spark_write_tfrecord

Write a Spark DataFrame to a TFRecord file

Description

Serialize a Spark DataFrame to the TensorFlow TFRecord format for training or inference.

Usage

spark_write_tfrecord(x, path, record_type = c("Example",
  "SequenceExample"), write_locality = c("distributed", "local"),
  mode = NULL)

Arguments

Argument Description
x A Spark DataFrame
path The path to the file. Needs to be accessible from the cluster. Supports the “hdfs://”, “s3a://”, and “file://” protocols.
record_type Output format of TensorFlow records. One of "Example" and "SequenceExample".
write_locality Determines whether the TensorFlow records are written locally on the workers or on a distributed file system. One of "distributed" and "local". See Details for more information.
mode A character element. Specifies the behavior when data or table already exists. Supported values include: ‘error’, ‘append’, ‘overwrite’ and ‘ignore’. Notice that ‘overwrite’ will also change the column structure. For more details see also http://spark.apache.org/docs/latest/sql-programming-guide.html#save-modes for your version of Spark.

Details

For write_locality = local, each of the workers stores on the local disk a subset of the data. The subset that is stored on each worker is determined by the partitioning of the DataFrame. Each of the partitions is coalesced into a single TFRecord file and written on the node where the partition lives. This is useful in the context of distributed training, in which each of the workers gets a subset of the data to work on. When this mode is activated, the path provided to the writer is interpreted as a base path that is created on each of the worker nodes, and that will be populated with data from the DataFrame.