Prepare a dataset for analysis

Transform a dataset with named columns into a list with features (x) and response (y) elements.

dataset_prepare(dataset, x, y = NULL, named = TRUE,
  named_features = FALSE, parallel_records = NULL, batch_size = NULL,
  num_parallel_batches = NULL, drop_remainder = FALSE)



A dataset


Features to include. When named_features is FALSE all features will be stacked into a single tensor so must have an identical data type.


(Optional). Response variable.


TRUE to name the dataset elements "x" and "y", FALSE to not name the dataset elements.


TRUE to yield features as a named list; FALSE to stack features into a single array. Note that in the case of FALSE (the default) all features will be stacked into a single 2D tensor so need to have the same underlying data type.


(Optional) An integer, representing the number of records to decode in parallel. If not specified, records will be processed sequentially.


(Optional). Batch size if you would like to fuse the dataset_prepare() operation together with a dataset_batch() (fusing generally improves overall training performance).


(Optional) An integer, representing the number of batches to create in parallel. On one hand, higher values can help mitigate the effect of stragglers. On the other hand, higher values can increase contention if CPU is scarce.


Ensure that batches have a fixed size by omitting any final smaller batch if it's present. Note that this is required for use with the Keras tensor inputs to fit/evaluate/etc.


A dataset. The dataset will have a structure of either:

  • When named_features is TRUE: list(x = list(feature_name = feature_values, ...), y = response_values)

  • When named_features is FALSE: list(x = features_array, y = response_values), where features_array is a Rank 2 array of (batch_size, num_features).

Note that the y element will be omitted when y is NULL.

See also

input_fn() for use with tfestimators.