splitType

The method to use to split the transform job's data files into smaller batches. Splitting is necessary when the total size of each object is too large to fit in a single request. You can also use data splitting to improve performance by processing multiple concurrent mini-batches. The default value for SplitType is None, which indicates that input data files are not split, and request payloads contain the entire contents of an input object. Set the value of this parameter to Line to split records on a newline character boundary. SplitType also supports a number of record-oriented binary data formats. Currently, the supported record formats are:

• RecordIO

• TFRecord

When splitting is enabled, the size of a mini-batch depends on the values of the BatchStrategy and MaxPayloadInMB parameters. When the value of BatchStrategy is MultiRecord, Amazon SageMaker sends the maximum number of records in each request, up to the MaxPayloadInMB limit. If the value of BatchStrategy is SingleRecord, Amazon SageMaker sends individual records in each request.

Some data formats represent a record as a binary payload wrapped with extra padding bytes. When splitting is applied to a binary data format, padding is removed if the value of BatchStrategy is set to SingleRecord. Padding is not removed if the value of BatchStrategy is set to MultiRecord.

For more information about RecordIO, see Create a Dataset Using RecordIO in the MXNet documentation. For more information about TFRecord, see Consuming TFRecord data in the TensorFlow documentation.