tar_format {targets}

Define a custom target storage format.

Description

Define a custom target storage format for the format argument of tar_target() or tar_option_set().

Usage

tar_format(
  read = NULL,
  write = NULL,
  marshal = NULL,
  unmarshal = NULL,
  convert = NULL,
  copy = NULL,
  repository = NULL
)

Arguments

read

A function with a single argument named path. This function should read and return the target stored at the file in the argument. It should have no side effects. See the "Format functions" section for specific requirements. If NULL, the read argument defaults to readRDS().

write

A function with two arguments: object and path, in that order. This function should save the R object object to the file path at path and have no other side effects. The function need not return a value, but the file written to path must be a single file, and it cannot be a directory. See the "Format functions" section for specific requirements. If NULL, the write argument defaults to saveRDS() with version = 3.

marshal

A function with a single argument named object. This function should marshal the R object and return an in-memory object that can be exported to remote parallel workers. It should not read or write any persistent files. See the Marshalling section for details. See the "Format functions" section for specific requirements. If NULL, the marshal argument defaults to just returning the original object without any modifications.

unmarshal

A function with a single argument named object. This function should unmarshal the (marshalled) R object and return an in-memory object that is appropriate and valid for use on a parallel worker. It should not read or write any persistent files. See the Marshalling section for details. See the "Format functions" section for specific requirements. If NULL, the unmarshal argument defaults to just returning the original object without any modifications.

convert

The convert argument is a function that accepts the object returned by the command of the target and changes it into an acceptable format (e.g. can be saved with the read function). The convert ensures the in-memory copy of an object during the running pipeline session is the same as the copy of the object that is saved to disk. The function should be idempotent, and it should handle edge cases like NULL values (especially for error = "null" in tar_target() or tar_option_set()). If NULL, the convert argument defaults to just returning the original object without any modifications.

copy

The copy argument is a function that accepts the object returned by the command of the target and makes a deep copy in memory. This method does is relevant to objects like data.tables that support in-place modification which could cause unpredictable side effects from target to target. In cases like these, the target should be deep-copied before a downstream target attempts to use it (in the case of data.table objects, using data.table::copy()). If NULL, the copy argument defaults to just returning the original object without any modifications.

repository

Deprecated. Use the repository argument of tar_target() or tar_option_set() instead.

Details

It is good practice to write formats that correctly handle NULL objects if you are planning to set error = "null" in tar_option_set().

Value

A character string of length 1 encoding the custom format. You can supply this string directly to the format argument of tar_target() or tar_option_set().

Marshalling

If an object can only be used in the R session where it was created, it is called "non-exportable". Examples of non-exportable R objects are Keras models, Torch objects, xgboost matrices, xml2 documents, rstan model objects, sparklyr data objects, and database connection objects. These objects cannot be exported to parallel workers (e.g. for tar_make_future()) without special treatment. To send an non-exportable object to a parallel worker, the object must be marshalled: converted into a form that can be exported safely (similar to serialization but not always the same). Then, the worker must unmarshal the object: convert it into a form that is usable and valid in the current R session. Arguments marshal and unmarshal of tar_format() let you control how marshalling and unmarshalling happens.

Format functions

In tar_format(), functions like read, write, marshal, and unmarshal must be perfectly pure and perfectly self-sufficient. They must load or namespace all their own packages, and they must not depend on any custom user-defined functions or objects in the global environment of your pipeline. targets converts each function to and from text, so it must not rely on any data in the closure. This disqualifies functions produced by Vectorize(), for example.

The write function must write only a single file, and the file it writes must not be a directory.

The functions to read and write the object should not do any conversions on the object. That is the job of the convert argument. The convert argument is a function that accepts the object returned by the command of the target and changes it into an acceptable format (e.g. can be saved with the read function). Working with the convert function is best because it ensures the in-memory copy of an object during the running pipeline session is the same as the copy of the object that is saved to disk.

See Also

Other targets: tar_cue(), tar_target(), tar_target_raw()

Examples

# The following target is equivalent to the current superseded
# tar_target(name, command(), format = "keras").
# An improved version of this would supply a `convert` argument
# to handle NULL objects, which are returned by the target if it
# errors and the error argument of tar_target() is "null".
tar_target(
  name = keras_target,
  command = your_function(),
  format = tar_format(
    read = function(path) {
      keras::load_model_hdf5(path)
    },
    write = function(object, path) {
      keras::save_model_hdf5(object = object, filepath = path)
    },
    marshal = function(object) {
      keras::serialize_model(object)
    },
    unmarshal = function(object) {
      keras::unserialize_model(object)
    }
  )
)
# And the following is equivalent to the current superseded
# tar_target(name, torch::torch_tensor(seq_len(4)), format = "torch"),
# except this version has a `convert` argument to handle
# cases when `NULL` is returned (e.g. if the target errors out
# and the `error` argument is "null" in tar_target()
# or tar_option_set())
tar_target(
  name = torch_target,
  command = torch::torch_tensor(),
  format = tar_format(
    read = function(path) {
      torch::torch_load(path)
    },
    write = function(object, path) {
      torch::torch_save(obj = object, path = path)
    },
    marshal = function(object) {
      con <- rawConnection(raw(), open = "wr")
      on.exit(close(con))
      torch::torch_save(object, con)
      rawConnectionValue(con)
    },
    unmarshal = function(object) {
      con <- rawConnection(object, open = "r")
      on.exit(close(con))
      torch::torch_load(con)
    }
  )
)