AlgorithmSpecification

The training input mode that the algorithm supports. For more information about input modes, see Algorithms.

Pipe mode

If an algorithm supports Pipe mode, Amazon SageMaker streams data directly from Amazon S3 to the container.

File mode

If an algorithm supports File mode, SageMaker downloads the training data from S3 to the provisioned ML storage volume, and mounts the directory to the Docker volume for the training container.

You must provision the ML storage volume with sufficient capacity to accommodate the data downloaded from S3. In addition to the training data, the ML storage volume also stores the output model. The algorithm container uses the ML storage volume to also store intermediate information, if any.

For distributed algorithms, training data is distributed uniformly. Your training duration is predictable if the input data objects sizes are approximately the same. SageMaker does not split the files any further for model training. If the object sizes are skewed, training won't be optimal as the data distribution is also skewed when one host in a training cluster is overloaded, thus becoming a bottleneck in training.

FastFile mode

If an algorithm supports FastFile mode, SageMaker streams data directly from S3 to the container with no code changes, and provides file system access to the data. Users can author their training script to interact with these files as if they were stored on disk.

FastFile mode works best when the data is read sequentially. Augmented manifest files aren't supported. The startup time is lower when there are fewer files in the S3 bucket provided.

Type: String

Valid Values: Pipe | File | FastFile

Required: Yes

The name of the algorithm resource to use for the training job. This must be an algorithm resource that you created or subscribe to on Amazon Marketplace.

Type: String

Length Constraints: Minimum length of 1. Maximum length of 170.

Pattern: (arn:aws[a-z\-]*:sagemaker:[a-z0-9\-]*:[0-9]{12}:[a-z\-]*\/)?([a-zA-Z0-9]([a-zA-Z0-9-]){0,62})(?<!-)$

Required: No

The arguments for a container used to run a training job. See How Amazon SageMaker Runs Your Training Image for additional information.

Type: Array of strings

Array Members: Minimum number of 1 item. Maximum number of 100 items.

Length Constraints: Maximum length of 256.

Pattern: .*

Required: No

The entrypoint script for a Docker container used to run a training job. This script takes precedence over the default train processing instructions. See How Amazon SageMaker Runs Your Training Image for more information.

Type: Array of strings

Array Members: Minimum number of 1 item. Maximum number of 100 items.

Length Constraints: Maximum length of 256.

Pattern: .*

Required: No

To generate and save time-series metrics during training, set to true. The default is false and time-series metrics aren't generated except in the following cases:

Type: Boolean

Required: No

A list of metric definition objects. Each object specifies the metric name and regular expressions used to parse algorithm logs. SageMaker publishes each metric to Amazon CloudWatch.

Type: Array of MetricDefinition objects

Array Members: Minimum number of 0 items. Maximum number of 40 items.

Required: No

The registry path of the Docker image that contains the training algorithm. For information about docker registry paths for SageMaker built-in algorithms, see Docker Registry Paths and Example Code in the Amazon SageMaker developer guide. SageMaker supports both registry/repository[:tag] and registry/repository[@digest] image path formats. For more information about using your custom training container, see Using Your Own Algorithms with Amazon SageMaker.

Type: String

Length Constraints: Maximum length of 255.

Pattern: .*

Required: No

The configuration to use an image from a private Docker registry for a training job.

Type: TrainingImageConfig object

Required: No