon A2I monitors the inference results of that service. For example, when running a job with Amazon Rekognition, Amazon A2I checks the inference confidence score for each image and compares it to the confidence thresholds specified in your flow definition. If the conditions to start a human review task are satisfied, or if you didn't specify conditions in your flow definition, a human review task is sent to workers.

Create an Amazon Textract Human Loop

Amazon A2I integrates with Amazon Textract so that you can configure and start a human loop using the Amazon Textract API. To send a document file to Amazon Textract for document analysis, you use the Amazon Textract AnalyzeDocument API operation. To add a human loop to this document analysis job, you must configure the parameter HumanLoopConfig.

When you configure your human loop, the flow definition you specify in FlowDefinitionArn of HumanLoopConfig must be located in the same AWS Region as the bucket identified in Bucket of the Documentparameter.

The following table shows examples of how to use this operation with the AWS CLI and AWS SDK for Python (Boto3).

AWS SDK for Python (Boto3)

The following request example uses the SDK for Python (Boto3). For more information, see analyze_document in the AWS SDK for Python (Boto) API Reference.

import boto3 textract = boto3.client('textract', aws_region) response = textract.analyze_document( Document={'S3Object': {'Bucket': bucket_name, 'Name': document_name}}, FeatureTypes=["TABLES", "FORMS"], HumanLoopConfig={ 'FlowDefinitionArn': 'arn:aws:sagemaker:aws_region:aws_account_number:flow-definition/flow_def_name', 'HumanLoopName': 'human_loop_name', 'DataAttributes': {'ContentClassifiers': ['FreeOfPersonallyIdentifiableInformation','FreeOfAdultContent']} } )
AWS CLI

The following request example uses the AWS CLI. For more information, see analyze-document in the AWS CLI Command Reference.

$ aws textract analyze-document \ --document '{"S3Object":{"Bucket":"bucket_name","Name":"document_name"}}' \ --human-loop-config HumanLoopName="human_loop_name",FlowDefinitionArn="arn:aws:sagemaker:aws-region:aws_account_number:flow-definition/flow_def_name",DataAttributes='{ContentClassifiers=["FreeOfPersonallyIdentifiableInformation", "FreeOfAdultContent"]}' \ --feature-types '["TABLES", "FORMS"]'
$ aws textract analyze-document \ --document '{"S3Object":{"Bucket":"bucket_name","Name":"document_name"}}' \ --human-loop-config \ '{"HumanLoopName":"human_loop_name","FlowDefinitionArn":"arn:aws:sagemaker:aws_region:aws_account_number:flow-definition/flow_def_name","DataAttributes": {"ContentClassifiers":["FreeOfPersonallyIdentifiableInformation","FreeOfAdultContent"]}}' \ --feature-types '["TABLES", "FORMS"]'

After you run AnalyzeDocument with a human loop configured, Amazon A2I monitors the results from AnalyzeDocument and checks it against the flow definition's activation conditions. If the Amazon Textract inference confidence score for one or more key-value pairs meets the conditions for review, Amazon A2I starts a human review loop and includes the HumanLoopActivationOutput object in the AnalyzeDocument response.

Create an Amazon Rekognition Human Loop

Amazon A2I integrates with Amazon Rekognition so that you can configure and start a human loop using the Amazon Rekognition API. To send images to Amazon Rekognition for content moderation, you use the Amazon Rekognition DetectModerationLabels API operation. To configure a human loop, set the HumanLoopConfig parameter when you configure DetectModerationLabels.

When you configure your human loop, the flow definition you specify in FlowDefinitionArn of HumanLoopConfig must be located in the same AWS Region as the S3 bucket identified in Bucket of the Image parameter.

The following table shows examples of how to use this operation with the AWS CLI and AWS SDK for Python (Boto3).

AWS SDK for Python (Boto3)

The following request example uses the SDK for Python (Boto3). For more information, see detect_moderation_labels in the AWS SDK for Python (Boto) API Reference.

import boto3 rekognition = boto3.client("rekognition", aws_region) response = rekognition.detect_moderation_labels( \ Image={'S3Object': {'Bucket': bucket_name, 'Name': image_name}}, \ HumanLoopConfig={ \ 'HumanLoopName': 'human_loop_name', \ 'FlowDefinitionArn': , "arn:aws:sagemaker:aws_region:aws_account_number:flow-definition/flow_def_name" \ 'DataAttributes': {'ContentClassifiers': ['FreeOfPersonallyIdentifiableInformation','FreeOfAdultContent']} })
AWS CLI

The following request example uses the AWS CLI. For more information, see detect-moderation-labels in the AWS CLI Command Reference.

$ aws rekognition detect-moderation-labels \ --image "S3Object={Bucket='bucket_name',Name='image_name'}" \ --human-loop-config HumanLoopName="human_loop_name",FlowDefinitionArn="arn:aws:sagemaker:aws_region:aws_account_number:flow-definition/flow_def_name",DataAttributes='{ContentClassifiers=["FreeOfPersonallyIdentifiableInformation", "FreeOfAdultContent"]}'
$ aws rekognition detect-moderation-labels \ --image "S3Object={Bucket='bucket_name',Name='image_name'}" \ --human-loop-config \ '{"HumanLoopName": "human_loop_name", "FlowDefinitionArn": "arn:aws:sagemaker:aws_region:aws_account_number:flow-definition/flow_def_name", "DataAttributes": {"ContentClassifiers": ["FreeOfPersonallyIdentifiableInformation", "FreeOfAdultContent"]}}'

After you run DetectModerationLabels with a human loop configured, Amazon A2I monitors the results from DetectModerationLabels and checks it against the flow definition's activation conditions. If the Amazon Rekognition inference confidence score for an image meets the conditions for review, Amazon A2I starts a human review loop and includes the response element HumanLoopActivationOutput in the DetectModerationLabels response.

Create and Start a Human Loop for a Custom Task Type

To configure a human loop for a custom human review task, use the StartHumanLoop operation within your application. This section provides an example of a human loop request using the AWS SDK for Python (Boto3) and the AWS Command Line Interface (AWS CLI). For documentation on other language-specific SDKs that support StartHumanLoop, use the See Also section of StartHumanLoop in the Amazon Augmented AI Runtime API documentation. Refer to Use Cases and Examples Using Amazon A2I to see examples that demonstrate how to use Amazon A2I with a custom task type.

Prerequisites

To complete this procedure, you need:

  • Input data formatted as a string representation of a JSON-formatted file

  • The Amazon Resource Name (ARN) of your flow definition

To configure the human loop
  1. For DataAttributes, specify a set of ContentClassifiers related to the input provided to the StartHumanLoop operation. Use content classifiers to declare that your content is free of personally identifiable information or adult content.

    To use Amazon Mechanical Turk, ensure your data is free of personally identifiable information, including protected health information under HIPAA, and include the FreeOfPersonallyIdentifiableInformation content classifier. If you do not use this content classifier, SageMaker does not send your task to Mechanical Turk. If your data is free of adult content, also include the 'FreeOfAdultContent' classifier. If you do not use these content classifiers, SageMaker may restrict the Mechanical Turk workers that can view your task.

  2. For FlowDefinitionArn, enter the Amazon Resource Name (ARN) of your flow definition.

  3. For HumanLoopInput, enter your input data as a string representation of a JSON-formatted file. Structure your input data and custom worker task template so that your input data is properly displayed to human workers when you start your human loop. See Preview a Worker Task Template to learn how to preview your custom worker task template.

  4. For HumanLoopName, enter a name for the human loop. The name must be unique within the Region in your account and can have up to 63 characters. Valid characters are a-z, 0-9, and - (hyphen).

To start a human loop
  • To start a human loop, submit a request similar to the following examples using your preferred language-specific SDK.

AWS SDK for Python (Boto3)

The following request example uses the SDK for Python (Boto3). For more information, see Boto 3 Augmented AI Runtime in the AWS SDK for Python (Boto) API Reference.

import boto3 a2i_runtime_client = boto3.client('sagemaker-a2i-runtime') response = a2i_runtime_client.start_human_loop( HumanLoopName='human_loop_name', FlowDefinitionArn='arn:aws:sagemaker:aws-region:xyz:flow-definition/flow_def_name', HumanLoopInput={ 'InputContent': '{"InputContent": {\"prompt\":\"What is the answer?\"}}' }, DataAttributes={ 'ContentClassifiers': [ 'FreeOfPersonallyIdentifiableInformation'|'FreeOfAdultContent', ] } )
AWS CLI

The following request example uses the AWS CLI. For more information, see start-human-loop in the AWS CLI Command Reference.

$ aws sagemaker-a2i-runtime start-human-loop --flow-definition-arn 'arn:aws:sagemaker:aws_region:xyz:flow-definition/flow_def_name' \ --human-loop-name 'human_loop_name' \ --human-loop-input '{"InputContent": "{\"prompt\":\"What is the answer?\"}"}' \ --data-attributes ContentClassifiers="FreeOfPersonallyIdentifiableInformation","FreeOfAdultContent" \

When you successfully start a human loop by invoking StartHumanLoop directly, the response includes a HumanLoopARN and a HumanLoopActivationResults object which is set to NULL. You can use this the human loop name to monitor and manage your human loop.

Next Steps:

After starting a human loop, you can manage and monitor it with the Amazon Augmented AI Runtime API and Amazon CloudWatch Events. To learn more, see Monitor and Manage Your Human Loop.