Amazon SageMaker
Developer Guide

CreateTrainingJob

Starts a model training job. After training completes, Amazon SageMaker saves the resulting model artifacts to an Amazon S3 location that you specify.

If you choose to host your model using Amazon SageMaker hosting services, you can use the resulting model artifacts as part of the model. You can also use the artifacts in a machine learning service other than Amazon SageMaker, provided that you know how to use them for inferences.

In the request body, you provide the following:

  • AlgorithmSpecification - Identifies the training algorithm to use.

  • HyperParameters - Specify these algorithm-specific parameters to enable the estimation of model parameters during training. Hyperparameters can be tuned to optimize this learning process. For a list of hyperparameters for each training algorithm provided by Amazon SageMaker, see Algorithms.

  • InputDataConfig - Describes the training dataset and the Amazon S3, EFS, or FSx location where it is stored.

  • OutputDataConfig - Identifies the Amazon S3 bucket where you want Amazon SageMaker to save the results of model training.

  • ResourceConfig - Identifies the resources, ML compute instances, and ML storage volumes to deploy for model training. In distributed training, you specify more than one instance.

  • EnableManagedSpotTraining - Optimize the cost of training machine learning models by up to 80% by using Amazon EC2 Spot instances. For more information, see Managed Spot Training.

  • RoleARN - The Amazon Resource Number (ARN) that Amazon SageMaker assumes to perform tasks on your behalf during model training. You must grant this role the necessary permissions so that Amazon SageMaker can successfully complete model training.

  • StoppingCondition - To help cap training costs, use MaxRuntimeInSeconds to set a time limit for training. Use MaxWaitTimeInSeconds to specify how long you are willing to wait for a managed spot training job to complete.

For more information about Amazon SageMaker, see How It Works.

Request Syntax

{ "AlgorithmSpecification": { "AlgorithmName": "string", "EnableSageMakerMetricsTimeSeries": boolean, "MetricDefinitions": [ { "Name": "string", "Regex": "string" } ], "TrainingImage": "string", "TrainingInputMode": "string" }, "CheckpointConfig": { "LocalPath": "string", "S3Uri": "string" }, "DebugHookConfig": { "CollectionConfigurations": [ { "CollectionName": "string", "CollectionParameters": { "string" : "string" } } ], "HookParameters": { "string" : "string" }, "LocalPath": "string", "S3OutputPath": "string" }, "DebugRuleConfigurations": [ { "InstanceType": "string", "LocalPath": "string", "RuleConfigurationName": "string", "RuleEvaluatorImage": "string", "RuleParameters": { "string" : "string" }, "S3OutputPath": "string", "VolumeSizeInGB": number } ], "EnableInterContainerTrafficEncryption": boolean, "EnableManagedSpotTraining": boolean, "EnableNetworkIsolation": boolean, "ExperimentConfig": { "ExperimentName": "string", "TrialComponentDisplayName": "string", "TrialName": "string" }, "HyperParameters": { "string" : "string" }, "InputDataConfig": [ { "ChannelName": "string", "CompressionType": "string", "ContentType": "string", "DataSource": { "FileSystemDataSource": { "DirectoryPath": "string", "FileSystemAccessMode": "string", "FileSystemId": "string", "FileSystemType": "string" }, "S3DataSource": { "AttributeNames": [ "string" ], "S3DataDistributionType": "string", "S3DataType": "string", "S3Uri": "string" } }, "InputMode": "string", "RecordWrapperType": "string", "ShuffleConfig": { "Seed": number } } ], "OutputDataConfig": { "KmsKeyId": "string", "S3OutputPath": "string" }, "ResourceConfig": { "InstanceCount": number, "InstanceType": "string", "VolumeKmsKeyId": "string", "VolumeSizeInGB": number }, "RoleArn": "string", "StoppingCondition": { "MaxRuntimeInSeconds": number, "MaxWaitTimeInSeconds": number }, "Tags": [ { "Key": "string", "Value": "string" } ], "TensorBoardOutputConfig": { "LocalPath": "string", "S3OutputPath": "string" }, "TrainingJobName": "string", "VpcConfig": { "SecurityGroupIds": [ "string" ], "Subnets": [ "string" ] } }

Request Parameters

For information about the parameters that are common to all actions, see Common Parameters.

The request accepts the following data in JSON format.

AlgorithmSpecification

The registry path of the Docker image that contains the training algorithm and algorithm-specific metadata, including the input mode. For more information about algorithms provided by Amazon SageMaker, see Algorithms. For information about providing your own algorithms, see Using Your Own Algorithms with Amazon SageMaker.

Type: AlgorithmSpecification object

Required: Yes

CheckpointConfig

Contains information about the output location for managed spot training checkpoint data.

Type: CheckpointConfig object

Required: No

DebugHookConfig

Configuration information for the debug hook parameters, collection configuration, and storage paths.

Type: DebugHookConfig object

Required: No

DebugRuleConfigurations

Configuration information for debugging rules.

Type: Array of DebugRuleConfiguration objects

Array Members: Minimum number of 0 items. Maximum number of 20 items.

Required: No

EnableInterContainerTrafficEncryption

To encrypt all communications between ML compute instances in distributed training, choose True. Encryption provides greater security for distributed training, but training might take longer. How long it takes depends on the amount of communication between compute instances, especially if you use a deep learning algorithm in distributed training. For more information, see Protect Communications Between ML Compute Instances in a Distributed Training Job.

Type: Boolean

Required: No

EnableManagedSpotTraining

To train models using managed spot training, choose True. Managed spot training provides a fully managed and scalable infrastructure for training machine learning models. this option is useful when training jobs can be interrupted and when there is flexibility when the training job is run.

The complete and intermediate results of jobs are stored in an Amazon S3 bucket, and can be used as a starting point to train models incrementally. Amazon SageMaker provides metrics and logs in CloudWatch. They can be used to see when managed spot training jobs are running, interrupted, resumed, or completed.

Type: Boolean

Required: No

EnableNetworkIsolation

Isolates the training container. No inbound or outbound network calls can be made, except for calls between peers within a training cluster for distributed training. If you enable network isolation for training jobs that are configured to use a VPC, Amazon SageMaker downloads and uploads customer data and model artifacts through the specified VPC, but the training container does not have network access.

Note

The Semantic Segmentation built-in algorithm does not support network isolation.

Type: Boolean

Required: No

ExperimentConfig

Configuration for the experiment.

Type: ExperimentConfig object

Required: No

HyperParameters

Algorithm-specific parameters that influence the quality of the model. You set hyperparameters before you start the learning process. For a list of hyperparameters for each training algorithm provided by Amazon SageMaker, see Algorithms.

You can specify a maximum of 100 hyperparameters. Each hyperparameter is a key-value pair. Each key and value is limited to 256 characters, as specified by the Length Constraint.

Type: String to string map

Key Length Constraints: Maximum length of 256.

Key Pattern: .*

Value Length Constraints: Maximum length of 256.

Value Pattern: .*

Required: No

InputDataConfig

An array of Channel objects. Each channel is a named input source. InputDataConfig describes the input data and its location.

Algorithms can accept input data from one or more channels. For example, an algorithm might have two channels of input data, training_data and validation_data. The configuration for each channel provides the S3, EFS, or FSx location where the input data is stored. It also provides information about the stored data: the MIME type, compression method, and whether the data is wrapped in RecordIO format.

Depending on the input mode that the algorithm supports, Amazon SageMaker either copies input data files from an S3 bucket to a local directory in the Docker container, or makes it available as input streams. For example, if you specify an EFS location, input data files will be made available as input streams. They do not need to be downloaded.

Type: Array of Channel objects

Array Members: Minimum number of 1 item. Maximum number of 20 items.

Required: No

OutputDataConfig

Specifies the path to the S3 location where you want to store model artifacts. Amazon SageMaker creates subfolders for the artifacts.

Type: OutputDataConfig object

Required: Yes

ResourceConfig

The resources, including the ML compute instances and ML storage volumes, to use for model training.

ML storage volumes store model artifacts and incremental states. Training algorithms might also use ML storage volumes for scratch space. If you want Amazon SageMaker to use the ML storage volume to store the training data, choose File as the TrainingInputMode in the algorithm specification. For distributed training algorithms, specify an instance count greater than 1.

Type: ResourceConfig object

Required: Yes

RoleArn

The Amazon Resource Name (ARN) of an IAM role that Amazon SageMaker can assume to perform tasks on your behalf.

During model training, Amazon SageMaker needs your permission to read input data from an S3 bucket, download a Docker image that contains training code, write model artifacts to an S3 bucket, write logs to Amazon CloudWatch Logs, and publish metrics to Amazon CloudWatch. You grant permissions for all of these tasks to an IAM role. For more information, see Amazon SageMaker Roles.

Note

To be able to pass this role to Amazon SageMaker, the caller of this API must have the iam:PassRole permission.

Type: String

Length Constraints: Minimum length of 20. Maximum length of 2048.

Pattern: ^arn:aws[a-z\-]*:iam::\d{12}:role/?[a-zA-Z_0-9+=,.@\-_/]+$

Required: Yes

StoppingCondition

Specifies a limit to how long a model training job can run. When the job reaches the time limit, Amazon SageMaker ends the training job. Use this API to cap model training costs.

To stop a job, Amazon SageMaker sends the algorithm the SIGTERM signal, which delays job termination for 120 seconds. Algorithms can use this 120-second window to save the model artifacts, so the results of training are not lost.

Type: StoppingCondition object

Required: Yes

Tags

An array of key-value pairs. For more information, see Using Cost Allocation Tags in the AWS Billing and Cost Management User Guide.

Type: Array of Tag objects

Array Members: Minimum number of 0 items. Maximum number of 50 items.

Required: No

TensorBoardOutputConfig

Configuration of storage locations for TensorBoard output.

Type: TensorBoardOutputConfig object

Required: No

TrainingJobName

The name of the training job. The name must be unique within an AWS Region in an AWS account.

Type: String

Length Constraints: Minimum length of 1. Maximum length of 63.

Pattern: ^[a-zA-Z0-9](-*[a-zA-Z0-9])*

Required: Yes

VpcConfig

A VpcConfig object that specifies the VPC that you want your training job to connect to. Control access to and from your training container by configuring the VPC. For more information, see Protect Training Jobs by Using an Amazon Virtual Private Cloud.

Type: VpcConfig object

Required: No

Response Syntax

{ "TrainingJobArn": "string" }

Response Elements

If the action is successful, the service sends back an HTTP 200 response.

The following data is returned in JSON format by the service.

TrainingJobArn

The Amazon Resource Name (ARN) of the training job.

Type: String

Length Constraints: Maximum length of 256.

Pattern: arn:aws[a-z\-]*:sagemaker:[a-z0-9\-]*:[0-9]{12}:training-job/.*

Errors

For information about the errors that are common to all actions, see Common Errors.

ResourceInUse

Resource being accessed is in use.

HTTP Status Code: 400

ResourceLimitExceeded

You have exceeded an Amazon SageMaker resource limit. For example, you might have too many training jobs created.

HTTP Status Code: 400

ResourceNotFound

Resource being access is not found.

HTTP Status Code: 400

See Also

For more information about using this API in one of the language-specific AWS SDKs, see the following: